Read Microsoft Word - Levi_diss_shortversion.doc text version

©Copyright 2004 Susannah V. Levi

The representation of underlying glides: a cross-linguistic study

Susannah V. Levi

A dissertation submitted in partial fulfillment of the requirements for the degree of

Doctor of Philosophy

University of Washington 2004

Program Authorized to Offer Degree: Department of Linguistics

University of Washington Graduate School

This is to certify that I have examined this copy of a doctoral dissertation by

Susannah V. Levi

and have found that it is complete and satisfactory in all respects, and that any and all revisions required by the final examining committee have been made.

Chair of Supervisory Committee: ________________________________________________________________ Ellen M. Kaisse Reading Committee: ________________________________________________________________ Ellen M. Kaisse ________________________________________________________________ Sharon Hargus ________________________________________________________________ Richard Wright Date: ____________________________

In presenting this dissertation in partial fulfillment of the requirements for the doctoral degree at the University of Washington, I agree that the Library shall make its copies freely available for inspection. I further agree that extensive copying of the dissertation is allowable only for scholarly purposes, consistent with "fair use" as prescribed in the U.S. Copyright Law. Requests for copying or reproduction of this dissertation may be referred to Proquest Information and Learning, 300 North Zeeb Road, Ann Arbor, MI 481061346, to whom the author has granted "the right to reproduce and sell (a) copies of the manuscript in microform and/or (b) printed copies of the manuscript made from microform." Signature ______________________________________ Date __________________________________________

University of Washington Abstract The representation of underlying glides: a cross-linguistic study Susannah V. Levi Chair of the Supervisory Committee: Professor Ellen M. Kaisse Department of Linguistics This dissertation examines languages with underlying glides and provides a crosslinguistic representation of these segments. Glides pose a problem to theories of features and natural classes by variably patterning with vowels or consonants. I maintain that the reason for their varied behavior lies in their phonemic status, or lack thereof. Phonemic (or underlying) glides pattern with consonants in a variety of phonological processes. Derived glides, on the other hand, are realizations of underlying vowels and pattern with them. The major contributions of this dissertation are (1) to prove that underlying glides do exist, (2) to provide a cross-linguistic representation for these segments that crucially distinguishes them from other similar segments (e.g. vowels), (3) to present a typology of mappings from underlying to surface vocoids, and (4) to provide a set of criteria that can be used to establish the existence of underlying glides. I show that the following nine languages have underlying glides: Imdlawn Tashlhiyt Berber (Afro-Asiatic), Cree (Algonquian), Karuk (Hokan), Sundanese (Austronesian), Yawelmani (Yokuts), Tahltan (Athabaskan), Pulaar (Niger-Congo), Turkish (Altaic), and Pashto (Indo-Iranian). The criteria for determining the phonemic

status of glides in these languages comes from (1) transparency to vowel harmony, (2) triggering of consonant harmony, (3) necessity of vowel epenthesis in unsyllabifiable clusters which include glides, (4) expected and unexpected syllabification of vocoids, (5) reverse sonority clusters, (6) blockage of nasal harmony, and (7) consonant gradation. I compare four representations for underlying glides. The first employs a lexical marking (LEX) which forces some underlying vocoids into non-syllabic positions. The second representation uses the feature [consonantal] to distinguish underlying vowels from underlying glides (CONS). This has been the most widely used method. However, both LEX and CONS fail to distinguish underlying from derived glides in processes which refer to distinctive features. Finally, I compare two feature geometries. The first segregates vowel and consonant place features on two different tiers (Vowel-Place Theory). The second uses only a single tier of place features and the markedness theory developed in Calabrese 1995 (Revised Articulator Theory (RAT)). When confronted with a variety of evidence, RAT fares better than Vowel-Place Theory.

TABLE OF CONTENTS 1 Introduction................................................................................................................. 1 1.1 Overview............................................................................................................. 1 1.2 Minimal Pairs...................................................................................................... 4 1.3 Overview of the languages used ......................................................................... 5 1.4 Criteria ................................................................................................................ 6 1.5 Typology ............................................................................................................. 8 1.6 Organization of this dissertation ....................................................................... 10 2 Background and history of glides in phonetics and phonology................................ 12 2.1 Introduction....................................................................................................... 12 2.2 History of glides: phonetics, features, and phonology...................................... 12 2.2.1 Phonetics of glides .................................................................................... 12 Constriction as a distinguishing property ............................................. 12 Duration as a distinguishing property ................................................... 14 Amplitude as a distinguishing property ................................................ 15 Position in the syllable as a distinguishing property............................. 15 Summary ............................................................................................... 15 2.2.2 Features and phonology ............................................................................ 16 Early generative phonology .................................................................. 16 Syllable-based phonology..................................................................... 17 Glides as [+consonantal]....................................................................... 21 Feature geometric glides ....................................................................... 23 2.3 Summary ........................................................................................................... 23 3 Possible Representations for Underlying Glides ...................................................... 25 3.1 Overview........................................................................................................... 25 3.2 Lexical Marking (LEX) .................................................................................... 25 3.2.1 Overview and Representation ................................................................... 25 3.2.2 Predictions................................................................................................. 26 3.3 [±consonantal] ([CONS]).................................................................................. 27 3.3.1 Overview and Representation ................................................................... 27 3.3.2 Predictions................................................................................................. 29 3.4 Vowel-Place Theory (VPT) .............................................................................. 29 3.4.1 Overview and Representation ................................................................... 29 3.4.2 Predictions................................................................................................. 33 3.5 Revised Articulator Theory (RAT)................................................................... 34 3.5.1 Overview and Representation ................................................................... 34 3.5.2 Predictions................................................................................................. 42 3.6 Summary ........................................................................................................... 44 4 Turkish ...................................................................................................................... 46 4.1 Introduction....................................................................................................... 46 4.1.1 Vowels and Vowel Harmony.................................................................... 46 i

4.1.2 Consonants and Vowel Harmony ............................................................. 48 4.2 Basic phonology of Turkish: syllabification..................................................... 50 4.3 Evidence for underlying glides ......................................................................... 54 4.3.1 Vowel Harmony........................................................................................ 54 The glide ............................................................................................... 54 Representing harmony: Evidence from laterals in Turkish .................. 54 4.3.2 Native roots............................................................................................... 56 4.4 Comparison of the representations.................................................................... 56 4.4.1 LEX........................................................................................................... 57 4.4.2 [CONS] ..................................................................................................... 58 4.4.3 Vowel-Place Theory ................................................................................. 59 4.4.4 Revised Articulator Theory....................................................................... 64 4.5 Discussion ......................................................................................................... 70 5 Pulaar ........................................................................................................................ 71 5.1 Introduction....................................................................................................... 71 5.2 Basic Pulaar Phonology .................................................................................... 72 5.2.1 Sonority Constraint ................................................................................... 73 5.2.2 Marker Shortening .................................................................................... 74 5.3 Evidence for underlying glides ......................................................................... 76 5.3.1 Consonant gradation ................................................................................. 76 Pulaar (Western dialect)........................................................................ 76 Basics of gradation.......................................................................... 76 Paradis's analysis............................................................................. 79 Unmarked Place in Pulaar............................................................... 79 Summary ......................................................................................... 82 Eastern dialects ..................................................................................... 83 Summary ............................................................................................... 83 5.3.2 Vowel Harmony........................................................................................ 84 5.3.3 Vowel Epenthesis...................................................................................... 86 5.3.4 Gemination................................................................................................ 86 5.4 Comparison of the representations.................................................................... 87 5.4.1 LEX........................................................................................................... 88 5.4.2 [CONS] ..................................................................................................... 89 5.4.3 Vowel-Place Theory ................................................................................. 89 5.4.4 Revised Articulator Theory....................................................................... 93 5.4.5 Summary ................................................................................................... 96 6 Yawelmani/Yowlumne ............................................................................................. 97 6.1 Introduction....................................................................................................... 97 6.2 Basic Yawelmani phonology ............................................................................ 99 6.2.1 Basic Processes ......................................................................................... 99 Epenthesis ............................................................................................. 99 Vowel Harmony.................................................................................... 99 ii Lowering ............................................................................................. 100 Shortening ........................................................................................... 101 Hiatus resolution ................................................................................. 101 6.2.2 Templates................................................................................................ 103 6.2.3 Tiers/Planes............................................................................................. 105 6.3 Evidence for underlying glides ....................................................................... 108 6.3.1 Vowel Harmony...................................................................................... 108 6.3.2 Epenthesis ............................................................................................... 110 6.3.3 Other parallels with consonants .............................................................. 112 6.3.4 Summary ................................................................................................. 113 6.4 Comparison of the representations.................................................................. 114 6.4.1 LEX......................................................................................................... 114 6.4.2 [CONS] ................................................................................................... 115 6.4.3 Vowel-Place Theory ............................................................................... 115 6.4.4 Revised Articulator Theory..................................................................... 117 6.5 Discussion ....................................................................................................... 120 7 Tahltan .................................................................................................................... 122 7.1 Introduction..................................................................................................... 122 7.2 Some basic phonology of Tahltan................................................................... 124 7.2.1 Syllabification ......................................................................................... 124 7.2.2 D-effect ................................................................................................... 125 7.2.3 Stem-initial consonants........................................................................... 126 7.2.4 Stem-final consonants............................................................................. 129 7.3 Evidence for underlying glides: Consonant Harmony.................................... 131 7.4 Previous Analyses........................................................................................... 139 7.5 Comparison of the representations.................................................................. 141 7.5.1 LEX......................................................................................................... 141 7.5.2 [CONS] ................................................................................................... 142 7.5.3 Vowel-Place Theory ............................................................................... 143 7.5.4 Revised Articulator Theory..................................................................... 145 7.6 Discussion ....................................................................................................... 149 8 Karuk....................................................................................................................... 151 8.1 Introduction..................................................................................................... 151 8.2 Syllabification in Karuk.................................................................................. 153 8.3 Evidence for underlying glides ....................................................................... 161 8.3.1 Sonorant Nasalization ............................................................................. 162 8.3.2 Vowel-initial suffixes.............................................................................. 163 8.3.3 Vowel harmony in suffixes..................................................................... 164 8.3.4 Allomorphy ............................................................................................. 166 8.3.5 [s]~[] allophony ..................................................................................... 167 8.3.6 Summary ................................................................................................. 169 8.4 Previous analyses of glides in Karuk .............................................................. 169 iii

8.5 Comparison of the representations.................................................................. 169 8.5.1 LEX......................................................................................................... 170 8.5.2 [CONS] ................................................................................................... 170 8.5.3 Vowel-Place Theory ............................................................................... 172 8.5.4 Revised Articulator Theory..................................................................... 176 8.6 Discussion ....................................................................................................... 180 9 Sundanese ............................................................................................................... 181 9.1 Introduction..................................................................................................... 181 9.2 Evidence for underlying glides: Nasal harmony............................................. 182 9.3 Comparing the representations ....................................................................... 185 9.3.1 LEX......................................................................................................... 186 9.3.2 [CONS] ................................................................................................... 186 9.3.3 Vowel-Place Theory ............................................................................... 187 9.3.4 Revised Articulator Theory..................................................................... 188 9.3.5 An alternative approach: Nasalizability Scale ........................................ 189 9.4 Discussion ....................................................................................................... 192 9.5 Appendix: Excursus on [continuant] and nasalization.................................... 192 10 Other languages with underlying glides ............................................................. 194 10.1 Overview......................................................................................................... 194 10.2 Moose Cree ..................................................................................................... 194 10.2.1 Introduction............................................................................................. 194 10.2.2 Basic phonology and morphology of Cree ............................................. 195 10.2.3 Evidence for underlying glides: reverse sonority clusters ...................... 199 10.3 Pashto.............................................................................................................. 202 10.3.1 Introduction............................................................................................. 202 10.3.2 Two aspects of Pashto phonology........................................................... 203 10.3.3 Evidence for underlying glides: reverse sonority clusters ...................... 204 10.4 Imdlawn Tashlhiyt Berber .............................................................................. 207 10.4.1 Introduction............................................................................................. 207 10.4.2 Basics of syllabification.......................................................................... 208 10.4.3 Evidence for underlying glides in ITB: Syllabification.......................... 212 Sequences of high vocoids.............................................................. 212 Interaction between underlying glides and other consonants ......... 214 Prepausal Annexation ..................................................................... 216 Geminate vocoids............................................................................ 217 Summary ......................................................................................... 219 10.4.4 Evidence for underlying glides in ITB: Morphology ............................. 219 Aorist............................................................................................... 219 Imperfect ......................................................................................... 222 Negative .......................................................................................... 224 Summary ......................................................................................... 226 10.5 Representation................................................................................................. 226 iv

11 Languages without underlying glides ................................................................. 228 11.1 Types of languages without underlying glides ............................................... 228 11.2 "Normal" languages ........................................................................................ 230 11.2.1 Kimatuumbi ............................................................................................ 230 11.2.2 Karuk....................................................................................................... 233 11.2.3 Sanskrit ................................................................................................... 236 11.3 "Vocalic" languages........................................................................................ 242 11.3.1 Ait Seghrouchen Berber.......................................................................... 242 11.3.2 Spanish.................................................................................................... 244 11.3.3 Korean..................................................................................................... 246 11.4 Summary ......................................................................................................... 248 12 Conclusion .......................................................................................................... 250 12.1 Typology ......................................................................................................... 250 12.2 Representations ............................................................................................... 251 12.3 Other languages .............................................................................................. 252 References....................................................................................................................... 253



There are several people who have had a profound and direct impact on this dissertation. I would first like to thank Ellen Kaisse, my advisor, who has never waivered in her support for me. Talking to Ellen was always easy because we were inevitably on the same wavelength. I was often able to start explaining something and have Ellen finish my thought. She read every word of my drafts and found even the teeniest problems. I would also like to thank Sharon Hargus who acted as my motivator and cheerleader. Whenever my interest in this topic waned, I knew that a quick trip to Sharon's office would assuage my fears and provide me with the motivation to continue. I would also like to thank Richard Wright for introducing me to phonetics and taking on a green graduate student who just wanted to learn. Though there is little phonetics in this dissertation, his insights have influenced my other work and will certainly continue to do so in my future work. I would also like to thank the other faculty members in the Department of Linguistics for being accessible to the students and creating an open atmosphere. In particular, Alicia Wassink was generous with her time and willing to talk about vowels at great length. Julia Herschensohn was genuinely interested in how things were progressing with me. Also, thanks to Kathryn for always finding the answer to administrative questions. The following linguists were especially helpful in answering questions in person and in email: Bert Vaux, Andrea Calabrese, and Morris Halle. I would also like to thank my family for their support. My mom has been especially understanding, knowing that I need time to work when I go home to visit and not making me feel guilty about it. Thanks also to Lizzie who kept me informed about my mother's goings-on so I didn't have to worry. Thanks to David and Jenny for always making me feel welcome and feeding me as though I were a queen. I would also like to thank Erik Vee for technical and moral support. He knows more than anyone else how much time went into this project. He also knew when I needed to take a break and get out of the house or when I just needed a candy fix. I would like to thank the graduate students in the Department of Linguistics. Special thanks to [d]ulia, [j]ulia, Lesley, Emily, Anna, Jay, Sylwia, Chia-Hui, and Darik. They have helped distract me from work in many ways: Carson, Minis!, 13 Coins, Sushi, DA-club. I'd also like to highlight my chlorine buddies who kept me sane and my stress level lower than it otherwise would have been: Alex, Carolyn, Gary, George, Ivan, Joe, and Sally. I'd also like to thank Rob for many long phone conversations and general advice about academics and life. Finally, I would like to thank the American Association of University Women and the Association for Women in Science for supporting my research financially.



To my father, Albert William Levi, who saw neither the beginning nor the end of this dissertation, but who motivated me throughout the entire process.



1 Introduction

1.1 Overview "Perhaps the most problematic segment type for all theories of phonology is the class of glides" (Hyman 1985). Despite ample work on glide/vowel alternations and underlying and surface vowels and glides (Chomsky & Halle 1968, Clements & Keyser 1983, Kaye & Lowenstamm 1984, Steriade 1984, Hyman 1985, Levin 1985, Deligiorgis 1988;1993, Hayes 1989, Waksler 1990, Rosenthall 1994, Hume 1995, Harris & Kaisse 1999), there is no consensus on either the existence or the representation of underlying glides. Furthermore, a cross-linguistic discussion of the differences between surface glides derived from vowels ("derived glides") and surface glides derived from underlying glides ("underlying glides") is absent. Glides are problematic for phonological theory because of their varied behavior. Sometimes they seem to be vowel-like, either alternating with vowels or patterning with them in various phonological processes. Other times they appear to pattern with consonants. It is this "duality" of glides that has lead to confusions in their status as underlying or derived.1 The problem of how to classify and deal with glides is evident partially in the variable names for them. One also finds the terms semivowel and semi-consonant, illustrating the fact that people variably elect to highlight the vocalic or consonantal quality of these segments. Though the presence of surface or output glides is not disputed, their existence as underlying or distinctive segments is. This dissertation will provide an analysis of surface and underlying glides, as well as a typology and representation of these segments. There have been two prominent ways to deal with glides. The first is to assume that underlying glides do not exist. Under this view, there is no underlying distinction between vowels and glides. Durand 1987 writes, " so many languages, the simplest analysis should treat these semi-vowels simply as the high vowels in a non-syllabic disguise." Such a view is also espoused in Clements & Keyser 1983, Kaye & Lowenstamm 1984, and Rosenthall 1994, among others. In this view, vowels and glides are not distinct on a featural or underlying level, but only on a structural or syllabic level. The difference between vowels and glides on the surface (i.e. syllabic vs. non-syllabic) is simply the result of syllabification. This view explicitly states that underlying glides do not exist because the assumption is that all glides are positional variants of vowels. Such a view is espoused explicitly by Kaye & Lowenstamm 1984 and Deligiorgis 1988.2 The second way to deal with glides is simply to not deal with them explicitly. Despite the rather prominent view above, many articles list glides in the consonant (and

1 2

The term "duality" is used in Hume 1995. Deligiorgis 1988 writes, "I therefore adopt here the traditional position that glides are not phonemic entities and that their underlying segments are high vowels" (pg. 27) (emphasis-SVL).

2 even phoneme) inventories of particular languages. An obvious question to ask is whether the authors actually believe that the language has underlying glides. Oftentimes, glides are not central to the issue and may therefore be simply listed in the inventories out of convenience. I believe that the inclusion of glides in inventories often stems from a lack of consensus about what criteria should be used in determining the underlying representation of a particular segment. The treatment of glides in Herman 1994 and Hume 1995 comes closest to dealing explicitly with this issue and attempts to establish a representation of contrasting glides. Hume 1995 argues for a distinction between consonantal glides on the one hand, and vocalic glides on the other. She argues that there can be two types of underlying glides-- vocalic and consonantal--and that they both have different representations which make different predictions. While Hume's conclusions are close to what I believe to be true, I maintain that she has missed the crucial step, which is to make the connection between the consonantal glides and underlying glides on the one hand, and vocalic glides and underlying vowels or derived glides on the other. I contend that the reason for the varied behavior of glides is due to the fact that some glides are derived from vowels while others are underlying. Under this analysis, there is in fact no "duality" of glides, as they come from two different sources. Their variable behavior is the result of their inherent vowel or glide status. The reason that glides have not been understood and have been viewed as behaving differently from other segments is because their variable behavior was assumed to be derived from a single class of segments. By realizing that glides can be derived from two distinct sources, the variable behavior of a single class is eliminated. The "duality" of glides is thus accounted for and the two different glides each behave in a non-variable and regular fashion. I will argue that underlying glides do exist and that they contrast with underlying vowels (and derived glides). I further contend that glides do not constitute a third class of segments different from vowels and consonants. Glides, or the vocoids in glide/vowel alternations, have sometimes been treated as a special third class of segments. For example in a CV-style article, Steriade 1984 assumes that consonants are dominated by C, vowels by V, and alternating vocoids by X. My findings are that glides are not special and do not behave differently from other segments. Glide/vowel alternations are determined by syllabification, and possibly other prosodic considerations. I maintain that derived glides pattern with vowels and have the same featural representation as they do. Underlying glides, on the other hand, can be grouped with the consonants and form a natural class with them. There is no need to assume a phonetic difference between derived and underlying glides on the surface. In fact, their phonetic equality is what has caused the confusion about their supposed variable and supposedly anomalous behavior. Instead, I argue that these two types of glides have structural differences underlyingly which are also maintained on the surface. Because of the phonetic overlap between underlying and derived glides, glides are fertile ground for

3 reanalysis by language learners. Surface glides can be reanalyzed as derived or underlying. Two dialects of Berber discussed in this dissertation treat glides differently.3 Other dialects of Berber discussed in Rosenthall 1994 exhibit additional patterns. Though I will argue that underlying glides do exist, this does not imply that all languages that contain surface glides also have underlying glides. Only the strictest criteria should be used to determine the status of glides in a particular language. In this dissertation, I will deal mainly with high vowels that alternate with glides ([i] ~ [j] or [u] ~ [w]) and only briefly with mid vowels that show high glide variants ([o] ~ [w]). There are of course other kinds of glides such as the offglide in [e] or the mid glides [o] and [e]. I am unaware of cases where any of these segments are said to be underlyingly distinct from vowels. The schwa off-glide exists in certain English dialects (Labov 1994) but it is the realization of a single unit, either /æ/ or /e/. The schwa offglide in German is a realization of //, as in [ze] 'very' (Hall 1992). In Romanian, midglides are derived from mid vowels and are not underlying (Chitoran 1997). In any case, only the high glides [j] and [w] will be dealt with in this dissertation, whether they are underlying or derived. Before delving into the details of this particular study, I would like to point out that establishing the existence and representation of underlying glides is crucial to any model of phonology. Regardless of what theoretical framework one adheres to, it is crucial to know what segments are present in the language and how those segments are distinguished from one another. In all theories, it is imperative that phonologists know what features distinguish the segments of the language and how those segments are differentiated from on another. In derivational theories that rely on rules and features, the featural representation of underlying glides affects whether or not a rule applies. For example in the rule in (1), it is necessary to know whether it applies to an underlying glide. Under standard interpretations of [-consonantal], it should. (1) [-cons] [round] / [round] _____

Similarly, within Optimality Theory (OT) (Prince & Smolensky 1993), it is necessary to know what features characterize a segment in order to assess violations of IDENT(FEATURE) or MAX(FEATURE) constraints. Without knowing both the features and their organization, OT constraints cannot be evaluated. Furthermore, most OT constraints that deal with vowel harmony make reference to vowels to the exclusion of other segments. In order to accurately evaluate these constraints, it is necessary to know which features are interacting in order to account for the interaction or lack thereof of glides and other consonants.

The dialects are Ait Seghrouchen (Guerssel 1986) and Imdlawn Tashlhiyt (various papers by Dell and Elmedlaoui).


4 (2) Sample harmony constraints (emphasis mine) a. AGREE [feature] Adjacent segments have the same value for a given feature (Bakovic 2003) (Here vowels are assumed to be adjacent and consonants are ignored) b. SPREAD [feature] Spread a feature to all vowels in a word. (Walker 2001) c. ALIGN ([closed], Suffix L, Stress, L) For all suffix vowels that are closed, there exists a [closed] such that the left edge of that occurrence of [closed] and the left edge of the stressed vowel coincide (Parkinson 1996) d. EXTEND [feature] The autosegment [feature] must be associated to all available vocalic positions within a word. (Kaun 1995) e. Combination of EXPRESSION[F] and *[F,Consonant] The feature [F] must be expressed on every element in an F-domain. And [F] cannot be realized on consonants. (Cole & Kisseberth 1997) This dissertation will (1) show definitively that underlying glides do exist, despite the previous doubts, (2) provide a cross-linguistic representation for them based on data from nine unrelated languages, (3) list criteria for determining whether a language has underlying glides, and (4) present a typological representation of how languages map underlying vocoids (vowels and glides) onto surface vocoids. 1.2 Minimal Pairs Contrasts between glides and vowels are inherently more difficult to grasp than other allophonic alternations for they do not form clear minimal pairs like two consonants ([ta] vs. [da]) or two vowels ([ta] vs. [ti]) would. The fact that glides and vowels have different positions in the syllable makes it impossible to find a pair where the two actually contrast in a true minimal pair. If a glide and a vowel appear in a near minimal pair such as Spanish vac[i.a]ba 's/he emptied' versus vic[ja]ba 's/he vitiated', there are actually at least two differences. First, one form contains a vowel and the other a glide. Second, the first word has four syllables, while the second has only three. Related to this is the presence/absence of vowel hiatus and possibly a different number of moras. The existence of near minimal pairs such as the Spanish example have two possible analyses, one of which does not include underlying glides. Consider the hypothetical pair [ta.i] and [taj]. Clearly there is some difference in the underlying representations of these words; otherwise they would not surface differently. The question is what the correct underlying distinction is. The first analysis rests on an underlying distinction between /j/ and /i/, hence we have /taj/ [taj] and /tai/ [ta.i]. This 'glide analysis' makes several predictions. First, we expect hiatus to be abundant on the surface since hiatus is allowed with a high vowel. Furthermore, since the distinction is between an underlying glide and an underlying vowel, they are featurally distinct and are expected to behave differently. We might also expect to find epenthesis in clusters

5 that include glides, if for example the language had consonant-only suffixes (e.g. /taj-t/ [taj-i-t], but /tai-t/ []). The second analysis treats the form with hiatus as anomalous. In this case, the underlying representations are /tai/ [taj] and /tai=N/ [ta.i]. Here, the notation /i=N/ symbolizes a vowel that is pre-linked to a syllable nucleus (Levin 1985). In Chapter 11, I call this type of language 'vocalic'. Featurally, /i/ and /i=N/ are identical because they are both underlyingly the same vowel; the difference is simply that the latter vowel must surface in the syllable nucleus. Since some high vowels are lexically marked, we might expect them to have a more marginal status, as is indeed the cases in Spanish (Harris & Kaisse 1999, and see Chapter 11). The normal situation in this language is for /i/ to surface as either a vowel or a glide depending on the environment and the parameters of the language, witness our form /tai/ [taj]. Therefore, we expect that hiatus will not be rampant on the surface. The most telling evidence for this analysis would be some featural process that treats both /i/ and /i=N/ the same. In the Pasiego dialect of Spanish, high stressed vowels cause harmonic raising of preceding mid vowels (Penny 1969a, Penny 1969b, Hualde 1989). For example, the infinitive [bebér] 'to drink' has a mid vowel in the first syllable, but in [bibís] 'you (pl) drink (indic)', the same vowel is raised by the high stressed vowel. Raising is also triggered by nuclear glides, as in [amfijón] 'infection' (c.f. [amfestár] 'to infect') (Kaisse & Levi 2004). Finally, lexically nuclear vowels also trigger harmony, as in [bibí.a] 'he was drinking (indic)' (Penny 1969b) or [kuxirí.a] 'I would take' (Hualde 1989). The fact that glides, plain high vowels, and lexically marked high vowels participate in vowel harmony is telling. Since they behave together with respect to harmony, we do not want to assume the 'glide analysis' above that contrasts /j/ and /i/. Instead, the 'vocalic analysis' is better for Spanish in that it unifies the analysis of vowel harmony to a single feature bundle /i/ regardless of whether it is lexically marked to occur in the nucleus. Spanish will be further discussed in Chapter 11. When faced with a near-minimal pair such as [taj] and [ta.i], it is necessary to look further into the language to determine the true representation of this contrast. Processes that target features and either group [j] and [i] together or differentiate between them help determine which analysis--'glide' or 'vocalic'--is better suited to the language. Similarly, finding the basic pattern of the language--hiatus or gliding--can also clarify which solution is better. 1.3 Overview of the languages used The nine languages with underlying glides that I will discuss are given in (3). The language classifications are based on the family divisions provided by Ethnologue.4 The languages that are dealt with in this dissertation form a diverse group, none coming from the same family.


6 (3) Languages with underlying glides Language Family Language Afro-Asiatic Imdlawn Tashlhiyt Berber Algic Cree Altaic Turkish Austronesian Sundanese Hokan Karuk Indo-European Pashto Na-Dene Tahltan Niger-Congo Pulaar Penutian Yawelmani In a cross-linguistic study designed to consider the distribution of a particular linguistic property, diversity is essential. In collecting a sample of languages for such a project, it is important to cull the set of possible languages and arrive at a sample that is representative of the languages that exist. Several methods for selecting such a language sample are discussed in Rijkhoff & Bakker 1998. The difference between a distributional study and the one outlined here lies in the type of question that is asked. In the former, the question asks about the distribution of a particular phenomenon (e.g. how common is voicing among obstruents). It is therefore necessary to consider disparate languages in order to arrive at a percentage which accurately represents the number and distribution of languages which exhibit a given property. In the current study, the question does not address the quantity or statistical frequency of underlying glides. In this dissertation, the question is the existence, not the prevalence, of a single phenomenon. Thus, all of the languages that are included have been selected explicitly because they fit a particular requirement, namely that of having underlying glides. For this reason, no sampling technique was used. Nonetheless, in searching for languages, I did try to include a diverse set in order to eliminate the possibility that underlying glides are a genetic or areal characteristic. Though the languages with underlying glides in this dissertation are well distributed across families, I do not believe that underlying glides are common or pervasive. Finding languages that satisfied the strict criteria I laid out was surprisingly difficult. It is my belief that in most cases, the evidence needed to posit underlying glides is not found; thus the language learner instead assumes that the surface glide is the result of universal syllabification tendencies. When both an analysis that assumes underlying glides and one that assumes none are possible, I assume that the latter is optimal. No extra segment need be stipulated in the inventory. 1.4 Criteria When considering the vocoid system in a language, the first characteristic to notice is the distribution of surface vocoids. In most systems, vowels and glides surface in a regular

7 manner, depending on the syllabification and/or the sonority of the surrounding environment. However, some languages appear to contrast two different sets of high vocoids. When there is an 'over-abundance' of surface vowels or glides, then some extra stipulation must be made. An 'over-abundance' of vowels (i.e. vowels surfacing in a location where glides are expected) signals lexical marking of the vowel as nuclear (Levin 1985). When there is an 'over-abundance' of glides (i.e. glides surfacing instead of the predicted vowels), an analysis with an underlying glide should be considered.5 Simply having a contrastive set of vocoids is not sufficient to assume a representation with underlying glides. In other words, lack of predictability does not in and of itself imply the existence of an underlying glide. For example, Spanish shows near minimal pairs vac[i.a]ba and vic[ja]ba (see Chapter 11), but does not have underlying glides. Researchers should also be cautioned when assuming that predictability is absent. For example, Waksler 1990 assumed underlying glides in Lenakel because of the existence of near minimal pairs, but Rosenthall (1994, 1997) points out that including metrical considerations yields a predictable glide/vowel distribution. Furthermore, for a language to be included in this study it had to show positive evidence for underlying glides. For example, if a language had surface glides but no glide/vowel alternations, I did not include it in the study. The types of processes and the sort of evidence that can motivate the existence of underlying glides are varied. In determining if a language has underlying glides, it is of course best to find converging evidence from several factors. The criteria for determining the phonemic status of glides in this dissertation are (1) transparency to vowel harmony (Karuk, Yawelmani, Pulaar, Turkish), (2) triggering of consonant harmony (Tahltan), (3) necessity of vowel epenthesis in unsyllabifiable clusters which include glides (IT Berber, Yawelmani), (4) syllabification (IT Berber, Karuk, Yawelmani), (5) reverse sonority clusters (Pashto, Cree), (6) blockage of nasal harmony (Sundanese), (7) consonant gradation (Pulaar), and (8) the presence of more than one series of glides (e.g. plain, voiceless, and glottalized) (Yawelmani).


The expectation of a vowel in a particular environment can be the result of syllabification (e.g. /atai/ [ataj]), but can also include language requirements on stress and word-minimality (e.g. two syllable minimality /tai/ [ta.i]).

8 (4) Evidence for underlying glides Yawelmani j Sundanese

VH CH Epenthesis Syllabification Reverse Sonority Clusters Nasal Harmony Consonant gradation Series The criteria I employed for including glides in the underlying inventory were strict. One type of evidence that I did not consider sufficient was the tri-consonantal roots of Arabic. It is possible to analyze the tri-consonantal roots as containing any segment with sonority less than or equal to high vowels. There may be other evidence from Arabic that merits including glides in the phoneme inventory, but the trilateral roots are not enough on their own. 1.5 Typology Languages differ as to how high vocoids are manifested on the surface. An underlying high vocoid can surface (a) with both a syllabic and non-syllabic parse ("both"), (b) as only itself, where "itself" refers to either a glide or a vowel ("itself"), or (c) as zero ("neither").6 Cases where the vocoid surfaces as "zero" are equivalent to assuming that the particular vocoid is not present in the underlying inventory. If it maps to zero, what evidence is there of its existence in the first place? (5) i


"Both" i/j j i

i j

"Itself" or j i j

"Neither" i/j i

There are cases where an underlying vocoid surfaces as neither a vowel nor a glide, as in onsets in Spanish /creio/ [cre.o] 's/he believed' (Harris & Kaisse 1999) or is deleted in certain contexts. Cases of this sort will not figure into the analysis.








9 There are therefore three different possible mappings from an underlying to a surface vocoid. Since there are two types of underlying vocoids (underlying vowels and glides), there are nine possible mappings from underlying to surface vocoids. The number nine is derived from the fact that an underlying vowel has three possible ways to surface times the fact that an underlying glide has three ways in which it can surface. For example, an underlying vowel can surface as "both" while the underlying glide surfaces as "both" (Type VIII), "itself" (Type VI), and "neither" (Type III). The full typology is presented in (6), where a high vowel is represented by V and a glide by G. The ninth possibility, where both underlying vocoids map to zero, is not included. This last example would be a language which had neither surface high vowels nor surface glides. Though such a language is possible, it has not been included since the primary concern is to consider languages that do have high vowels and glides. (6) Typology of mappings -III-I-II/G/ /V/ /V/ [V] [G] [V] [G]

-IV/G/ [V]

-V/V/ /G/

-VI/V/ /G/ [V] [G]

-VII/V/ /G/ [V] [G]

-VIII/V/ /G/ [V] [G]

[G] [V] [G]

This dissertation will not discuss Types I and III since they do not have underlying glides. A language exemplifying each of the types in (6) that include an underlying glide is given in (7). It is not surprising that Type V is the most common. This is the type where there is no overlap between the underlying and surface vowels and glides. Perhaps such a system is more stable, allowing less confusion between the two underlying vocoids. In Type II, there is no high vowel in the vowel inventory, but there is a corresponding surface glide. In Type VI, the standard vowel/glide alternation from an underlying vowel occurs, but the language also contains an underlying glide which is never syllabic. The representation in (6) where the realization of the derived glide and the underlying glide is neutralized is perhaps misleading. The surface glides sound the same, but can actually be distinguished by the way they pattern in the phonology. Type VIII is a symmetric case where both the underlying glide and vowel show alternations. (7) Examples of the typology Type II Deg Xinag Type IV Type V Turkish, Cree, Tahltan, Yawelmani, (Pulaar), (Sundanese) Type VI Karuk, Pashto, (Pulaar), (Sundanese) Type VII Type VIII Imdlawn Tashlhiyt Berber

10 Thus far, Types IV and VII are unattested in my search. In fact, these types are the two in which an underlying glide surfaces as "both", while not being balanced by a "both" alternation in the corresponding vowel. Type IV could be lacking from my language sample due to the relative paucity of languages which lack a high vowel.7 Since languages with an underlying glide are also rare, the probability of finding a language with both of these low-probability properties is even lower. In order to establish a Type IV classification, the language would also need some featural process (e.g. vowel harmony) which could distinguish it from Type III. The absence of Type VII could be due to the existence of an alternative analysis. A language in which some vocoids exhibit a glide/vowel alternation on the surface, while another set of vocoids always surfaces as a vowel, is subject to the analysis established in Levin 1985 and elsewhere (e.g. Harris & Kaisse 1999 for Spanish). Indeed, a language exhibiting an 'over-abundance' of vowels is more likely explained by lexically marking those overly abundant vowels (notated i=N). Exhibiting different behavior among vocoids is not a sufficient reason to assume the existence of underlying glides. The only way to distinguish between an analysis of Type VII and one like (8b) would be if the language also contained a process that featurally distinguished the two vocoids. In (8b), the vocoids are predicted to behave the same with respect to featural processes, whereas those in Type VII are expected to pattern differently. Chapter 11 will deal with languages like those in (8b). (8) a. /?/ [i] /?/ [j] b. /i=N/ /i/ [i] [j]

As a final note, it is not necessary for different high vowels or glides in a single language to be of the same type. In Karuk, for example, the labial vocoids /u/ and /w/ exhibit a Type VI pattern, while /i/ shows a Type III alternation. 1.6 Organization of this dissertation This dissertation is laid out as follows. Chapter 2 looks at previous phonetic and phonological treatments of glides. Chapter 3 presents four possible representations that can distinguish underlying glides from vowels. Chapters 4-9 provide detailed data and analysis of six languages with underlying glides. These six chapters first provide evidence for underlying glides and then compare each of the four models from Chapter 3. Chapter 10 provides data from three other languages with underlying glides. The languages in Chapter 10 do not have any featural processes that could help determine the

According to Maddieson 1984, 8.5% of languages in his sample of 317 lack /i/ and 16.1% lack /u/. When this set is expanded to the more general class of high-front-unrounded vowels (thereby including ), it drops to 1% (3/317). When any high-back-round vowel is included (u and ), only 6.6% (21/317) lack such a vowel.


11 best representation for underlying glides. Chapter 11 looks at six languages that do not have underlying glides. Three are Type III languages and three are like (367b). Finally, Chapter 12 discusses the best representation.


2 Background and history of glides in phonetics and phonology

2.1 Introduction One cannot talk about glides without at once broaching the subject of what distinguishes vowels and consonants. Indeed, when dealing with glides, the problem is whether to determine if they should best be represented as part of the class of consonants or of vowels. This issue is not as simple as it appears. In terms of their featural definition as [-consonantal] segments, glides are generally considered not to be consonants (e.g. Chomsky & Halle 1968). However, the question will be whether we need this feature or definition at all. Regardless of whether we decide that glides should be featurally distinct from high vowels or not, we must still consider if they pattern with consonants or not in terms of the phonology. If they do act like consonants, then the question will be how to represent them theoretically. 2.2 History of glides: phonetics, features, and phonology 2.2.1 Phonetics of glides As with all issues involving glides, there is little consensus as to the proper way to classify glides in the phonetics. It is important to note that when we talk of the phonetic properties of glides, we can only consider output glides. Therefore, this discussion will not provide any insight into the presence or absence of underlying glides. Nonetheless, it is important to see how glides have been treated phonetically in order to understand the history of certain features, and as a result, the history of phonology. Various authors have cited constriction, timing (i.e. duration), amplitude, and even syllable position as the phonetic cues which distinguish glides from vowels. Each one of these criteria will be treated in turn. Constriction as a distinguishing property It has been a long held belief that constriction or manner of articulation is what distinguishes glides from vowels. Indeed, for more than 100 years from Whitney 1865 to Maddieson & Emmory 1985, we see this notion perpetuated. In discussing English, Whitney notes that "we have vowels of so close a position that they verge upon the consonants" and that we "regard them as consonantal and not vocalic" (Whitney 1865:239). Similarly, Sweet provides a definition of the difference between consonantal and vocalic sounds that relies on differences in constriction, and that will provide the basis for the features [consonantal] and [vocalic] in The Sound Pattern of English (SPE) and other such sources. Not only does Sweet consider constriction as relevant for the distinction between vowels and consonants, but he also considers its effect on spontaneous voicing (Sweet 1877/1907). He writes,

13 "A consonant is the result of audible friction, squeezing, or stopping of the breath in some part of the mouth (or occasionally of the throat). The main distinction between vowels and consonants is that while in the former the mouth configuration merely modifies the vocalized breath, which is therefore an essential element of the vowels, in consonants the narrowing or stopping of the oral passage is the foundation of the sound and the state of the glottis is something secondary." (Sweet 1877: 31) Given his definition, he groups both [j] and [w] as consonants. He does not, however, omit the relation between the glides and their corresponding vowels, for he does note that "(w) is a consonantalized (u) with a narrower lip passage" (Sweet 1877: 42). Much later, Maddieson & Emmory 1985 provide actual acoustic evidence that at least some glides are produced with greater constriction than the corresponding high vowels. Their study looks at glides in Amharic, Yoruba, and Zuni. First, the authors find that there are different targets for the glides that parallel the differences among the vowels in these three languages. Because languages can vary in terms of the amount of coarticulation of segments and in terms of which segments are more or less likely to yield to coarticulation, they restrict themselves to the tokens [iji] and [uwu]. If the glides are produced with greater constriction, then (1) the first formant (F1) should be lower for the glides than for the vowels, (2) the second formant (F2) should be lower for [w] than for [u], and (3) the third formant (F3) should be higher for [j] than for [i]. Indeed, all three of the predictions are borne out for the three languages. Thus, in [iji] and [uwu] in these three languages, the glides are produced with greater constriction than the vowels. Though the constriction theory is likely the most prevalent in the literature, there is some evidence that this is not the whole picture. Straka 1964 shows that glides are not always produced with greater constriction than their corresponding vowels. His goal in this article is to show that glides are consonants. He draws on a variety of phonetic sources to argue for this stance. First, he considers constriction. Straka shows on various palatograms that there is not a stable difference between the aperture of glides and that of vowels.8 He writes that it is not possible to trace a line in the mouth and say that less than X is a vowel and more is a consonant. When [j] is uttered between two high vowels, as in [iji], it is produced with greater constriction than the surrounding vowels (as was also found in Maddieson & Emmory 1985). However, when next to a non-high vowel, as in [aja], it is produced with less constriction than [i] between two consonants. He also argues that when we compare [z, , j], we actually find more constriction for the glide than for the fricatives.9 He also cites radiographs that show that the tongue is more

The data in the paper come from French, Estonian, Polish, and Albanian, but he also cites other languages where the same results have been found (German, English, Czech, Slovak, and Albanian). 9 Straka uses tracings of palatograms. He is not explicit in whether 'more constriction' means added tongue bracing or the size of the actual opening.


14 elevated in [j] than in [s, , x]. He therefore concludes that neither the size of the constriction (how much of the oral cavity is open), nor the height of the tongue can determine the difference between glides, fricatives, and vowels. Second, Straka examines the behavior of glides and vowels when expressed under increased effort in a variety of languages including French, Estonian, and Polish. When vowels are produced with greater effort, they show less constriction than in normal circumstances. Consonants, on the other hand, show greater constriction when produced under effort. Offering continuing support to his claim that glides are truly consonants, he shows that under effort, glides pattern with consonants in being produced with greater constriction. Third, Straka considers the amount of oral airflow using kymographs that indicate the amount of airflow in the mouth. He argues again that the glides pattern with the consonants to the exclusion of vowels. He also writes that glides have a lower amplitude of voicing and generally lower laryngeal frequency than do vowels. Based on these three characteristics, Straka concludes that glides are in fact consonants, and are not vowels and are also not a separate third category. The second two points above do lend credence to the claim that glides are consonants. That is, they pattern with consonants with respect to articulation under effort and to the amount of airflow. The first point, however, seems to highlight the variable nature of glides in that their constriction can be more or less than their corresponding vowel. This could potentially be viewed as a contradiction to the notion that glides are consonants. Duration as a distinguishing property Second only to constriction, duration is the most cited difference between glides and vowels. Jones 1918 considered glides as different from vowels partially because of their "gliding nature." Liberman et al. 1956 and O'Connor et al. 1957 show that the duration of the transition changes the perception of the segment. Liberman et al. 1956 show that as the duration of the transition increases from 10 to 300 ms, listeners perceive a stop, a glide, and a vowel. O'Connor et al. 1957 show that some steady state portion is needed to perceive a glide in initial position. With no steady state, listeners perceive a stop. When the duration of the steady state is increased to 40+ ms, listeners perceive it as a vowel. Catford 1977, who seems to be unaware of O'Connor, believes that the duration and timing of glides differs from those of vowels. He believes that glides do not have any steady state portion, but that we nonetheless perceive them as segments. He argues that the reason that glides are perceived as segments at a place in the signal where there is no steady state is because listeners are used to listening to the "onglide" (i.e. transition into the segment) and the "offglide" (i.e. transition out of the segment). If the steady state portion is reduced to zero, Catford claims that we still perceive a segment because of the transitions. He believes that vowels and glides are identical in the type of stricture, but that they can be differentiated by whether or not they have a "noticeable" duration.

15 I should point out, however, that the duration argument runs into problems when faced with languages that allow geminate glides. In languages that have sequences such as [ajja], clearly the glide is not simply a transition into and out of a constriction. Amplitude as a distinguishing property The notion that prominence or amplitude distinguishes vowels and glides is promoted by Jones 1918. He writes that "[glides] have the organic positions of close vowels...but they are uttered with very little breath force as compared with the normally pronounced vowels which adjoin them in connected speech." Like Sweet, he considers glides to be consonantal because of their "small inherent sonority", their short duration, and their "gliding nature".10 In a sequence of GV, the stress (i.e. prominence) is on the vowel. If it is shifted to the glide, then it "is at once recognized as [an] independent element" (Sweet 1890). Position in the syllable as a distinguishing property The last often cited difference between glides and vowels is their position in the syllable (Stetson 1928, Grammont 1933, Troubetzkoy 1939, Pike 1943). Indeed, this notion will function as a segue into the discussion of phonology. Though this seems like a phonological issue, it is first treated in phonetic descriptions. Pike 1943 argues against the constriction approach, saying that there is not a point at which one can say more constriction is a consonant and less is a vowel. He argues that what Jones 1918 meant when he said that glides are consonants is that the difference is based on how the segments behave in the syllable. Jakobson, Fant, & Halle 1952 do not give phonemic status to glides in English. Instead, they assume that syllable position determines the difference. Thus, woo [wu] is represented as /u'uu/. As we have seen, it is not uncommon for authors to use a combination of these characteristics to distinguish glides from vowels. Stetson 1928 believes that the difference between a vowel and a consonant is based on their function in the syllable. He also writes that when vowels are "forced to the limit [i.e. constriction--SVL] and quickly released [i.e. duration--SVL]" that they become consonants (Stetson 1928: 91). In his view, the glides are considered consonantal. Summary In this section, we have seen that the phonetic treatment of glides varies. Constriction, amplitude, duration, and syllable position have all been used as the cues which distinguish glides from vowels. Some authors have argued explicitly that glides are consonants while others merely imply that this is the case. When we talk of the phonetics, we are only able to look at output glides to determine if they look and behave similarly to vowels or to consonants. Those who believe that the difference lies in their


"Their rapid gliding nature...combined with the use of rather weak force of exhalation...renders [glides] consonantal." (Jones 1918)

16 position in the syllable have taken a step beyond the phonetics and realized that glides and vowels are related and that their difference can also be represented in the phonology. In the next section we will turn to the notion of distinctive features and the impact this has had on the phonology of glides and vowels. 2.2.2 Features and phonology Glides are inherently tied to the notion of syllabicity and the difference between consonants and vowels. Given the various acoustic and articulatory characteristics of glides, as well as consonants and vowels, several definitions have been proposed to divide sounds in to natural classes in the phonology. For the question of how to partition the set of consonants and vowels, the features consonantal, vocalic, and syllabic have been most widely used. As early as Sweet 1877, we have the notion that consonants must be made either with some amount of closure or enough of a constriction to cause friction. Furthermore, in vowels the voicing is considered primary, while in consonants it is not. For Jones 1918, obstruction and frication are also the relevant features for consonants. Even prior to these, Whitney 1865 looked at constriction as a key to the consonantal nature of sounds and considered glides to be consonants. Early generative phonology Within traditional accounts of distinctive features, [consonantal] and [vocalic] have been used to divide sounds into four classes: non-liquid consonants, liquids, glides, and vowels. Jakobson, Fant, & Halle 1952 lay out the following definitions based on acoustics and articulation to distinguish consonantal and vocalic sounds. (9) (Jakobson, Fant, & Halle 1952) a. Vocalic/Non-vocalic: Phonemes possessing the vocalic feature have a single period ("voice") source whose onset is not abrupt. b. Consonant/Non-consonant: phonemes possessing the consonantal feature are acoustically characterized by the presence of zeros that affect the entire spectrum.

In terms of production, "vowels have no obstructive barrier along the median line of the mouth cavity, whereas consonants have a barrier sufficient to produce either complete occlusion or turbulent noise source" (Jakobson, Fant, & Halle 1952). Using these two features of acoustics and articulation, they can characterize the four classes in (10) (10) Descriptions of segments (Jakobson, Fant, & Halle 1952) a. Vowels [+vocalic, -consonantal]: phonemes possessing the vocalic feature and having no consonantal feature. b. Consonants [-vocalic, +consonantal]: phonemes possessing the consonantal feature and having no vocalic feature.

17 c. Liquids [+vocalic, +consonantal]: the laterals and the various intermittent rsounds, have the vocalic as well as the consonantal feature: like consonants, they have significant zeros in their spectrum envelope. The formant structure of liquids is basically similar to that of the vowels, but the configuration of their first three formants usually differs from that of any vowels. The overall intensity of the liquids is considerably lower than that of the vowels. d. Glides [-vocalic, -consonantal]: the glides are distinguished from the vowels in that they have either a nonharmonic source (h in English) or a transient onset of the source ( in English). They are distinct from the consonants in that they have no significant zeros in their spectra. Jakobson, Fant, & Halle 1952 do not consider the glides [j] and [w] to be phonemic (at least in English). In their view these glides are merely desyllabified high vowels. Jakobson & Halle 1956 provide similar definitions and classifications to those in (10). Chomsky & Halle 1968 draw on these same definitions. Early in SPE, they use the features [±consonantal] and [±vocalic] to distinguish these same four groups, though now the glides [j] and [w] are included in the class of glides. They have the class of "true" consonants which includes the obstruents and the nasals [+consonantal, -vocalic], the liquids [+consonantal, +vocalic], the glides (j, w, , h) [-consonantal, -vocalic], and the vowels [-consonantal, +vocalic] (Chomsky & Halle 1968: 68). Here, vocalic sounds are those "produced in the oral cavity in which the most radical constriction does not exceed that found in the high vowels [i] and [u] and with vocal cords that are positioned so as to allow spontaneous voicing." Consonantal sounds, on the other hand, are "produced with a radical obstruction in the midsagittal region of the vocal tract" whose constriction must be as narrow as that "found in fricative consonants." Later in SPE, they replace the feature [vocalic] with the two features [syllabic] and [sonorant] (p. 354). Under such a representation, glides are [-consonantal, -syllabic, +sonorant]. The difference now between vowels and glides is in the feature [syllabic] instead of [vocalic]. This feature differs from the others in not having an articulatory or acoustic basis. Instead, it is a relative feature, which "would characterize all segments constituting a syllable peak." (SPE: 354). Syllable-based phonology With the reintroduction of the syllable in phonology (Kahn 1976) came the push to eliminate the feature [syllabic] from the theory. Kaye & Lowenstamm 1984 argue against [syllabic], saying that syllabification processes determine syllable position. They further argue that the specification of glides as [-consonantal, -vocalic] actually allows no place in the syllable to be available to them. Thus in their theory, glides and vowels are given the same features ([-consonantal, +vocalic]) and are supposed to never be distinct segments. Kaye & Lowenstamm represent them as /I/ and /U/, stating that syllabification is what determines whether the glide or the vowel surfaces.

18 In theories that make use of the syllable, it is structure and not features that distinguishes glides from vowels. Clements & Keyser 1983 introduce the CV tier between the segmental tier and the syllable tier. In their theory, V stands for a syllable peak, and C a non-peak. In this way, they eliminate the need for the feature [syllabic], stating that the difference between vowels and glides is whether they are dominated by a C or a V node. Here, vowels and glides are not distinct on the featural level, and are only distinguished on the CV-tier. CV theory can distinguish the outputs in (11). (11) V | i = [i] VV \/ i = [i] C | i = [j] CC \/ i = [j]

One of the problems with CV theory is that it can distinguish more structures than are actually necessary (Waksler 1990). In (12), we see two different representations for [ji]. (12) C V | | i i = [ji] C V \ / i = [ji]

Another problem with CV theory is that it allows a vowel to be linked to consecutive V and C slots, as in the compensatory lengthening in (13), but it does not explain why there are no cases where a vowel links to consecutive C and V slots (Levin 1985). (13) a. C V C b. *C V C | | | | d a = [da] 'mountain' (Turkish) i p = [ip] Perhaps the biggest problem with CV theory is that it required largely redundant information (Levin 1985). Whether a segment surfaces in a peak or non-peak position within the syllable is largely predictable and need not be encoded on the CV-tier. CV theory also contains the skeletal changing rules (e.g. C V or V C) which Waksler 1990 argues obscure the fact that these changes occur to satisfy syllable well-formedness. In a move to eliminate some of this redundancy, Steriade 1984 uses C, V, and X on the skeletal tier. Noting the predictability of the glide/vowel alternation in Latin, Steriade 1984 amends the CV tier to include an X for the vocoids that alternate. She represents the high vocoids in Latin as [-consonantal, + high, 0 syllabic] on the segmental level and as X on the CV tier. This representation allows the high vocoids to alternate between peak and non-peak elements in the syllable without requiring specific rules that change skeletal slots (e.g. C V). Under Steriade's proposal, there could be structures as in (14).

19 (14) a. / \ O R | | X V | | i a = [ja] b. O | C | t / \ R | X | i = [ti]

Noting the redundancy of CV, Levin 1985 eliminates CV in favor of X for all skeletal positions. Whereas in CV theory it was still possible to distinguish glides from vowels on the CV tier, in X-theory an additional level is needed. Syllable nuclei are Xslots linked to an N-node as shown in (15). Non-peak elements are now linked either to N" (onsets) or to N' (codas). The difference between vowels and glides is now represented on a higher level. Vowels' X-slots are linked to N and are the heads of syllables, while glides' X-slots are not. For the majority of output vowels and glides, it is "the system of distinctive features which is used to define categories of obligatory, possible, and impossible syllable heads on a language specific basis" (Levin 1985). (15) a. N"= N' N i a p = [jap] t b. N"= N' N i p = [tip]

Peak (N) vs. non-peak (N" or N') differs from [±syllabic] in four ways (Levin 1985). First, [+syllabic] is never distinctive on the surface because it is always a property of the syllable nucleus. Second, X-theory predicts that any segment can be a head. Third, the need for a three-way distinction [±/0syllabic] is unnecessary because it is simply a question of whether a segment is a head or non-head. Fourth, syllabicity is a syntagmatic property, unlike most other features. Levin 1985 does, however, remark that not all cases of vowels and glides can be predicted by these rules alone. She writes that the vocoids "must often be specified underlyingly as [+syllabic], but never as [-syllabic]." This is because heads "can be marked in the underlying lexical representations, while the property of being a non-head cannot." In this theory, there are three ways in which a segment can surface as a head. The first is via a Redundancy Rule. In these cases, a segment, /a/ for example, always surfaces as a head regardless of the surrounding segments. The second is via a Phonological Rule. In these cases, the syllabification of segments is determined by the environment. Finally, a segment can surface as a head if it is Lexically Marked as such.

20 This lexical pre-linking, Levin argues, can only exist at morpheme edges, presumably because this is the only place where glide/vowel alternations occur. In X-theory, glides and vowels are featurally nondistinct and the difference is whether the vocoid is in a head or non-head position. This lexical pre-linking is used by Guerssel 1986, Roca 1997, and Harris & Kaisse 1999.11 Instead of using syllabic structure to distinguish alternating vocoids from those that always surface as vowels, Deligiorgis 1988/1993 reverts to the feature [consonantal]. She argues that [-consonantal] "true" vowels always surface as syllabic or nuclear. For vocoids which alternate between vowels and glides on the surface, she assumes that they are [0consonantal]. Though the logical possibility of allowing [+consonantal] vocoids exists in this system, she argues that there are not [+consonantal] vocoids (=glides) and espouses the view that something like underlying glides do not exist. In order to unify weight-related phenomena, a moraic representation was developed in which skeletal slots were abandoned. Hayes 1989 assumes a possible threeway distinction in high vocoids. (16) µ | i = [i] µ µ \ / i = [i]

i = [j]

As in CV-theory and X-theory, glides and vowels are featurally identical. Hayes argues that this moraic theory is better than the previous ones in that it does not require any syllable structure in the underlying representation. However, in the moraic theory promoted in Hayes 1989, there is no way to distinguish a geminate glide from a short vowel. That is, the representation in (17) has two possible outputs (Waksler 1990).12 (17) µ | i = [i] or [j]


Guerssel's analysis differs slightly from Levin's stance. In his analysis of Ait Seghrouchen Berber, Guerssel examines the differences between glides and vowels. He represents the vocoids that have different surface alternants (e.g. [i-ru] 'he cried' [y-ari] 'he writes') as regular unlinked X slots. For those vocoids that always surface as vowels, he uses the lexical pre-linking. He uses the pre-linking both for cases such as /arba-u/ [arba-j-u] 'this boy' *[arbaw] in which gliding of the final /u/ is not allowed and for cases such as the first vowel in /tusi/ [tusi] 'carry'. In a slight reanalysis of these facts, Deligiorgis 1988 uses the lexical pre-linking for only those cases such as [arba-j-u] and not for [tusi]. See Chapter 11 for more discussion of this example. 12 In Hayes 1989, geminates differ from single consonants in being underlying associated with a mora. We have the representation in (i). (i) µ | n = [n] n = [n]

21 Finally, Optimality Theory has been used to explain the alternations of vowels and glides (Rosenthall 1994). By reordering constraints such as those in (18), Rosenthall is able to explain the resolution of vowel-vowel sequences in a variety of languages. (18) a. b. c. d. e. f. Sample of the constraints used in Rosenthall 1994 ONSET: Syllables must have onsets. NOLONGVOWELS: Two moras cannot link to a single vocalic root node. NODIPTH: Tautosyllabic moras cannot link to two separate vocalic root nodes. V-MORA: For every vocalic root node rti, there is a mora µi. SECART: Consonants should not have a secondary articulation. PARSE-µ: Moras shall not be deleted (essentially MAXµ)

For example, the input /tia/ might yield [ta], [ta], [tja], [tja], [ta], [ta], or [ti.a] depending on the ranking of the constraints. That is, some languages will preserve the mora count when the high vocoid is not in a nuclear or moraic position, while others will not. Some languages will allow secondary articulation, while others will choose to allow an onset cluster or rising diphthong. Some languages will simply delete the high vocoid, while others will allow hiatus. Rosenthall's basic claim is that glides are derived from vowels. He writes, "glides are simply underlying high vowels that best satisfy the constraints by being parsed nonmoraically" (pg. 8). Roca 1997 also uses an OT account to characterize the glide/vowel alternations in Spanish and concludes that glides as distinct segments do not exist in Spanish and possibly universally. Roca writes, "In the absence of this feature [i.e. [±syllabic]--SVL], the label "glide" becomes redundant, indeed undesirable, since it misleadingly points to a segmental category deprived of existence." Glides as [+consonantal] Though these three theories of the syllable (CV, X, µ) adhere to the basic assumption that glides are featurally identical to high vowels, that they are derived from the same source, and that principles of syllable well-formedness are what determine the output of an underlying high vocoid, several of these same authors have needed to resort to features in order to distinguish vowels from glides in some languages. For languages that exhibit an over-abundance of vowels (i.e. vowels in places where glides are expected based on normal syllabification), Levin's lexical head marking is sufficient to prevent these segments from surfacing as glides. Deligiorgis 1988 uses a similar approach, but targets the feature [consonantal] to account for such alternations. She considers segments that never surface in a nuclear position to be [+consonantal]. Those that always surface in a

22 nuclear position are [-consonantal]. And those segments that vary are [0consonantal].13 She considers "true vowels" to be those segments that are [-consonantal] and "non-true vowels" to be those that are [0consonantal]. Though she does not believe that underlying glides exist, within a framework such as hers that uses a three-way distinction of [consonantal], it would be simple to use this feature to distinguish underlying glides ([+consonantal]) from "true vowels" ([-consonantal]) and from alternating vowels ([0consonantal]). Hyman 1985 proposes a system where all glides are [+consonantal] on the surface. He theorizes that glides are either underlyingly [+consonantal] or [-consonantal], but become [+consonantal] later in the derivation. Under such an approach, vocoids can alternate between [-consonantal] and [+consonantal]. Hyman 1985 and Hayes 1989 use the tri-consonantal Semitic roots (such as /jjj/ 'to write the letter y' and /jsr/ 'to play with a dreydl') to argue for [+consonantal] glides. Hayes writes, "representing the Semitic glides as [+consonantal] can solve the long-standing problem of how to indicate that they are to be mapped onto syllable-peripheral rather than nuclear positions" (1989:302). Recall that in the moraic theory of Hayes 1989, there was no way to distinguish short vowels from geminate glides (see (17)). In order to account for the data in Central Alaskan Yupik where the language must contrast geminate glides, singleton glides, and long and short vowels, Hayes resorts to the feature [consonantal] to avoid the problem posed in (17). (19) [anjalijuluni] 'he was excellent at making boats' (Woodbury 1987: 687) [anjalijjuluni] 'he was EXCELLENT at making boats'

Waklser 1990 too relies on [consonantal] for a three-way distinction in Lenakel. She shows that X-theory and µ-theory have shortcomings in that they cannot account for a language in which there are predictably alternating vocoids, true vowels, and true glides. X-theory can only distinguish two types of vocoids, those that are lexically linked to a head and those that are not. Similarly, µ-theory has problems because it can only exhibit a two-way contrast, those segments that are underlyingly linked to a mora and those that are not. Waksler shows that in Lenakel, a three-way distinction is needed. In order to accommodate this data, she simply utilizes [±,0consonantal]. Rosenthall 1994 must also appeal to this feature. He first argues against the validity of the Lenakel data presented in Waksler 1990. He does use [consonantal], however, to distinguish various high vocoids in languages such as Berber. Rosenthall also argues against what he calls the "protected vowel" method (i.e. lexically pre-linked to N), saying that it is the ordering of the constraints which dictates how the vocoid will

Deligiorgis 1988 treats syllabic consonants, such as in 'button' or 'little', as actually having a schwa in the syllable. This way, she can allow the nasal and the lateral to still be [+consonantal] and never surface as a nuclear head.


23 surface. He does not, however, mention that he has traded the lexical pre-linking for a distinction within the feature [consonantal]. Feature geometric glides An alternative to using [consonantal] is to instead use a feature geometry (Clements 1990, Clements & Hume 1995). In fact, Hume & Odden 1996 suggest that [consonantal] be eliminated altogether in favor of two different representations.14 In their particular model, vowels are differentiated from glides in having V-place features instead of Cplace features. (20) vowel root | C-place ... V-place | [coronal] glide root | C-place | [coronal]

They argue that calling glides [+consonantal] goes against the very definition of that feature. They do not consider glides to be consonantal vowels and write, "we have a hard time imagining what a [+consonantal] vowel would be." They find that the feature can be dispensed with when explaining lenition, fortition, and natural classes (but see Kaisse 1992 for an alternative analysis). This particular feature geometry, as well as another version, will be described in more detail in the next chapter. 2.3 Summary This chapter has considered different ideas about what distinguishes vowels from glides in the phonetics. The most common articulatory and acoustic properties that are cited include constriction, duration, and amplitude. Constriction is the most cited characteristic that distinguishes vowels from glides. Nonetheless, it is clear that there is no line along the palate where we can definitively say that less constriction is a vowel and more constriction is a glide. This constriction idea finds its way into the phonology by way of the feature [consonantal] which is defined by the amount of constriction. There has been debate back and forth about the number of features that should be used to divide consonants, vowels, and glides. Early on, phonologists used both [±vocalic] and [±consonantal] to develop a four-way split in the phonemes. Once syllabification was considered to be predictable, [vocalic] (or [syllabic]) was abandoned and only [consonantal] remained. CV-theory, X-slot, moraic theory, and OT have all


Also see Herman 1994 and Hume 1995.

24 needed to resort to the feature [consonantal] when trying to account for all cases of glide/vowel alternations or non-alternations. Therefore, it is clear that not all syllabification of high vocoids is purely predictable. A side debate considered a threeway distinction of the feature to include segments that were [+consonantal], [-consonantal], and [0consonantal].


3 Possible Representations for Underlying Glides

3.1 Overview This chapter describes four approaches to representing underlying glides. The choice of these four methods does not preclude the possibility of other representations. However, these four possibilities mark the most commonly used methods in the current literature for distinguishing underlying glides from underlying vowels. The four representations are: lexical marking, the use of the feature [±consonantal], and two alternatives within feature geometry (Vowel-Place Theory and Revised Articulator Theory). Before beginning the discussion of the four representations, a note on terminology is necessary. In the following sections, I will often use the terms underlying vowel and underlying glide. The term underlying vowel will refer to an underlying segment with vocalic/vowel features. These underlying vowels can surface as either a vowel or a glide depending on the parameter settings of the language and on the environment (see Chapter 1). Some authors prefer to use the term vocoid to refer to these segments in order to highlight their alternating behavior. The reason for referring to these segments as vowels is that they have similar features to other vowels. Calling them underlying vowels also distinguishes them from underlying glides. The term derived glide will refer to a glide whose source is an underlying vowel. The term underlying glide will be reserved for true underlying glides. In other words, these are glides that are not simply derived from underlying vowels. 3.2 Lexical Marking (LEX) 3.2.1 Overview and Representation The Lexical Marking Representation (LEX) of underlying glides targets only the syllabic or moraic structure of these sounds. In this representation the high vowels have no special lexical marking. They are free to surface as glides or as vowels based on their environment. The lexically marked underlying glides, on the other hand, are marked as unlinkable to a nuclear position. Such a representation would look like the following hypothetical case. I use the symbol * as a shorthand to show that this segment is not allowed to link to a nuclear position within the syllable. (21) Underlying glide /t i p/ * N Underlying vowel /t i p/

The difference between underlying vowels and underlying glides is based solely on possible syllabic affiliation. That is, the underlying vocoid is lexically marked to be

26 banned from a nuclear or syllabic position. A way to think about the representation in LEX is as an anti-N where the underlying glide is prohibited from being nuclear. In a language with a word such as the one in (21), it is likely that the word would not be allowed to surface as *[tjp]. Instead, the language would likely epenthesize a vowel in order to produce a legal output, such as [tjap] or [tajp]. The Lexical Marking in (21) is in essence the reverse of Levin's lexically marked nuclear vowels.15 In fact, this representation of lexically unlinked vocoids is exactly what Levin 1985 argues against. She writes, "vowel/glide alternations must often be specified underlyingly as [+syllabic], but never as [-syllabic]... [Heads] can be marked in underlying lexical representations, while the property of being a non-head cannot" (Levin 1985: 2). However, representations like (21) have certainly been contemplated in other work. The closest representation discussed in the literature that resembles LEX is that of CV-theory discussed in Clements & Keyser 1983. In CV-theory, underlying glides link to a C-slot and are therefore banned from nuclear positions. CV theory is what Levin argues against, stating that only X-slots are needed to represent all segments. Despite Levin's claim that LEX is not a plausible representation, it will be considered alongside the other three approaches in order to determine the best representation of underlying glides cross-linguistically. Extending Levin's notion of lexically marked heads, we now have the following: (22) a. Normal / i / (variable surface representation) b. Levin 1985: N | / i / (always surfaces as a vowel) c. LEX: N * / i / (always surfaces as a glide)

The only difference between (22a) and (22b) is whether or not the underlying vocoid is pre-linked to a nuclear position. Featurally, these two segments are identical. Levin lists both of these segments as [+sonorant, -consonantal]. Similarly, the underlying vocoid in (22c) is featurally non-distinct from the other two. The only difference between the three segments in (22) is in their range of permitted syllabic affiliations. 3.2.2 Predictions Given the representation above, what kinds of predictions does LEX make with respect to underlying glides? In particular, we are interested in what types of phonological processes will distinguish between underlying vowels and underlying glides given the representation of LEX presented in (22c). Because the featural content of underlying vowels and underlying glides is identical, phonological processes that refer to features will not be able to differentiate between them. The only types of phonological processes


Levin's proposals are discussed in Chapter 2.

27 that could distinguish between the segments in (22) are those that refer exclusively to syllable position, for this is the only difference between the three forms in (22). If we assume the representation of LEX, then both the derived and underlying glide will be treated the same by all phonological processes that refer to features. LEX cannot distinguish between the non-syllabic surface reflexes of (22a) and that of (22c). In a language that has both underlying glides like those in (22c) and a glide formation rule that applies to segments like (22a), the surface representations of these two segments will be identical. Once these two segments have been syllabified, their underlying difference is completely neutralized and they are predicted to pattern identically since they are both featurally and structurally identical. From the surface representation, it is not possible to determine whether a surface glide was derived from (22a) or (22c). 3.3 [±consonantal] ([CONS]) 3.3.1 Overview and Representation Perhaps the most frequent method of representing underlying glides is to make use of the feature [consonantal] (Hyman 1985, Hayes 1989, Waksler 1990, Rosenthall 1994, among others). In the representation that utilizes the feature [±consonantal] ([CONS]), underlying vowels are [-consonantal], while underlying glides are [+consonantal]. The use of the feature [consonantal] to distinguish between vowels and glides has not gone without criticism. Hume & Odden 1996 argue against using this feature, stating "we have a hard time imagining what a [+consonantal] vowel would be." Hume 1995 actually argues for the complete elimination of this feature in favor of a feature geometry that uses a segregation between consonant and vowel place features (see §3.4). In terms of the validity of targeting this particular feature, there are also some potential problems. First, the feature [consonantal] is explicitly defined so as to exclude segments such as glides from the definition. The definition from SPE below expressly omits segments that do not have a constriction at least as narrow as that of a fricative. (23) Consonantal sounds are produced with a radical obstruction in the midsagittal region of the vocal least as narrow as a fricative. (Chomsky & Halle 1968)

Using a constriction-based approach to the feature [consonantal] is potentially problematic when we consider the actual constriction of glides cross-linguistically. As discussed in Chapter 2, Straka 1964 shows that glides can be produced with more or less constriction than their corresponding vowels, depending on the vocalic environment and emphasis.16 Thus, using a stricture account (namely [CONS]) to distinguish between vowels and glides either underlyingly or on the surface could be problematic. The very


See Chapter 2 § for discussion of Straka's results.

28 definition of the feature [consonantal] eliminates glides from being [+consonantal] and the palatograms of Straka 1964 show that degree of constriction of glides varies greatly. Despite these potential problems with using [CONS] to differentiate underlying vowels from glides, this approach will be considered alongside the other three. In particular, it is important to look at the ramifications of this approach because of its widespread use throughout the literature. Under [CONS], the only difference between glides and vowels is in the feature [consonantal]. All other features remain the same. For example, a language with underlying /i/ and /j/ might have the pair in (24). (24) a. /i/ -cons +son -back -round +high -low b. /j/ +cons +son -back -round +high -low

If the language allowed glide formation of (24a), its representation would remain unchanged. That is, the [-consonantal] feature of the basic vowel remains, but the segment surfaces phonetically as a glide because of its position in the syllable. (25) a. /tai/ [taj] N" N' t N a i = [taj] -cons +son -back -round +high -low

29 b. /taj/ [taj] N" N' t N a j = [taj] +cons +son -back -round +high -low

A second possible representation of [CONS] is a sort of hybrid approach. A feature geometry could be used instead of the stacked SPE-style features of (24). In this version of [CONS], underlying vowels and glides would have the same feature geometry (e.g. features under V-place in Vowel-Place Theory), but differ in their value of [consonantal]. 3.3.2 Predictions This representation predicts that the only phonological processes that can distinguish between underlying vowels and glides are those that either target (1) the feature [consonantal] or (2) syllable position. Processes that are based on the feature [consonantal] will treat underlying glides differently than derived glides and vowels. On the other hand, processes that consider syllable position will differentiate surface vowels from both underlying and derived glides. Crucially, however, these syllable-based processes would not distinguish derived from underlying glides because they would both be in non-syllable-head positions. Thus, the only processes that could see the difference between underlying glides and derived glides would be those that refer to their consonantal status. If [CONS] is correct, then phonological processes which target features other than [consonantal] should treat these two segments identically. 3.4 Vowel-Place Theory (VPT) 3.4.1 Overview and Representation Vowel-Place Theory is the version of feature geometry that segregates consonant and vowel features into two different tiers (Clements 1989, Clements & Hume 1995). It captures the interactions between vowels and consonants by using the same terminal features to describe both classes of sounds (Unified Feature Theory, Clements 1989). The main departure of Vowel-Place Theory (VPT) from previous models of feature geometry (e.g. Sagey 1986) is the separation of vowel features onto a lower tier than consonant features. This segregation is used to explain, for example, long distance

30 assimilations such as vowel harmony, while simultaneously explaining the lack of long distance consonant interactions.17 Sample hierarchies for consonants and vowels are given in (26) and (27) (slightly modified from Clements & Hume 1995: 292). Consonants with a secondary articulation are represented with their primary articulation under the C-place node and their secondary articulation under the V-place node as in (28). (26) Consonants: root [±sonorant] [+consonantal] Oral Cavity [continuant] C-place [dorsal] [nasal] Laryngeal [spread] [constricted] [voice]

[labial] [coronal] [anterior] (27)


Vowels: root [+sonorant] [-consonantal] Oral Cavity [nasal] [spread] Laryngeal [constricted] [voice]


C-place Vocalic V-place

[labial] [coronal] [-anterior]


Aperture | [open]



Some long distance consonant harmonies do exist (e.g. Shaw 1991 for Tahltan). Consonant harmony or long distance consonant interactions can be dealt with by targeting a single feature (Clements & Hume 1995: 289). For example, coronal harmony can be seen as the spreading of [coronal] and not as spreading higher nodes such as C-place or Oral Cavity. Shaw 1991 argues against the segregation of place features, arguing that it would actually over-predict the quantity and types of consonant harmonies.

31 (28) Consonants with secondary articulation: root [±sonorant] [+consonantal] Oral Cavity [continuant] C-place Vocalic V-place [labial] [coronal] [dorsal] Aperture | [open] [nasal] [spread] Laryngeal [constricted] [voice]

[labial] etc.

[-anterior] [distributed] Within this framework, nodes of the tree can spread from one segment to another segment with the stipulation that there can be no intervening segment that is also specified for that node. For example, the structure in (29) is banned by the No Crossing Constraint (Goldsmith 1979), while the one in (30) is allowed (adapted from Clements & Hume 1995: 268). (29) *C root | Oral Cavity | C-place V root | Oral Cavity | C-place | Vocalic ... C root | Oral Cavity | C-place C root | Oral Cavity | C-place


V root | Oral Cavity | C-place | Vocalic | V-place

V root | Oral Cavity | C-place | Vocalic | V-place

32 In VPT, the same features that characterize consonants also characterize vowels. That is, vowels have the place features [labial], [coronal], or [dorsal] in addition to height features. Furthermore, these features are taken to be privative in both vowels and consonants. The mechanics of VPT follow the general mechanisms used in all feature geometries. Assimilatory processes are explained as spreading a node or a leaf (a single feature) of the tree from one segment to another. Delinking of a node or feature is also possible. In these cases, default or unmarked features surface in the feature tree. For example, delinking explains coda neutralization in Korean. In Korean, the laryngeal node delinks in coda position and default laryngeal features are filled in. Thus, tense, aspirated, and plain stops are all realized as plain in coda. Because of the segregation of place features onto two tiers, VPT allows for two possible representations for glides (Herman 1994, Hume 1995, Levi 2000). Underlying /i/ must have vowel features, and thus the representation in (31). With the two tiers of place features, a different representation for underlying glides is possible, as in (32). (31) Underlying Vowel: /i/ (and output [j]) root [+sonorant] Oral Cavity [+continuant] C-place | Vocalic V-place | [coronal] | [-anterior] (32) Laryngeal | [voice] /u/ (and output [w]) root [+sonorant] Oral Cavity [+continuant] C-place | Vocalic V-place | [dorsal] Laryngeal | [voice]

Aperture | (height fts)

Aperture | (height fts)

Underlying Glide: /j/ root [+sonorant] Oral Cavity Laryngeal | [voice]

/w/ root [+sonorant] Oral Cavity [+continuant] C-place | [labial] Laryngeal | [voice]

[+continuant] C-place | [coronal] | [-anterior]

33 The representation in (32) has features under C-place instead of under V-place. Derived glides, on the other hand, do have features under the V-place node. In fact, they are indistinguishable from vowels except with respect to syllabic association. A further difference between these segments is that underlying glides lack aperture/height features which are only represented as dependents of the Vocalic node.18 The representations in (31) and (32) use the feature [coronal] to represent both underlying glides and vowels. Abundant work (e.g. Hume 1994) has shown the relationship between coronal consonants and front vowels. However, this particular theoretical bias is not required by VPT. That is, a representation that uses [dorsal, -back] for /j/ and /i/ would work in the same fashion. We will see in §3.5 that some versions of feature geometry rely crucially on the representation of glides as [coronal] and front vowels as [dorsal]. Hume 1995 and Herman 1994 use representations like (31) and (32) to differentiate what they call consonantal vs. vocalic glides. Hume 1995 assumes that these two types of glides are both underlying. The difference between their representation and the one that I assume is that they do not make the link between the consonantal glide and the underlying glide on the one hand, and the vocalic and derived glide on the other. Under my analysis, the "consonantal" glides have place features under C-place because they are underlying. The "vocalic" glides, on the other hand, have place features under V-place because they are derived. Thus, in the version espoused here, only true underlying glides have features under C-place, while derived glides have features under V-place. Hume 1995 also argues for the elimination of the feature [consonantal] from the tree. She argues that the difference between consonants and vowels should be determined by the presence or absence of place features under the C-place node. 3.4.2 Predictions It is immediately apparent that the VPT distinguishes underlying vowels and glides to a greater extent than the two preceding theories, LEX and [CONS]. Because of the division of features into V-Place and C-place, there are two types of processes that could distinguish underlying glides from underlying vowels and derived glides. First, the underlying glides are expected to be transparent with respect to processes that target Vplace features. For example, underlying glides should be transparent to vowel harmony processes, while derived glides should not be. Second, underlying glides should not be transparent to processes that affect or make reference to C-place features. That is, if a process looks at the level of C-place, the underlying glide should behave with the other supralaryngeal consonants.


This predicts that no language can have underlying glides that contrast for height.

34 3.5 Revised Articulator Theory (RAT) 3.5.1 Overview and Representation Like Vowel-Place Theory, Revised Articulator Theory (RAT) (Halle 1995 and Halle, Vaux, & Wolfe 2000) is a type of feature geometry. However, it differs from VPT in several respects. The most notable difference in RAT is that there is no segregation between vowel and consonant place features. All features are organized on the same level of the hierarchy. The feature tree assumed by RAT is given in (33). (33) RAT feature tree (Halle et al. 2000) [suction] [continuant] [strident] [lateral] [round] [labial] [anterior] [distributed] [coronal] [high] [low] [back] [dorsal] [nasal] [rhinal] [ATR] [RTR] [radical] [spread glottis] [constricted glottis] [stiff vocal folds] [slack vocal folds] [glottal] Lips

Tongue Blade


Tongue Body [consonantal] [sonorant] Soft Palate

Tongue Root Guttural Larynx

RAT is based on the idea that there are six articulators: the lips, the tongue blade, the tongue body, the soft palate, the tongue root, and the larynx. Each of these articulators gives rise to a designated articulator feature. In the tree in (33), the

35 articulator features are [labial], [coronal], [dorsal], [rhinal], [radical], and [glottal]. It is immediately apparent from the tree in (33) that there are two different classes of features. There are those features that attach directly to the root node and there are those that attach under one of the articulator nodes. The former are called articulator-free features because they can be articulated by several different articulators. For example, [+continuant] can be articulated by the tongue blade to produce [s] or the tongue body to produce [x]. The other features are called articulator-bound features because they can only be produced by one articulator. For example, [+round] can only be articulated by the lips. The designated articulator (e.g. [dorsal], [labial], etc.) shows which articulator is executing the articulator-free features. Electing to use full specification of features reflects a second departure from previous models of feature geometry.19 Not all versions of articulator theories of feature geometry require full specification. One of the central tenets of RAT, however, is the presence of all features at all points of representation. Calabrese 1995, whose results are accepted by RAT, shows that full specification is required in order to explain all of the facts of Japanese Rendaku and Lyman's Law. The basic facts surrounding Rendaku are as follows. Word-initial obstruents surface as voiced in compounds (e.g. ori kami origami 'paper folding'). Lyman's Law states that Rendaku does not apply when the word already contains a voiced obstruent (e.g. kami kaze *kamigaze 'divine wind'). What is important up to this point is that the voicing of sonorants does not prevent Rendaku (c.f. origami). An additional process of Post-nasal Voicing occurs (e.g. tombo 'dragonfly' and kam-te kande 'chewing'). Of significance is the fact that these post-nasal obstruents also block Rendaku (e.g. iroot kagae *irootogagae 'layman's idea'). If underspecification, instead of full specification, is adopted, an ordering paradox ensues. In order for post-nasal obstruents to be visible to Rendaku, Post-nasal Voicing must apply before Rendaku. For Post-nasal Voicing to occur at all, the voice feature on sonorants must be inserted via a redundancy rule. We therefore have the following order: (1) Redundant Sonorant Voicing, (2) Post-nasal Voicing, and (3) Rendaku. However, we know that the voicing of sonorants does not affect Rendaku, therefore Redundant Sonorant Voicing must follow Rendaku, hence the paradox. The paradox, however, only exists because of the assumption of underspecification. Calabrese 1995 shows that full-specification does not encounter the same problem. He argues that Lyman's Law, which prevents Rendaku when the word contains a voiced obstruent, is only sensitive to marked features (i.e. [+voice] on obstruents). Post-nasal Voicing is sensitive to all features and simply spreads voice features to a following obstruent. (See following discussion for sensitivity to different classes of features.) All that is needed in the full-specification is for Post-nasal Voicing


The type of full specification discussed in Calabrese 1995 could of course be implemented in other versions of feature geometry, for example in VPT.

36 to occur before Rendaku and Lyman's Law. Several of the examples in the following chapters will rely crucially on a full representation of segments. Full specification, however, does not imply that all features are visible at all stages of the representation. Features are determined to be marked, contrastive, or neither in the representation of each phoneme of the language. The basis for determining whether a feature is marked, contrastive, or neither is dependent on a set of universal marking statements (Calabrese 1995). A simple example of this is voicing in obstruents. It is a generally held belief that voiced obstruents are more marked than voiceless obstruents, yielding the marking statement in (34). (34) [-sonorant, +voice]

The marking statement should be read as follows: Among obstruents, voiced obstruents are marked. The underlined feature is that feature which could be contrastive in the set of [-sonorant] segments. If a language allows voiced obstruents, it is said to have deactivated the marking statement. Consider two languages, A and B. In Language A, both voiced and voiceless stops occur (e.g. [p b t d k g]). In Language B, on the other hand, only the voiceless stops [p t k] are present. Because Language A has voiced stops, the marking statement in (34) is deactivated. Therefore, both values of [voice] are contrastive in obstruents in this language. The contrastive features in bold (35). Additionally, the value [+voice] is marked for obstruents, which is indicated with an underline. Irrelevant features have been omitted. (35) Language A /p/ root Place | Lips [labial] [-round] Guttural | Larynx | [-voice] Place | Lips [labial] [-round] /b/ root Guttural | Larynx | [+voice]

Now consider Language B, which has only one series of stops. In this language, the marking statement in (34) is still active. As a result, neither value of [voice] is contrastive for obstruents in the language. Furthermore, there is no marked value for the feature [voice]. This is of course intuitive because the obstruents in the language do not contrast for voicing. Notice that even though the language does not contrast voicing in obstruents, the feature [-voice] is still present on this segment.

37 (36) Language B /p/ root Place | Lips [labial] [-round] Guttural | Larynx | [-voice]

The voicing example illustrated above is quite simple and intuitive. However, not all segments are obviously contrastive or non-contrastive for different features.20 In addition to full specification, the other innovation used in RAT, developed in Calabrese 1995, is the way in which phonological processes apply. A phonological rule can be sensitive to either marked features, contrastive features, or all features. Phonological transparency is explained by these divisions. Consider Language A again. Assume that it has a rule of voicing assimilation which only targets contrastive [voice] features. Furthermore, Language A has a sonorant series in which all segments are voiced (e.g. /n, l, r/). Since the marking statement [+sonorant, -voice] is still active (i.e. there are no voiceless sonorants), sonorants are not contrastive for [voice]. Therefore, their [voice] features are invisible or transparent to the voicing rule.21 Similarly, if a rule targets only marked features, then contrastive features are transparent. Marked and contrastive feature specifications are subsets of the set of all feature specifications. Furthermore, marked feature specifications are a subset of the contrastive feature specifications. If a feature is marked then it is necessarily also contrastive, even if the segment with that contrastive feature does not exist in the language. (37) Set of features (Calabrese 1995) All feature specifications Contrastive feature specifications Marked feature specifications


There are other ways in which a feature can be contrastive on a given segment. Calabrese (1995: 435) explains these other methods, which will not be discussed here. 21 In this simple example, underspecification is also a possible explanation for the transparency of voicing on sonorants. Calabrese 1995 discusses more complex examples in which an appeal to underspecification is not possible.

38 There are several other differences between RAT and other versions of feature geometry. First, previous articulator models (e.g. Sagey 1986) depended on a pointer to distinguish multiply articulated consonants (e.g. [kp]) from those with secondary articulations (e.g. [kw]). The presence of articulator features in RAT, such as [labial] and [dorsal], eliminates the need for these pointers. In RAT, a multiply articulated consonant such as [kp] will have two articulator features, while [kw] will only have the articulator feature [dorsal]. In (38), irrelevant nodes have been omitted. (38) kw root | Place Lips | [+rd] Body | [dorsal] Lips [-rd] [lab] kp (Halle et al. 2000: 392) root | Place Body | [dorsal]

Second, the articulator features are not nodes in the tree with other features dependent upon them, but are leaves, just as other features are (Halle et al. 2000).22 Third, though the articulator features are not binary, all other features are. The fourth difference between RAT and other versions of feature geometry lies in what areas of the tree are allowed to spread. In other versions of feature geometry, such as VPT, intermediate nodes are allowed to spread. The reason for this is to avoid spreading random sets of features in a single rule. By limiting spreading to a single feature (leaf) or to a node, the possible spreading rules are limited. The very organization of the feature tree was designed to facilitate spreading that is natural. Though RAT allows only leaves to spread, the insights from other models still hold. The spreading rules in RAT are of two types. First, a single feature may spread. Second, all features dominated by a particular node may spread. In other words, it is not possible to spread a subset of features under a particular node without spreading the others. Likewise, it is not possible to spread a few features from under different nodes in a single rule. Halle 1995 and Halle et al. 2000 argue not only that this type of spreading captures the identical generalizations to node-spreading, but also that it can account for certain processes that node-spreading cannot. To formalize the spreading in RAT, Halle proposes the following convention.


This marks a return to representing the articulator features as leaves in the tree. VPT in §3.4 represented the features [coronal] and [dorsal], for example, as nodes. Halle 1995 also considered these features to be nodes in the tree, but in Halle et al. 2000, these features are again considered to be leaves.

39 (39) Spreading Convention (Halle 1995: 20) The linking lines that are spread from one segment to another by an assimilation rule are those of terminal nodes in the tree, with the restriction that terminal nodes spread in a given rule are all and only those dominated by a single nonterminal node.

In order to illustrate that leaf-spreading exhibits the same properties as nodespreading, Halle provides an example from n-retroflexion in Sanskrit. (Also see Chapter 11 for more discussion of Sanskrit.) In (40), spreading the Tongue Blade node from the initial consonant to the final nasal is blocked by the intervening coronal consonant because of a line crossing violation. (40) Node Spreading (Halle 1995: 23)23 * [+cons] [+cons] | | Place Place | | Blade Blade [-ant] [-distr] [ ant] [ distr] [+cons] Place | Blade [-ant] [+distr] Soft Palate | [+nasal]

Similarly, in (41) spreading from the initial segment is blocked. In this example it is not the node itself that spreads or blocks, but is the terminal features that are dominated by the Tongue Blade node. (41) Leaf Spreading (Halle 1995: 23) * [+cons] [+cons] | | Place Place | | Blade Blade [-ant] [-distr] [ ant] [ distr] [+cons] Place | Blade [-ant] [+distr] Soft Palate | [+nasal]

In order to show that leaf-spreading is in fact superior to node-spreading, Halle et al. discuss vowel copy in Barra Gaelic. In their analysis, all features dominated by the Place node spread to the epenthetic vowel. When an intervening consonant specified for


In Halle 1995, the Blade node is labeled Coronal. This aspect of RAT was changed in the Halle et al. 2000 article.

40 backness intervenes, only the [back] feature of the stem vowel is blocked; all other vowel features are allowed to spread. (42) Terminal feature spreading in Barra Gaelic (Halle et al. 2000: 432) Vowel v /mrjv/ [mrjev] m rj Root Root Root 'the dead' | | | Place Place Place | Lips Body Body Lips Body | [-rd] [-hi] [-low] [+bk] [-back] Since both dependents of the Lips and the Tongue Body nodes spread, a theory that relies on node-spreading would have to spread the Place node itself. However, spreading the Place node incorrectly predicts that all features from the stem vowel are blocked by intervening consonants that contain this node (or at least those that have contrastive secondary articulations). In fact, all features are not blocked. The only feature of the stem vowel that does not spread to the epenthetic vowel is [+back] since the intervening consonant contains this feature. These two brief examples show that (1) leaf-spreading has the same predictions as node-spreading in most cases, so long as leaf-spreading is subject the convention in (39) and (2) that leaf-spreading actually fares better than node-spreading in the case of Barra Gaelic. Finally, we turn to the representation of glides and vowels in RAT. In RAT, backness features among vowels are represented as [±back] under the Tongue Body node and vowels have the designated articulator [dorsal].24 A high front rounded vowel /y/ would be represented as in (43). Irrelevant nodes have been omitted.


The existence of consonant harmony and the fact that it is restricted to coronal segments is easily explained in RAT, as in other articulator theories. Vowels are generally contrastive for labial and dorsal features, but not for coronal features, which explains the transparency of vowels in coronal consonant harmonies (Shaw 1991, Halle et al. 2000: 424).

41 (43) Place Lips | [+round] Body [dorsal] [+high] [-low] [-back] /y/ root [-consonantal] Guttural | Root | [+ATR]

Whereas vowels have [dorsal] as their designated articulator, glides do not (Halle 1995: 13-4). The designated articulator for the palatal glide /j/ is [coronal] and for the labio-velar glide /w/, [labial]. As with the other representations, an underlying vowel that surfaces as a glide due to syllable position has not altered its representation. Therefore, the surface glide derived from /i/ would still have [dorsal] as its designated articulator, whereas the underlying glide would have [coronal] as its designated articulator. In (44)(47) irrelevant features and nodes have been omitted (e.g. [+voice]). (44) Place Lips | [-round] (45) Body [dorsal] [+high] [-low] [-back] /j/ root [-consonantal] | Place Lips | [-round] Blade [coronal] [-anterior] [+distrib] /i/ (and derived [j]) root [-consonantal] Guttural | Root | [+ATR]

42 (46) Place Lips | [+round] (47) Body [dorsal] [+hi] [-lo] [+bk] /w/25 root [-consonantal] | Place Lips [lab] [+rd] Body [+back] /u/ (and derived [w]) root [-consonantal] Guttural | Root | [+ATR]

3.5.2 Predictions Given the representations of vowels and underlying glides in the previous section, there are several ways in which their behavior should differ. First, in processes that target Tongue Body (dorsal) dependents, /j/ should be transparent. If the language also has consonants which contrast for dorsal features (e.g. front and back velars or palatalized vs. plain consonants), then these segments could interact with a process like vowel harmony that targets contrastive dorsal features. For example, in (48c), the contrastive [back] feature on the intervening consonant will block the spreading from the preceding vowel to the following vowel. (48) a. Access contrastive feature specification. b. [-cons] [-cons] (irrelevant features are omitted) ... ... Body Body [dorsal] [back] [dorsal]


As with VPT, height features are not represented on underlying glides.

43 c. * V ... Body [dorsal] [back] C ... Body [dorsal] [back] V ... Body [dorsal]

Second, if a vocalic process targets contrastive values for [round], then /w/ and /j/ should be transparent because they are not contrastive for [round]. However, if the language has both /j/ and its rounded counterpart // (as in (49) and (50)), then a rule sensitive to contrastive [round] should interact with these two segments.26, 27 (49) /j/ root [-consonantal] | Place Lips | [-round] (50) Blade [cor] [-ant] [+distrib] // root [-consonantal] | Place Lips | [+round] Blade [cor] [-ant] [+distrib]

Third, processes that affect the Tongue Blade dependents might interact with /j/ depending on whether the rule targets marked, contrastive, or all features, but would not interact with /i/.

According to Maddieson 1984, only 1.3% (4/317) of languages in his sample contain (Breton, French, Mandarin, and Ga). 27 Actually, in RAT, an underlying potentially has two representations. It can either be the round counterpart of a basically coronal glide as in (50) or it can be the front counterpart of the basically round glide. None of the articles on RAT discuss this glide.


44 3.6 Summary This chapter considered four different approaches to represent underlying glides. The lexical marking (LEX) option does not target any features of the segments but relies only on syllable affiliation to distinguish underlying vowels from underlying glides. [CONS] uses only the feature [consonantal] to differentiate underlying vowels from glides. VPT targets the segregation of consonant and vowel features to separate the place features of these two segments. Finally, RAT relies on a single tiered feature hierarchy, full specification, and different designated articulators to differentiate between underlying vowels and glides. Any process that relies purely on syllabic position can be represented in all four approaches. Consider for example a cluster that requires epenthesis and does not allow the glide to surface as syllabic, as in /tj/ [taj], *[ti]. In LEX, the final glide is explicitly banned from being in the nucleus and forces epenthesis. In [CONS], the epenthesis rule can be stated to apply between two [+consonantal] segments. In VPT, characterizing epenthesis is slightly more complicated. Epenthesis in VPT is affected by the inventory of segments. The most general process of epenthesis must be stated to apply to segments with C-place features or to segments that lack V-place features. If the former stipulation is omitted and the process simplified to apply to segments that lack V-place features, then it would incorrectly not apply next to consonants with a secondary articulation. If the latter stipulation is omitted and epenthesis occurs only next to segments with C-place features, then it incorrectly does not apply adjacent to laryngeals, which lack C-place features. Of course, the inventory is relevant. If the language lacks either laryngeals or consonants with secondary articulation then epenthesis can be simplified. In RAT, an epenthesis rule can be stated to apply between two segments which are not characterized as [-consonantal], [dorsal] (designated articulator). This way the epenthesis will apply to [-consonantal] glides, to [dorsal] consonants, and to laryngeals. The types of processes that will help to establish a difference between these theories are primarily harmonic processes, as well as others that target features. Harmony processes that specifically target vowel features to the exclusion of glides and consonants will fare better in the two feature geometries than in LEX or [CONS]. The explanation for this type of division in a harmony process is embedded in the feature representation of glides and vowels. If the harmony process targets vowel features, then the only way to explain the transparency of glides is with a representation that uses a different set of features for the two segments. In LEX and [CONS], the features of glides and vowels are identical. Of course [CONS] can mention this feature in the harmony. Each of the representations of underlying glides requires some stipulation. This is perhaps not surprising given that the theories were not developed to represent glides of any kind. The differentiation of glides and vowels in [CONS] is dependent upon some dubious claims about the constriction of glides. VPT requires that the place features exist on different tiers but does not explain why underlying and derived glides should be perceived as the same. Do the different tiers of place features make any claims about constriction? RAT must rely on glides and vowels having different designated

45 articulators and it is not clear why /j/ should be coronal while /i/ is dorsal. Underlying glides are inherently elusive segments, overlapping perceptually with derived glides and being vowel-like but generally patterning with consonants. Furthermore, none of these representations were developed to account for underlying glides, so it is not surprising that stipulations must be made. Ultimately, we would like the representation of underlying glides to be the same cross-linguistically. One of the goals of this dissertation is to determine a universal representation of underlying glides. By looking at the glides in a variety of languages, we will be able to evaluate the quality of each of the four approaches in representing underlying glides.


4 Turkish

4.1 Introduction The first language to be discussed here that provides evidence for the existence of underlying glides is Turkish. Turkish is a Southern Turkic language (Altaic) spoken primarily in Turkey. The basis of the analysis in this chapter comes from Levi 2000;2001. The primary evidence for underlying glides in Turkish comes from vowel harmony and the distribution of glides in native roots. Turkish will serve as an introduction to the analysis of underlying glides and will provide the representation of vowel harmony that will be assumed throughout this dissertation. A distinction will be drawn between underlying glides, which do not participate in the harmony, and the lateral // which does. Some dialects of Turkish also contain front and back velars which condition harmony. A brief discussion of these segments is included for completeness, but they are not part of the dialect that will be treated here. We will see that Turkish is a Type V language where the underlying glide always surfaces as a glide and the underlying vowel always surfaces as a vowel. (51) Type V language /i/ /j/ [i] [j]

4.1.1 Vowels and Vowel Harmony Turkish has a symmetric eight vowel system: four vowels are round, four back, and four high. The vowel inventory in given in (52). In Turkish, [a] and [e] pattern as a backfront pair with respect to backness harmony. In the underlying representation, the symbols <I> and <E> will be used to represent a high or non-high suffix vowel. (52) Turkish Vowels i y u e ø o a Turkish exhibits two types of vowel harmony (VH): backness harmony and labial harmony. Backness harmony affects all suffixes, regardless of the height of the suffixal vowel. High vowel suffixes alternate between i ~ y ~ ~ u and, as is common in the literature on Turkish, such suffixes will be written as /I/ to indicate that the vowel is [+high]. Non-high vowel suffixes alternate between e ~ a and will be written as /E/,

47 showing that the suffix is [-high]. As the preceding description indicates, labial harmony only affects [+high] suffixes. The non-high suffixes never undergo rounding harmony in Turkish. Harmony proceeds rightward throughout the phonological word (Underhill 1976: 25, Underhill 1986: 13, Comrie 1997: 886). The chart in (53) shows which vowel surfaces after a particular stem vowel. Notice that the height of the triggering vowel does not affect the realization of the suffix; only its backness and labiality are relevant. In general, the same values for backness are present throughout the entire word. Since non-high vowels are never rounded as the result of harmony, vowels may not necessarily agree in labiality within a word. The data in (54) show some examples of backness and labial harmony (Clements & Sezer 1982: 216). (53) Surface vowels resulting from harmony Stem Vowel /I/ /E/ i, e i e y e y, ø a , a u, o u a (54) Nom. Sg. ip kz jyz son Gen. Sg. ip-in kz-n jyz-yn son-un Nom. Pl. ip-er kz-lar jyz-er son-lar Gen. Pl. ip-er-in kz-lar-n jyz-er-in son-lar-n Gloss 'rope' 'girl' 'face' 'end'

The data from the genitive singular confirms that the high suffix vowel alternates in both backness and rounding. Furthermore, the plural suffix surfaces with only two allomorphs: front and back. In the genitive plural forms where a non-high suffix intervenes between the root and the high vowel in the genitive, the /I/ surfaces as unround. In other words, rounding cannot skip the non-high vowel. In the case of underlying mismatches in backness and rounding, it is the rightmost vowel that spreads its features. In disharmonic roots, such as (55), the suffixes surface with back vowels since the final root vowel is back. In the present progressive suffix /-Ijor/ in (56), the second vowel is fully specified and does not alternate. That is, it does not surface as front after a front vowel root. Furthermore, it begins a new harmonic domain forcing the first person plural suffix to surface as [-uz] with a back round vowel, as in (56). (55) a. kitap-lar b. kitab- *kitap-er *kitab-i 'book-pl.' 'book-acc.'

48 (56) a. gid-ijor-uz *gid-ijer-iz b. kal-jor-uz *gid-ijør-yz 'we are going'28 'we are staying'

4.1.2 Consonants and Vowel Harmony The consonant inventory is provided in (57). (57) Consonant phonemes in Turkish (Levi 2000) Labial Dental, Palato- Palatal Palatalized Velar Glottal Alveolar alveolar Stop p, b t, d (kj gj) k, g Affricate t d Fricative f, v s, z h Nasal m n Liquid l, r Glide j The palatalized velars have a secondary articulation at the palate and are pronounced [k] and [g] (Lewis 1967, Underhill 1976, Kornfilt 1997). The palatal lateral is made near the palate, while the plain or velarized [l] is produced near the alveolar ridge and with some additional constriction near the velum (Lewis 1967, Underhill 1976, Kornfilt 1997). In this chapter, these laterals may be referred to as "front" or "back" respectively.29 The symbol <l> will be used to denote the 'back' lateral, but it should be noted that it is produced with a dorsal articulation. The palatal lateral [] and the palatalized velars [kj] and [gj] were formerly only allophones of the alveolar lateral and the plain velars, respectively.30 At an earlier stage, the palatal and palatalized allophones were only found in front vowel words, while their plain counterparts were found in back vowel words. Due to abundant loans in the Turkish lexicon, the palatalized sounds can now contrast in back vowel words, as in (58). Additionally, the plain lateral /l/ surfaces as [] in word initial position, regardless of the backness of the following vowel (Underhill 1976, Kornfilt 1997). The data in (58) give examples of minimal and near minimal pairs for these consonants.

28 29

These words have the morpheme structure verb-pres. The exact articulation of these laterals is unclear. Clements & Sezer (1982: 233) describe the two laterals as front (palatal) versus back (velar). Comrie (1997: 885) describes the back lateral as velarized and the front one as non-velarized. Zimmer & Orgun 1992 contrast the two laterals as palatalized postalveolar versus velarized dental. 30 This is still the case in the native vocabulary.

49 (58) kjar gjavur petro 'profit' 'infidel' 'gasoline' kar gaz bol 'snow' 'gas' 'abundant'

Clements & Sezer 1982 give examples of roots with underlying word final palatalized velars in back vowel words, as in /idrakj/ 'perception' and /imsakj/ 'fasting'. In their analysis, these forms surface with plain velars in unsuffixed forms, but condition front vowel suffixes and surface as palatalized in suffixed forms. The five speakers I consulted, however, regularized these forms and produced them with non-palatalized velars in all morphological environments. Furthermore, these speakers treated these forms as regular when suffixed as well, producing back vowel suffixes. In addition to the front velars in back vowel roots, there were formerly some morphemes with back velars in front vowel words. Among those cited as containing a word final plain velar in a front vowel environment are /evk/ 'desire', /fevk/ 'top', and /sevk/ 'driving (n.)' (Clements & Sezer 1982: 242). Clements & Sezer acknowledge that these forms are found primarily in the speech of older speakers. When presented with these forms, Clements' & Sezer's college-aged informants (as well as my own) invariably produced these with the front velars. In other words, these words were treated as fully harmonic. A summary of this distribution is given in (59). The normal pattern is shaded. In this chapter, the palatalized velars will be considered purely allophonic and will not be marked in the text. (59) Summary of the distribution of velars in Turkish (Levi 2000) kj k unsuffixed suffixed unsuffixed suffixed 31 regularized in the speech of bebek bebe-i front 'baby' 'baby acc.' younger speakers vowel word (e.g. [evk] [evkj] 'desire') regularized in the speech of ok ok-u back 'arrow' 'arrow acc.' younger speakers vowel word (e.g. [idrakj] [idrak] 'perception') Whereas the underlying contrasts between velars have been eliminated for many speakers, the palatal lateral can contrast in final position. In this position it can trigger harmony, as in (60).


Velars delete intervocalically in most roots, with the few exceptions (generally monosyllabic), such as [ok-u] 'arrow acc.'.

50 (60) a. /petro/ petro petro-y petro-de b. /megu/ megu megu-dym megu-ym 'gasoline' *petrol *petro-u *petro-da 'busy' *megul *megu-dum *megu-um

Nom. Sg. Acc. Sg. Loc. Sg.

'he is ...' 'I was ...' 'I am ...'

The other lateral in Turkish shows allophonic alternations depending on the vocalic environment. In native morphemes, the alternation is purely allophonic, as in the plural suffix /-LEr/ in (61). [] surfaces in front-vowel domains, while [l] surfaces in back vowel words. (61) Sg. ip kz jyz son Pl. ip-er kz-lar jyz-er son-lar Gloss 'rope' 'girl' 'face' 'end'

4.2 Basic phonology of Turkish: syllabification This section will discuss four types of evidence for the location of glides in Turkish syllables (Levi 2000, 2001). The location of glides in Turkish syllables will not provide evidence for their underlying status. Knowing the location of glides, however, will be necessary for the discussion of the representation of harmony that will be discussed in The data will show that the glide in Turkish does not form part of a nuclear diphthong, but instead is syllabified in either onset or coda. The first type of evidence comes from their distribution. The data in (62) illustrate that the palatal glide can occur both before and after each of the eight vowels. In addition to the verbal stem [kj-], my consultants produced the nonsense word [pj] with the expected back vowel suffixes, e.g. [pj-da] 'pj-locative'. (62) a. Glide before each vowel jiv 'groove' jyn 'wool' jemek 'food' jøre 'neighborhood'

jl julaf jad jok

'year' 'oats' 'strange' 'absent'

51 b. Glide after each vowel gij'dress' (verb) tyj 'feather' bej 'gentleman' køj 'village'

kjhuj taj boj

'chop, massacre (verb)' 'disposition, temper' 'stream' 'length'

If the glide were nuclear, then Turkish would have, in addition to the eight vowels in (52), eight rising diphthongs (those in (62a)) and eight falling diphthongs (those in (62b)). That the glide can occur before or after each vowel implies that it is not nuclear, but rather in syllable peripheral positions (onset or coda). The second type of evidence comes from disallowed consonant clusters. Turkish does not allow any onset clusters. The lack of words beginning with CjV is further evidence that the glide does not form part of a diphthong, but instead occupies peripheral positions. While coda clusters of the form sonorant-obstruent are allowed, sequences of two sonorants are not. Nonetheless, some Turkish roots end in two underlying sonorants (Clements & Sezer 1982: 243, Kornfilt 1997: 497). When these roots occur in unsuffixed forms or before consonant-initial suffixes, an epenthetic vowel is inserted between the two sonorants. Before a vowel-initial suffix, no epenthesis is required since the sonorants can be syllabified in two different syllables. In (63), the epenthetic vowel is underlined and a period indicates the syllable boundary.32 (63) UR /aln/ /burn/ /bojn/ Nom. Sg. a.ln bo.jun Nom. Pl. a.ln-.lar bo.jun-.lar Acc. Sg. al.n- bur.n-u boj.n-u Gloss 'forehead' 'nose' 'neck'

What is important about this data is that /bojn/ surfaces as [bojun] and not as *[bojn]. If the glide were syllabified as part of a nuclear diphthong, then only the nasal would occupy the coda and epenthesis would not be necessary. Because epenthesis does occur, it is clear that the glide is non-nuclear.


An alternate account of the data in (63) is vowel deletion (Clements & Sezer 1982: 244). In this analysis, the vowel is underlying and is only deleted when vowel-initial suffixes are added. For example, this analysis would assume /bojun-I/ and /burun-I/ would surface as [boj_nu] and [bur_nu] as the result of vowel deletion. When consonant-initial suffixes are attached, the second vowel remains because deletion would create an impermissible cluster. Therefore, /bojun-LEr/ and /burun-LEr/ surface as [bojunlar] and [burunlar]. Crucially, the form *[bojnlar] is disallowed because the coda cluster j-n is disallowed. In the deletion account, the analysis of the glide is the same; it cannot be parsed in the syllable nucleus. This deletion analysis cannot account for forms such as [kojun] 'sheep' and [bakr] 'copper' which do not alternate. The genitive singular forms of these words are [kojunun] and [bakrn], not *[koj_nun] and *[bak_rn].

52 Additional evidence comes from the loans in (64). As before, these forms are relevant because of the final sonorant-sonorant sequences. In addition to the Arabic data in (64a), when asked how a certain English brand name would be pronounced in Turkish, my consultants responded as in (64b). In all of the data in (64), the lending language allows the sequence VjR (R=sonorant) to surface. When these forms are borrowed into Turkish, however, they surface with an epenthetic vowel between the glide and the sonorant. As was the case in (63), the glide cannot be syllabified in the nucleus, and epenthesis results. If the glide in Turkish were nuclear, these sequences would surface unchanged. (64) a. Gloss 'good' 'man's name' 'tendency' b. Gloss '(C.) Klein' Arabic form khajr husajn mejl English klajn Turkish loan (Lewis 1967: 18) hajr hysejin mejil Turkish loan (Levi 2000: 22) kejin

The third type of evidence showing the location of the glide in Turkish syllables comes from quantity sensitive stress. One class of words in Turkish, consisting mostly of place names and loans, exhibits quantity sensitive stress (Sezer 1983: 65, Kaisse 1985: 200, Kornfilt 1997: 504, Inkelas 1999). In this class, stress falls on the penultimate syllable unless the penultimate syllable is light and the antepenultimate syllable is heavy (HL, else ).33 In particular, words ending in HH are stressed on the penult. The examples in (65) illustrate this pattern. Of particular interest is the form in (65b). Here, the glide must occupy the onset position, forcing the lateral into the coda of the preceding syllable. If [ja] were a nuclear diphthong, then the lateral would occupy the onset and the penultimate syllable would be light, exhibiting the pattern HL. Given this pattern, stress would be expected on the heavy antepenultimate syllable as in (65e). Because stress falls on the penultimate syllable in a HX pattern, the penultimate must be heavy as well.

This class of words is variably described as building iambs (with final syllable extrametricality) (Barker 1989), as building moraic trochees (with final consonant extrametricality) and invoking foot extrametricality in clash (Hayes 1995), or as the interaction of the constraints TROCHAIC, NON-FINALITY, and CONTOUR (a language specific ban on H(L) in a special co-phonology) (Inkelas 1999). Not all authors believe that this class of words has predictable or productive stress (Kabak & Vogel 2001).


53 (65) a. is.tan.bul b. an.tal.ja c. va.ik.ton d. e. 'city in Turkey' "city in Turkey" "Washington" "Kennedy" "window" HH (Sezer 1983) HH *an.ta.lja (*HL) LH LL HL

The last type of evidence for the syllabic position of Turkish glides comes from allomorph selection. This evidence comes from two suffixes: the third person possessive suffix and the reflexive suffix. The third person possessive suffix has two forms: /-sI/ and /-I/. The s-initial allomorph surfaces after vowel-final stems (open syllables) (see (66a)), while the vowel-initial allomorph surfaces after consonant-final stems (closed syllables) (see (66b). Crucially, when the stem ends in a glide, the vowel-initial variant surfaces (see (66c)), indicating that the stems ends in a closed syllable. As before, if the glide were part of a diphthong, the s-initial allomorph would surface. Since it does not, the glide must be in a non-nuclear position.34 (66) a. boru b. jlan c. saraj boru-su jlan- saraj- 'his/her pipe' (Lees 1961) 'his/her snake' 'his/her palace'

Next, consider the case of the reflexive verbal suffix /-(I)n/ in (67). Roots which end in a consonant surface with the full suffix /-In/ (67a). After vowel-final roots, on the other hand, only the nasal surfaces (67b). Roots ending in a palatal glide pattern with consonant-final roots in taking the full suffix /-In/ (67c). Consonant-final stems take the vowel-initial allomorph because the consonant-initial allomorph would result in an impermissible cluster. As was the case with the data in (64), the glide-nasal clusters are impermissible, showing that the glide cannot occupy nuclear positions.35 (67) Reflexive suffix (Underhill 1976: 357, Kornfilt 1997: 139) a. gør- "to see" gør-yn"to seem" b. jka- "to wash" jka-n"to wash oneself" c. gij- "to wear" gij-in"to dress oneself"


Another way to characterize this data is to say that the different allomorphs are sensitive to whether or not there is a branching structure in the rhyme. In this analysis, it would be possible to assume that the glides form part of a branching nucleus. However, this representation does not account for the other data in this section. 35 Whether these suffixes are assumed to delete the vowel in some cases or epenthesize it in the other does not matter. Either analysis is possible. As for the consonant that surfaces, Turkish has two different default/epenthetic consonants, s and n (as in ­(s)I or ­(n)In).

54 4.3 Evidence for underlying glides 4.3.1 Vowel Harmony The fact that the palatal glide always surfaces in syllable margins does not prove that the glide is underlying. This section will provide the first evidence for the underlying nature of glides in Turkish by examining their behavior in vowel harmony. Evidence comes from glides and laterals will show that harmony is not simply a nuclear-to-nuclear phenomenon, but is instead dependent on the features of segments. The glide The only place where /j/ could affect the harmony is in final position, because it would be the rightmost potential trigger. The data in (68), however, show that the glide is completely transparent. (68) Sg. koj paj køj Pl. koj-lar paj-lar køj-er Acc. koj-u paj- køj-y Gloss 'cove' 'share' 'village'

*koj-er *paj-er

*koj-i *paj-i *køj-i

If the glide were to interact, we would expect the suffixes to surface with front vowels, which they do not. The glide also does not cause the following vowel to unround as is evident in the accusative forms. If the glide were simply /i/ then it would have the same features as /i/ and would be expected to participate in the harmony. Therefore, it cannot be an underlying vowel. In addition to the data in (68), my consultants used back vowel suffixes in nonsense words such as [pj] and [puj], indicating that the transparency of the glide is not a lexically specified phenomenon, but is productive within the phonology of Turkish. Representing harmony: Evidence from laterals in Turkish Before continuing, it is necessary to consider the representation of harmony. That is, is it legitimate to assume that harmony should interact with all segments which have the appropriate features, regardless of their position in the syllable? Some might consider the arguments in the previous section premature. What if harmony did not simply spread features between segments, but instead expressed some relationship between syllabic or nuclear elements (Hulst & Weijer 1995)? If this were the correct representation, then the transparency of glides in Turkish Vowel Harmony would be explained since glides in Turkish are non-nuclear (see §4.2) (Levi 2000, 2001). However, an analysis that is based on the notion that only syllabic or nuclear segments participate in vowel harmony cannot explain the interaction of the non-syllabic, non-nuclear lateral in Turkish. We have seen that in addition to vowels, the palatal lateral // acts as a blocker and a trigger to backness harmony in Turkish. When // is word final, it can trigger a

55 new harmonic domain, as seen in (69). Notice that this segment need not be strictly adjacent to the harmonic vowel. (69) a. /petro/ petro petro-y petro-de b. 'gasoline' *petro-u *petro-da Nom. Sg. Acc. Sg. Loc. Sg.

/megu/ 'busy' megu megu-dym *megu-dum megu-ym *megu-um

'he is ...' 'I was ...' 'I am ...'

Because the palatal lateral interacts with the harmony, it is not possible to analyze the harmony as a phenomenon that only occurs between syllabic segments (Levi 2000, 2001). Therefore, harmony must be analyzed on the segmental level. Since harmony is a relationship between segments, it should affect all segments which have the relevant features. The question is which segments have the relevant features. Recall from Chapters 1 and 3 that languages can have two different underlying vocoids: underlying vowels and underlying glides. More importantly, surface glides can be one of two types: derived glides (from underlying vowels) or underlying glides. Since derived glides are reflexes of underlying vowels, their features are the same. Given these types of segments, which have features that will participate in vowel harmony? Certainly vowels have the harmonic features. If glides in Turkish were underlying vowels, then they would also be forced to participate in the harmony (in light of the fact that harmony must occur on the segmental, featural level). Because surface glides do not participate, they cannot be underlying vowels, but must be underlying glides. Ideally, the representation of harmony, that is, the level at which the harmony propagates, should be the same across all languages. Since representing harmony on a syllable or prosodic level fails in Turkish, such a representation cannot be the correct way to characterize vowel harmony. Though the other languages in this dissertation that exhibit vowel harmony do not have consonants which interact, I will conclude that the propagation of harmony must be represented cross-linguistically on a featural level.36


My general assumption is that vowel harmony must propagate on the level of the features. Kaisse & Levi 2004 discuss Pasiego, a language in which harmony cannot be explained with the same featural representation. The harmony in this language, however, is dependent upon stress. Because the harmony already focuses on prosodic structure, we assume that the harmony may also target this level of the representation, but only if the prosodic level is already salient in the harmony. Therefore, Pasiego does not represent a counterexample to the claim that harmony must occur at the featural level.

56 4.3.2 Native roots The second type of evidence for underlying glides comes from their distribution in the native vocabulary. Whereas borrowed roots may be disharmonic with different values of backness on vowels or with the front lateral in back vowel words (e.g. [petro] 'gas', [ote] 'hotel' [kitap] 'book'), virtually all native Turkish roots are harmonic (Lees 1961, Yavas 1980, Comrie 1997).37 Assume for a moment that Turkish does not have underlying glides and that all surface glides are derived from /i/. If this were the case, any native root with a back vowel and [j] would be underlyingly disharmonic. The disharmony in the native vocabulary would then be due to underlying /i/ which always surfaces as [j]. Many native words exist with back vowels and glides, such as [jol] 'road', [jurt] 'dorm', [bojnuz] 'horn', [jajn] 'sheatfish/sheathfish', [jaj] 'bow', [oja] 'embroidery', [kaj-dr-mak] 'to slide', [boj] 'stature', [bajram] 'religious feast', [boja] 'paint', [ja] 'fat', [kaja] 'huge rock', [saj-] 'count', and [daj] 'maternal uncle' among many others.38 Forms [koj] 'bay', [soj] 'family', and [taj] 'colt' would have the underlying forms /koi/, /soi/, and /tai/. Furthermore, this disharmony would only occur in roots where the /i/ surfaces as [j]. Finally, assuming that the surface glides are derived from underlying vowels cannot explain why labial harmony is blocked by the /i/ in /pois-I/ [pois-i] *[pois-y] 'police acc.', but not by the "/i/" in /koi-I/ [koj-u] *[koj-i] 'bay acc.'. Assuming that the surface glides are derived from /i/ faces serious problems. If instead we assume that the surface glides are derived from underlying glides /j/, these issues are no longer problematic. The roots [koj], [soj], and [taj] are no longer disharmonic. Furthermore, the transparency of the glide in labial harmony is also explained since it has different features than the vowel. The conclusion is therefore that the glide in Turkish is not derived, but rather underlying. 4.4 Comparison of the representations This section will compare each of the four representations discussed in Chapter 3. There are four aspects of Turkish vowel harmony that need to be considered when comparing the models. First, most suffix vowels take on the backness of the preceding vowel unless they are themselves fully specified vowels (e.g. the /o/ in /-Ijor/ does not alternate). Second, the palatal glide /j/ is completely transparent to the harmony, acting like /p, m, t, n/ for example. Third, the velar stops /k, g/ exhibit two allophones, one in front vowel environments, the other in back vowel environments. Fourth, the 'plain' lateral /l/ or /L/ also alternates between two allophones, front and back whereas the front/palatal lateral

There are a few native roots that are disharmonic, [elma] 'apple' and [tamur] 'mud' (Lees 1961: 12), as well as [karde] 'sibling' (Comrie 1997: 888). 38 All of the words in this list are of Turkic origin, according to Sergei Starostin's online Altaic StarLing database.


57 can occur in back vowel environments. When it is in final position, it triggers a new harmonic domain, thereby acting as opaque. (70) Summary of Turkish facts a. Suffix vowels /I/ /E/ [i] [y] [] [u] b. Glide /j/ [j] c. Velars /K/ [kj] d. [k] [e] [a]

Laterals /L/ [l] []

// []

The challenge for each of the models is to describe vowel harmony in a single rule, as a single phenomenon, not as multiple disparate phenomena. 4.4.1 LEX Specifying underlying glides as non-nuclear, as LEX requires, immediately runs into problems with the Turkish data. LEX assumes that the featural composition of /i/ and /j/ is identical. Thus, any rule that refers to the features of vowels, cannot exclude the features of the glides. The only way for LEX to account for the difference between vowels and glides is to refer to the syllabic affiliation of these two sets of segments. If the harmony rule could specify that only nuclear or syllabic segments could interact with the harmony, then the transparency of underlying glides in Turkish could be explained. The glides are non-nuclear and therefore do not interact. The problem, as we saw in §, is that front laterals are clearly non-nuclear and yet are not transparent. Thus, making reference to the syllabic position of possible targets and triggers is impossible. Furthermore, LEX misses the generalization that any segment with a particular feature or property will interact with the harmony. Because it cannot adequately distinguish the glide from the vowel with respect to harmony, it fails to explain the necessary facts of Turkish.


4.4.2 [CONS] Not surprisingly, distinguishing underlying glides from vowels by using only the feature [consonantal] also runs into problems. Under this representation, all features are shared between vowels and glides except that the glide is [+consonantal] whereas the vowel is [-consonantal]. In this model, either an SPE-style representation with unordered features or a feature geometric representation can be used. In neither case can the sole use of the feature [consonantal] adequately distinguish glides from vowels and account for the facts of harmony in Turkish. A brief outline of the relevant consonants is necessary. The velar stops are assumed to be [dorsal] segments. The front lateral is [-back] while the back lateral is [+back]. The glide is of course [+consonantal] but otherwise agrees with /i/ in its features. A first attempt at explaining the transparency of glides could use a rule like that in (71). Since the rule only targets [-consonantal] segments, only the vowels undergo the harmony, leaving glides unaffected. Furthermore, the rule specifies the trigger as [-consonantal], eliminating the possibility that the glide could condition a new harmonic domain since it is [+consonantal]. (71) Vowel Harmony in Turkish (First attempt)39 [-cons] [ back] / -cons C0 _____ back The problem with limiting the harmony rule to only [-consonantal] segments is that it cannot account for the fact that [+consonantal] velar stops are targets and that [+consonantal] front laterals are triggers. Either the rule must be amended or an additional one posited. An additional rule is not ideal as it misses the point that the effects of harmony are all due to a single process. Thus, an amendment to the rule in (71) must be made. The second attempt at the harmony rule is given in (72). (72) Vowel Harmony in Turkish (Second attempt) [dorsal] [ back] / [ back] C0 _____ This rule has several benefits over the first one. It allows all dorsals to be targets, explaining why velars and vowels are targets. It also allows the front lateral to be a trigger since it is [-back] and since the target is not specified as [dorsal]. The problem is that now the glide will incorrectly be a target of this rule since it has all of the features of a vowel. Since both [+consonantal] and [-consonantal] segments are triggers and targets

A corresponding OT constraint, e.g. AGREE-BACK(-CONSONANTAL), would also make reference to the feature [consonantal]. Chapter 1, section 1.1 has more examples of harmony constraints in OT.


59 of the harmony, any rule of harmony cannot use this feature to characterize the trigger or target. If the rule cannot refer to this feature, then it cannot differentiate between glides and vowels, since this is the only feature which distinguishes them. Therefore, [CONS] cannot account for all of the facts of harmony in Turkish. Before moving to the next model, I would like to point out that using a feature hierarchy is not a way to fix the problem of [CONS]. For example, in a segregated model of the feature geometry, such as Vowel-Place Theory, that relied only on [±consonantal], both vowels and glides would have V-place features but different values for [consonantal]. Even if all of the stipulations and other representations of VPT were the same as those we will see in §4.4.3, it would not be possible to model the facts of Turkish. In particular, any segment with Vocalic Node dependents acts as a target or trigger regardless of whether it is [+consonantal] (e.g. /k/) or [-consonantal] (e.g. /i/). Therefore, the underlying glide, having an identical representation to /i/ with the exception of [consonantal] would be expected to similarly participate in the harmony. We conclude that using only [consonantal] to differentiate underlying glides from vowels is problematic; it is not sufficient to distinguish these segments. 4.4.3 Vowel-Place Theory In contrast to LEX and [CONS], the richer representation of Vowel-Place Theory (VPT) can account for the facts of Turkish Vowel Harmony. The representations of the relevant segments are provided in (73). Irrelevant features are omitted. What is particularly important in (73) is that some segments have an empty V-place node. These are exactly the segments that alternate in backness depending on the backness of the preceding vowel (or lateral). The underlyingly front lateral has a vocalic [coronal] feature, while the underlying glide has only C-place features. (73) Representation of segments in VPT Suffix Vowels Vocalic | V-place /l/ (allophonic) C-place | Vocalic | V-place

/j/ C-place | [coronal] /k/ (allophonic) C-place | Vocalic | V-place

60 // C-place | Vocalic | V-place | Lingual | [coronal] Each of these representations is necessary in order to accurately account for harmony in Turkish, as we will see shortly. We now turn to the representation of harmony. The spreading which accounts for the backness harmony in Turkish is given in (74). (74) Backness harmony in VPT Vocalic | V-place | Lingual [coronal] [dorsal] Vocalic | V-place

The representation in (74) differs slightly from that in Chapter 2 by the addition of a Lingual Node which intervenes between V-place and the two place features [coronal] and [dorsal]. The addition of this node is necessary in order to account for the fact that suffix vowels surface as both front and back and that both front and back vowels block spreading from a preceding vowel. It is not possible to analyze the Turkish harmony as simply spreading one of [coronal] or [dorsal] and making the other a default. Evidence for the Lingual Node as a dependent of V-place comes from languages that spread both front and back in harmony (Browman & Goldstein 1989, Clements & Hume 1995).40 Spreading the lingual node explains how vowels of different backness block the harmony. Furthermore, the Lingual Node is necessary in order to accurately account for the behavior of labial harmony (Levi 2000, 2001). It is not possible to combine the labial


According to Clements & Hume, there is also evidence from Mandarin and Slovak for a Lingual Node under C-place.

61 harmony and the backness harmony in a single rule spreading the higher V-place node. A combined rule would spread the entire V-place node, as in (75).41 (75) Combined labial and backness harmony Vocalic Vocalic = V-place V-place [labial] [coronal] [dorsal] [labial] [coronal] [dorsal]

The problem with a combined spreading rule is that it incorrectly predicts an unrounded vowel in the suffix in (76) since the entire V-place node spreads. Therefore two separate rules must account for the labial and backness harmony. Linking [coronal] and [dorsal] under the Lingual node as in (74) allows for these features to spread together to the exclusion of the feature [labial]. (76) Problem with combined spreading p e tr o ... ... ... Vocalic Vocalic Vocalic | | | V-place V-place V-place | | [coronal] [labial] [dorsal] [coronal]

I ... Vocalic


The correct formulation of harmony in (74) accounts for backness harmony in Turkish. We now consider the adequacy of the combined representation and spreading rule in accounting for the facts of Turkish. In (77), several of these factors are presented. First, the glide is completely transparent to the harmony. Because it lacks a Vocalic node, it does not act as a blocker or target. This example also illustrates that suffix vowels, with their empty V-place nodes function as targets of the harmony. In (77), the Lingual node spreads from the root vowel to the suffix vowel which surfaces correctly as [dorsal]. Finally, (77) shows that the plain lateral is a target of backness harmony because it has an empty V-place node. If the root vowel were [coronal], the suffix vowel would surface as [e], the lateral as [], and the glide would still be transparent.


An additional mechanism would have to account for the delinking of [labial] on non-high suffix vowels, which only alternate between the unround [a] and [e].

62 (77) /taj-lEr/ [taj-lar] t a j C-pl C-pl | | Voc. [cor] | V-pl | Lingual | [dor]


l C-pl | Voc. | V-pl

E C-pl | Voc. | V-pl


The example in (78) shows how the front lateral blocks backness harmony from a preceding vowel. The Lingual Node from the /o/ is blocked from spreading across the lateral, because this would result in a violation of the No Crossing Constraint (Goldsmith 1979). This example further illustrates the blocking of fully specified vowels. Just as /o/ is not allowed to spread across //, initial /e/ is barred from spreading across or to /o/. The /o/, being fully specified, cannot be the target of the rule since the rule only allows for the spreading to empty V-place nodes. (78) /petro-dE / [petro-de] p e tr o Vocalic Vocalic Vocalic | | | V-place V-place V-place | | | Lingual Lingual Lingual | | | [coronal] [dorsal] [coronal]

I Vocalic | V-place


Finally, the data in (79) demonstrate the allophony of the velar stop. In the suffix /-dIk/ '1 p. pl. past', the final velar takes on the vocalic features from the preceding specified vowel and surfaces either with a front [kj] or back [k] allophone.

63 (79) /san-dIk/ [san-dk] 'we imagined' /et-dIk/ [et-tikj] 'we did' s a n - d I k e t - d I k Voc. Voc. Voc. Voc. Voc. Voc. | | | | | | V-pl V-pl V-pl V-pl V-pl V-pl | | Ling. Ling. | | [dor] [cor] The analysis of Turkish relies crucially on the representations in (73). In particular, it is necessary for some segments to be specified with an empty vocalic node (suffix vowels, velars, plain lateral). It is also necessary to posit a Lingual node to account for the front/back alternations. VPT can account for the five facts of Turkish backness harmony. First, the suffix vowels alternate in backness because they have an empty V-place node. Second, the transparency of the glide is explained by its lack of Vocalic features. Third, the allophony of the velars is accounted for by their empty Vplace node. Fourth, the alternating laterals are similarly represented with an empty Vplace node, accounting for their front and back surface forms. Finally, the front lateral has V-place features and therefore blocks harmony from a preceding vowel and initiates a new harmonic domain. Before moving to the last model, I would like to point out how contrastively back velars (c.f. §4.1.2) could be accommodated in this analysis. In a dialect where these velars are contrastive in final position, they can easily be represented in VPT with contrastive V-place features. An underlyingly front velar would have a [coronal] feature under V-place, while a back velar would have a [dorsal] feature under V-place. The alternating velars could still be represented with an empty V-place node, thereby accounting for their allophony. Finally, I would like to point out that rounding harmony can be easily explained in VPT. The rounding harmony simply spreads V-place [labial] from one segment to subsequent segments targeting the V-place node. Rounding harmony is not blocked by any intervening segments from (73) because they are not specified as [-labial]. It is also not blocked by plain labial consonants such as /b/ because it lacks V-place features. In fact, in VPT the place features are monovalent. Thus, rounding harmony can spread to the suffix vowel in [petro-y] and is not blocked by the final lateral even though it has some V-place features.42


The reader may wonder if the consonants with empty V-place nodes are also targets of rounding harmony. Boyce 1990 provides some evidence for this. In her analysis of lip protrusion, she found that Turkish speakers maintained a plateau in the amount of lip protrusion through the nonsense word [kuktluk].

64 4.4.4 Revised Articulator Theory It is also possible to use Revised Articulator Theory (RAT) to represent all of the facts of Turkish backness harmony. Since the basis of Turkish backness harmony is the set of vowels, we will start by looking at their features. The eight vowels of Turkish can be described using the three features [high], [back], and [round]. In the strictest Calabresean (1995) marking statements, the low vowel /a/ is not contrastive for [round] or [back], or even [high]. Since it is the only low vowel, it would only be contrastive for [low]. However, when we consider the vowels in (80) it is clear that the three features [high], [back], and [round] evenly divide the set of vowels into two sets. That is, there are four high vowels and four non-high. Likewise, there are four round and four back vowels. Looking at the vowels from this perspective, we see that all of the vowels are contrastive for these three features, as indicated by the boxed features in (80). (80) Contrastive Features for Turkish Vowels i y e a ø High Low Back Round + + + + + + o + + + + u + + +

Now that the contrastive features of vowels are clear, it is possible to establish a rule of spreading. In (81), only contrastive features are visible. The rule spreads contrastive [back] rightward from a segment with contrastive [back] to another segment which has a designated articulator [dorsal]. Since it is the feature [back] which spreads, only segments with contrastive [back] features will block the rule. Notice the rule does not spread to a segment which is itself designated with a [back] feature and does not delink this feature. This lack of delinking explains why fully specified vowels do not undergo harmony.43 (81) a. View contrastive features. b. Place Place | | Body Body [dorsal] [ back] [dorsal]


This same assumption was used in VPT.

65 The representations of the relevant segments are given in (82). In the representations in (82), the features [high] and [low] from the Tongue Body node have simply been omitted since they are not relevant for the backness rule in question, though they are actually present. The suffix vowels, whose backness is completely predictable from the preceding environment, are not specified for [back] in the underlying representation. Though Calabrese 1995 argues that all segments must be fully specified in the underlying form, I believe that it is not aberrant to assume that harmonic suffix vowels are not specified. Because they are completely predictable, there is no way to decide on their underlying representation, aside from their height. As described in Chapter 3, the palatal glide /j/ is represented as [coronal] in this theory. (82) Representations in RAT a. Suffix Vowels Place | Body | [dorsal] c. /l/ (allophonic) Place Blade | [cor] Body | [dor] b.

/j/ Place | Blade

[cor] [-ant] [+distr]


// Place

Blade Body | [cor] [dor] [-bk]

e. /k/ (allophonic) Place | Body | [dor] What is important to notice in these representations in (82) is that the suffix vowels (82a), the plain lateral (82c), and the velars (82e) have a designated articulator [dorsal], but do not have [ back] features. This too is a slight departure from the standard RAT and Calabresean requirement of full specification. I will argue that this does not undermine RAT's fundamental notion of full specification because of the particular segments that are involved. In (82), the only segments that are not fully specified are those that have a designated articulator [dorsal]. Flemming 2002 remarks

66 that cross-linguistically the exact location of the constriction of velars varies with the vocalic environment. That is, in a front-vowel environment, the constriction is more forward than in a back-vowel environment.44 In such a situation, the velars cannot be fully specified in the underlying representation because their value for the feature [back] is dependent on the environment. Thus, it is reasonable for the velar consonants to be unspecified for [back] in their underlying representations and to gain this value as a part of the spreading rule. This explains the representation of velars in (82e) as lacking [back]. We now move to the representations of the laterals in (82c&d). The laterals in Turkish are distinct, one having an allophonic distribution with a front and back version depending on the vocalic environment (as in (82c)) and the other always surfacing as front (as in (82d)). Their representations show both Tongue Blade and Tongue Body features. As the Tongue Blade features are not relevant to the harmony, these have been omitted.45 From the descriptions of the laterals (see Fn. 29) it is clear that the back lateral has both a velar and a dental articulation. The front lateral also has a dorsal articulation, but it is closer to the palate. There are three logical possibilities to represent the laterals in Turkish. First, they could have a primary articulation of [coronal] (Designated Articulator) with a secondary articulation of [±back] based on the vowel environment. Second, they could be doubly articulated with two designated articulators [coronal] and [dorsal]. Third, they could have a primary articulation of [dorsal] with a secondary articulation with the tongue blade. The third possibility is rather strange. In fact, if secondary articulations are thought of as vocalic articulations super-imposed on the consonantal articulation, then in RAT, it would not be possible to have a tongue blade secondary articulation since vowels have dependents only under the Lips and Tongue Body nodes.46 I do not know of any consonants that have a secondary articulation made with the tongue blade. The choice between a doubly articulated lateral and one with a secondary articulation of the tongue body requires some discussion. The allophonic/plain lateral in Turkish has two realizations, front and back. The tongue body articulation is completely dependent on the vocalic environment, just as it is in the velars. This seems to imply that


Flemming 2001 cites Katt 1987 and Sussman et al. 1991 who argue that velars have two F2 loci, a high locus in front-vowel contexts and a lower one in back-vowel contexts. Since F2 is the formant that corresponds to backness, this is essentially equivalent to saying that velars have two place targets, one further front than the other. 45 In particular, I make no claims about what the values of [anterior] and [distributed] are. I assume that both laterals are [+anterior] [-distributed]. The back lateral then surfaces with a dental and back velar constriction. The front lateral would phonologically have the representation as a dental lateral with a front velar/palatal constriction. It may surface as post-alveolar (Zimmer & Orgun 1992), but this is explained since the fronted tongue body constriction results in also pulling the tongue blade back from its target as dental. 46 Vowels may also have articulations under the Soft Palate node if they are nasalized or under the Tongue Root node for [ATR]/[RTR] values.

67 there is a dorsal articulation which is required in this segment and that this articulation varies with the environment. The lateral is fundamentally different from the coronals /t/ or /n/ which do not have a tongue body articulation.47 Thus, it seems reasonable to assume that it is not the case that plain coronals surface with a secondary articulation but that the laterals are inherently different from the other coronal segments. This eliminates the representation where the lateral has only the designated articulator [coronal]. Because it is necessary to produce the laterals with a tongue body articulation, it seems reasonable to assume that there is a double articulation in these segments. Therefore, the plain lateral, allowing front and back velar articulations, is doubly articulated [coronal]/[dorsal]. The feature [back] is not represented on this segment because the value of [back] is dependent on the vocalic environment (see preceding discussion of velar stops). The front lateral differs in that it always surfaces as front and therefore has a [-back] specification. Now that the representations of the relevant segments have been established and the spreading rule has been stated, we can move on to looking at how they interact to account for the data. The way suffix vowels undergo harmony is illustrated in (83). (83) /taj-lEr/ [taj-lar] t a j Place Place | | Body Blade


[dor] [+bk] [cor] [-ant] [+distr]

l E r Place Place | | Body Body | | [dor] [dor]

This example also shows the transparency of the glide and the allophony of the lateral. The root vowel /a/ has a contrastive [+back] feature which spreads to any segment with a designated articulator [dorsal]. In this example, the lateral in the suffix and the suffix vowel are both specified with a [dorsal] articulator. The lateral also has features under the Tongue Blade node, but these have been left off of the diagram in order to conserve space. The Tongue Blade node dependents have no effect on the spreading because they are on a different plane. This is also why the glide is transparent to the harmony in (83). It is not specified for [back] ergo [back] can spread across it. This example further illustrates the need for the harmony rule to specify that the target segment have a [dorsal] articulator in order for the lateral and the vowel to undergo it. The example in (84) shows velar allophony, as well as suffix vowel alternations. As in the previous example, segments with no [dorsal] articulator are not represented, though of course they have features. Since only [dorsal] and [±back] are relevant to the


One can contrast Turkish with Russian, where every consonant is obligatorily velarized or palatalized (Rubach 2000 and references therein).

68 harmony, these other segments do not participate. As expected, the segments with a [dorsal] articulator are targets of the harmony. The velar stops surface as either front or back as the result of spreading from the preceding vowel. (84) /san-dIk/ [san-dk] s a n-d I Place Place | | Body Body | [dor] [+bk] [dor] /et-dIk/ [et-tikj] e t-t I Place Place | | Body Body | [dor] [-bk] [dor]

k Place | Body | [dor]

k Place | Body | [dor]

The example in (85) shows that fully specified vowels block harmony. The initial front vowel spreads its [-back] to the initial suffix vowel /I/, but cannot and does not spread to or across /o/. It is only allowed to spread to a segment with a designated articulator [dorsal] that does not have a [back] feature (see (81b). It cannot spread across /o/ to the final suffix vowel because the presence of [+back] blocks spreading. The final vowel gets its backness from the fully specified suffix vowel. (85) /ed-Ijor-Im/ [ed-ijor-um] 'I am doing (do-pres. progr.-1 p. sg.)' e d I j o r I m ... ... ... ... Body Body Body Body [dor] [-bk] [dor] [dor] [+bk] [dor]

The next example deals with the underlyingly front lateral. From the data, it is clear that this segment prevents spreading from the preceding vowel and also initiates a new harmonic domain by spreading its own features to the following targets. Since the spreading rule in (81) explicitly states that contrastive features are visible, it is necessary to further consider the representation of this segment. The alternating laterals and velars have no [back] feature and therefore cannot block harmony. Likewise, the suffix vowels, which also lack a [back] feature in their underlying representation, do not block harmony. The fact that the front lateral does block harmony must therefore be examined. The key to the visibility of [-back] on the front lateral comes from the hierarchy of visibility. As illustrated in Chapter 3, §3.5.1, marked features are a subset of contrastive features which are in turn, a subset of all features. Explicit in this formulation is the notion that marked features are visible to rules that target all, contrastive, or marked features. That is, if a feature is marked, it has the highest visibility and cannot be avoided or made invisible by any rule. The important thing to notice about the underlyingly front lateral is that it is a marked segment. Of the features it has, it is the [-back] feature that is marked. Since

69 coronal laterals are unmarked, the features under the Tongue Blade node are not marked. What makes this segment marked is that it has a fronted dorsal constriction, i.e. [-back].48 Since the feature [-back] is marked in this segment, it is automatically also contrastive, even though it does not actually contrast with another [+back] segment. Therefore, when the spreading rule specifies that only contrastive features are visible, the [-back] from this lateral remains visible and therefore participates in the harmony. The [+back] from /o/ cannot spread across the lateral to the following suffix vowel because it would spread across the visible [-back] of //, resulting in a line crossing violation. In this way, the lateral acts as a blocker to the backness harmony. The [-back] of the lateral can spread to the following suffix vowel, thereby initiating a new harmonic domain. (86) /petro-dE/ [petro-de] p e tr o Place Place | | Body Body [dor] [-bk] [dor] [+bk]



Blade Body | [cor] [dor] [-bk]

E Place | Body | [dor]

Before concluding this section, I would like to address the question of contrastive/blocking velar stops. As discussed in §4.1.2 some descriptions of Turkish discuss velars which contrast in final position. As with VPT, RAT can account for these segments. Front and back velars would simply have a [±back] feature specified in their underlying representation. These features would be contrastive and the segments would therefore act as blockers in the harmony. All of the facts of Turkish backness harmony can be accounted for in RAT. First, the suffix vowels are targets of harmony because they are not specified for [back]. Second, the glides are transparent because they lack the features of the spreading rule. Third, the velars alternate because they have the designated articulator [dorsal] and are not specified for [back] in their underlying representation. Fourth, the allophony of the plain laterals is explained by their also having a designated articulator [dorsal] and being unspecified for [back]. Finally, the underlyingly front lateral, being a marked segment, blocked the spread of harmony and initiated a new harmonic domain. In order to explain these facts, there were several requirements on the representations. The suffix vowels, the velars, and the plain lateral were unspecified for [back] since this feature was completely predictable from the vocalic context. Furthermore, the laterals were


It might also be the case that the other Tongue Body features ([-low], [+high]) are also marked. Whether or not this is the case does not affect the spreading since it is only sensitive to the presence or absence (visibility/invisibility) of [back].

70 represented as doubly articulated segments with both Tongue Blade and Tongue Body features. 4.5 Discussion Of the four representations, only VPT and RAT are able to deal with the varied behavior of consonants and vowels in Turkish backness harmony. Both LEX and [CONS] are insufficient in distinguishing glides from vowels with respect to harmony. VPT and RAT, on the other hand have a richer representation and can therefore distinguish the subtle difference between these segments. Parallel requirements were made in the representations in RAT and VPT. To account for the allophony of laterals and velars, an empty V-place node or a [dorsal] designated articulator was required. The front lateral was represented with a filled Vplace node in VPT and with a double articulation [dorsal]/[coronal] and [-back] in RAT. The transparency of the glide is expected in both feature geometries because it lacks the spreading feature in RAT and the location of the spreading in VPT. However, the difference between RAT and VPT is in the motivation for these representations. In RAT, only segments with a dorsal articulation participate in the harmony. Velars are expected to gain backness features from the neighboring vowels because they are dorsal segments. In VPT, on the other hand, it must be stipulated that velars have an empty V-place node. In RAT, the reason coronals such as [t] and [s] do not acquire backness values is because they have no designated articulator [dorsal]. But in VPT, their transparency is part of the stipulation that they do not have an empty V-place node.


5 Pulaar

5.1 Introduction The data in this chapter come from the Western Pulaar dialect of Kaedi Pulaar (or Futankoore Pulaar) spoken in the Futa Toro area of Mauritania. Pulaar is a dialect of Fula, a language of the Niger-Congo family (Paradis 1992: 5). When other dialects from other sources are discussed, it will be explicitly stated. The Fula dialects are known for their consonant gradation. Pulaar shows two patterns of glide alternations, j~c/w~b and j~g/w~g. This chapter will show that the latter pairs are derived via assimilation from vowels while the former are underlying. Evidence for the underlying glides will also come from their transparency to ATR harmony, from epenthesis, and from gemination. Pulaar is a Type VI language, where some surface glides are derived from vowels, while others are underlying. Pulaar is slightly different from other Type VI languages because the derived glides are the result of spreading features, not of a separate segment that alternates. (87) Type VI /i/ /j/ [i] [j] /u/ [u] /w/ [w]

Pulaar has the five phonemic vowels listed in (88). The two high vowels are [+ATR], while the remaining ones are [-ATR] (Paradis 1992: 84). On the surface, two additional [+ATR] vowels [e, o] exist as the result of vowel harmony. All of the vowels can appear long or short. (88) i a The surface consonant inventory is listed in (89). The two segments [c ] are actually the affricates [t, d], but I follow Paradis 1992 in using the former symbols. Paradis 1992 argues that the prenasalized stops are not phonemic (pp. 109-119). All vowel initial words are realized with a [] (Paradis 1992: 104, Diop 1993: 39), thus Paradis 1992 argues that [] is not phonemic. u

72 (89) Consonant Inventory Labial Oral stops pb Implosives Prenasalized (mb) Nasals m Fricatives f Liquids Glides w Dental td (nd) n s lr Palatal49 c () Velar kg (g) h j () Glottal ()

The rounded glides are in complementary distribution, with [] surfacing before front vowels and [w] elsewhere, as shown in (90). Because their distribution is regular, I will follow Paradis in not marking the difference in the remaining data.50 (90) (Paradis 1992: 107) cc- 'chest' ib- 'wing' w-r 'hare' wukk-uru 'pompom' waa-r 'beard'

Pulaar is probably best known for its consonant alternations. Consonants alternate between several different grades: continuant, stop, or nasalized. Two aspects that affect this gradation are the lexical stratum of the word and its class marker. These will be described more extensively in the next section. The phonology of Pulaar is divided into two lexical strata. Nominal morphology is part of Stratum I, while verbal morphology (including deverbal nouns) is part of Stratum II (Paradis 1992: 124). We will see in the following sections that different phonology applies to these strata. There are approximately 25 class marker suffixes. These class markers are what determine the grade of the initial consonant. Syllables never contain branching onsets or codas. Geminates do occur, but only intervocalically. 5.2 Basic Pulaar Phonology This section will deal with four properties of Pulaar phonology which will be necessary to understanding the evidence for underlying glides in §5.3. The first section deals with a

49 50

In Paradis & Prunet 1989a, this column is listed as "Alveopalatal". Paradis 1992 does argue that the inchoative suffix is underlyingly the front /-/, though this seems to be the only place where it is underlying.

73 sonority constraint. It will show that the palatal glide [j] patterns as a coronal segment. This sonority requirement is also necessary to understand why (suffix) marker shortening does not apply in all cases. The section on (suffix) marker shortening is included primarily to illustrate the types of suffix alternations that are possible. 5.2.1 Sonority Constraint Pulaar has a place dependent syllable contact constraint. Across a syllable boundary, segments must fall in sonority if they are of the same major place. This sonority constraint applies only at Stratum I. Presumably (91) would apply to all major places of articulation, but evidence from labial and dorsal segments is lacking. (91) Sonority Constraint In a cluster of two coronal consonants, the first consonant must be more sonorous than the second one. (Paradis 1992: 124) The sonority hierarchy of Pulaar is given in (92). To illustrate the hierarchy, the list of possible and impossible cross-syllable clusters is presented in (93), where O represents a coronal obstruent and N a coronal nasal. From (93) it is clear that obstruents cannot be followed by other segments because the result would be either a plateau or a rise in sonority across the syllable boundary. By contrast, a glide can be followed by any other coronal consonant because the result is always a fall in sonority. (92) V > Glides > r > l > Nasals > Obstruents (Paradis 1992: 124, Paradis 1988: 14) (93) Glide + C: Rhotic + C: Lateral + C: Nasal + C: Obs + C: Possible Clusters jO, jN, jl, jr, rO, rl, r-N lO, lN NO ---Impossible Clusters ---*rj, r-r51 *lr, lj *Nl, Nr, Nj *dl, Or, Ol, ON, OO

The sonority constraint only applies to segments with the same major place of articulation. In (94), the labial nasal can be followed by [r], whereas the coronal nasal cannot. In these examples, the class marker alternates between [d] and [r] depending on the preceding segment (see §5.2.2 for more information). With respect to major place, the segments listed in the Palatal column of (89) count as coronal. For example, *cr, *r, *lj, *nj, and *sj are not allowed at Stratum I.


I have added r-r, as it is implied in Paradis (1992: 129).

74 (94) am-r an-d 'turtle' 'syllable'


5.2.2 Marker Shortening This section provides a brief description of marker shortening. The data and analysis in this section do not provide independent evidence for underlying glides but do provide some information that will be used in the discussion of the unmarked segment in Pulaar (see §5.3.1). The analysis of marker shortening in this section comes from Paradis 1992. As mentioned above, Pulaar has approximately 25 class marker suffixes. The choice of which marker a stem takes is arbitrary; each stem is subcategorized for the class marker (Paradis 1992: 22). While the choice of marker is arbitrary, Paradis shows convincingly that the form of that marker is not. The class marker suffixes have often been described as having four unpredictable grades: zero grade, continuant grade, stop grade, and nasal grade. Paradis 1992, however, shows that there are actually only two grades for these suffixes (strong and weak) and that their realization is predictable. The nasal grade can be quickly eliminated as a separate alternation once it is realized that this fourth grade is simply the result of a floating nasal found on all perfective agentive constructions and on participial adjectives (Paradis 1992: 74-5). The floating nasal is also found at the end of a few stems (approximately 2.5%) (Paradis 1992: 79). According to Paradis 1992, the strong form of the class markers is underlying. In (95), gV refers to the markers [g, g, gu, gel, gal, gol], kV refers to [ka, ki, k, kal, kn], V to [, i, am, um, ], V to [], bV to [ba], and dV to [d, di, du] (Paradis 1992: 70). (95) Strong Variant: Weak Variant: gV V kV V V V V V bV V dV rV

At Stratum I, a process of marker shortening occurs where the first slot of the marker is deleted. At Stratum II, there is no shortening and all stems take the strong variant. Marker shortening only applies at Stratum I when the stem ends in a consonant. Where the stem ends in a vowel, no shortening occurs.52 This is not surprising since the shortened form would result in vowel hiatus (at least for most of the variants) and hiatus is banned at Stratum I. The rule is formalized in (96) where C is the initial consonant of the marker and S is a non-nuclear segment. Some examples of the shortened markers are given in (97).


"Vowel-final stems always select a strong marker." (Paradis 1992: 79)

75 (96) Marker Shortening (Stratum I) skeletal tier X - X | (|) segmental tier S C Strong dV (Paradis 1992: 79) X - Ø | S C


kV V gV V

Shortened (Stratum I) (Paradis 1992: 78) lew-ru 'moon' gaw-ri 'millet' nw-r 'palm' gaw-al 'millet (dim)' caal-i 'shelter' bajl- 'blacksmith' krl- 'hocks' maaj- 'river' laac-el 'tail (dim)' wuj- 'thief' af- 'elder child'

The two markers dV and V require further explanation. The weak forms of these markers still contain an initial consonant. According to Paradis 1992, these two markers contain an extra slot. In the weak form when the slot is deleted, there are still enough slots for the remaining segments, allowing the initial consonant to surface. (98) XXX r V XXX V

According to Paradis, when the shortening rule does not apply, the /r/ spreads onto the remaining slot. However, continuant geminates do not occur in Pulaar. When continuant geminates are expected to occur (see §5.3.4), they are replaced by [-continuant] segments. The [-continuant] form of /r/ is [d].53 Neither of the markers in (98) ever actually surface as geminate. There is one place where marker shortening does not apply to dV markers and that is when the resulting [rV] would violate the sonority constraint from §5.2.1. It is clear that these forms in (99a) are in Stratum I, because when other markers are added as


Paradis argues that the process is for /r/ to geminate, occlusify, and then subsequently degeminate. It is not clear why the degemination occurs in vowel-final stems which take the strong marker. After a vowelfinal stem, a geminate would be allowed because it would be intervocalic. That degemination applies after consonant-final stems is not surprising given the ban on consonant clusters. I do not have an explanation for this.

76 in (99b&c), they surface in their weak form. The markers in parentheses in (99b&c) show the strong/underlying form. (99) a. Singular b. Plural c. Diminutive fal-d (*fal-r) pal- () pal-el (gel) ton-du (*ton-ru) ton-i (i) ton-el (gel) Gloss (Paradis 1992: 77) 'river bank' 'lip'

The final grade that needs to be explained is the so called continuant grade. Paradis 1992 argues that this is not a separate grade, but is the combination of the inchoative suffix /-w/ or the paradigm suffix /-j/ which trigger the shortening rule, creating what looks like markers that begin with a glide. In this way, Paradis has succeeded in reducing the four grades to two and shown that they are fully predictable. One final place where marker shortening does not apply is following "nuclear diphthongs". These cases make up only about 5% of words. Here, sequences of a vowel plus a sonorant do not cause the expected marker shortening. Paradis 1992 explains this lack of shortening by arguing that these sequences constitute nuclear diphthongs. Because they are nuclear elements, they do not require the marker shortening. These forms will be relevant to the discussion of vowel harmony in §5.3.2. As far as I can tell, there is no independent evidence for the argument that these are nuclear diphthongs. They could also be analyzed as stems which exceptionally take the strong variant. (100) caw-gu tr-gal lam-d jl-d da-ki *caw-u *tr-al *lam- *jl- *da-i 'skin' (Paradis 1992: 78) 'limb' 'vagina (vulg.)' 'gap between two teeth' 'stool'

5.3 Evidence for underlying glides In Pulaar, there are three types of evidence that show the existence of underlying glides. The most well-known examples come from consonant gradation where stem initial glides alternate with different consonants depending on whether these glides are underlying or derived. The consonant gradation in the eastern dialects of Fula has been discussed extensively in the literature. For this reason, a brief section on these dialects is included in § The comparison of the representations in 5.3.4 will only discuss the Western dialect of Pulaar described for Paradis. Evidence will also be presented from vowel harmony and vowel epenthesis. 5.3.1 Consonant gradation Pulaar (Western dialect) Basics of gradation

77 Pulaar is famous for its consonant gradation, which occurs at the beginning of adjectives, nouns, and verbs, and also at the end of stems (Paradis 1992: 38). Only the gradation that begins Stratum I nouns will be discussed here. Consonant alternations can involve up to three grades. It is the class marker suffix that determines the grade of the initial consonant. Effect 1 markers cause the initial consonant to be [-continuant]. Effect 2 markers add prenasalization to the initial consonant if it is a voiced stop. Effect 0 markers cause no change in the initial consonant. Paradis argues that the nasal grade is not a separate grade, but is simply the addition of a nasal prefix (Paradis 1992: 51). The nasal prefix cannot surface unless the following segment is a voiced stop, which explains the lack of nasalization on [c] or [k]. A schematized picture of the gradation is given in (101). (101) Stratum I noun consonant gradation (Paradis 1992: 40) rV sV jV w/wu Effect 0 wV j/i a | dV cV V gV Effect 1 bV (n) () () dV cV V gV Effect 2 (m)bV hV V

kV kV


Several examples that relate to glides are presented in (102). In these examples, the segment of interest is word initial. The suffixes are either singular, plural, or diminutive. (102) Effect 0 a. w ~ b w-r wukk-uru wib- wil-d waad-r b. w ~ g wr- wur- woot-uru wuuf-r Effect 1 b- bukk-i bib-el bil- baad- Effect 2

m m

b-n bukk-n

'hare' (Paradis 1992: 41) 'pompom' (Paradis 1992: 41) 'wing' (Paradis 1987b: 329) 'trap' (Paradis 1987b: 329) 'drop (of rain)' (Paradis 1987b: 330)

gr-k- gur- gott-um guuf-

gr-n gur-n

'man' 'village' 'same' 'mouthful'

(Paradis 1992: 41) (Paradis 1992: 41) (Paradis 1987b: 331) (Paradis 1987b: 331)

78 c. j ~ j-r jim-r ju- jamir--r d. j ~ g jiit-r jert-r ji- jnaa-nd e. Ø ~ g allaa-du au-g abb-r am-r f. Ø ~ Ø en-du ullun-du af- in-d r-nd

e-el im-el u- amir--- giit- gert- gi- gnaa-l

-n im-n

'seed' 'poem' 'system' 'order'

(Paradis 1992: 41) (Paradis 1992: 41) (Paradis 1987b: 330) (Paradis 1987b: 330)

giit-n grt-n

'eye' 'peanut' 'friend' 'tomb'

(Paradis 1992: 42) (Paradis 1992: 42) (Paradis 1987b: 331) (Paradis 1987b: 331)

gallaa-i gau-l gabb-el gam-


'horn' 'cheek' 'seed' 'turtle'

(Paradis 1992: 36) (Paradis 1992: 36) (Paradis 1992: 42) (Paradis 1992: 47)

en-i ullu-i af- ill- r-

n-kn ullu-kn

'breast' 'cat' 'eldest' 'name' 'herd'

(Paradis 1992: 42) (Paradis 1992: 42) (Paradis 1987b: 336) (Paradis 1987b: 336) (Paradis 1987b: 336)

There are a few important aspects of the alternations in (101). First, the segments which alternate are all [+continuant], and alternate with a [-continuant] consonant. The remaining consonants (stops, affricates, nasals, and l) are all [-continuant] and do not show any alternation in word-initial position. Second, the basic pattern is for [w] to alternate with [b], for [j] to alternate with [], and for vowel-initial words to alternate with vowel-initial words (Paradis 1987b: 330).54 Third, the only time when a glide alternates with [g] is when the following vowel is similar (i.e. w with /u, / or j with /i, /). This is of course not a sufficient condition since [j] alternates with [] before all vowels including [i, ], and [w] can alternate with [b] before [u, ].


Paradis also assumes that b~w and ~j (as opposed to w/j~g) are basic because w and j never alternate with [g] in other contexts, e.g. gemination.

79 The problem is how to explain the different behavior of the two stem initial glides. What explains why some jV sequences alternate with [g] and others with []? Or wV with [g] or [b]? Similarly, what explains why some a-initial stems alternate with [g] and others with Ø? Paradis's analysis Paradis operates under the assumption that there is no distinction between glides and vowels in the underlying representation (Paradis 1987b: 332). She assumes explicitly that there are no underlying glides in Pulaar. In order to account for the varied alternations, she must assume three things (Paradis 1987b, 1992). First, she assumes that the words which alternate with [g] have an empty segmental slot word-initially. Second, she must assume that the default segment in Pulaar is velar. Third, she must assume that the empty slot is specified as [+voice]. What follows is a brief sketch of Paradis's analysis. Words like [wur-]/[gur-] 'village', [jiit-]/[giit-] 'eye', and [am-]/[gam-] 'turtle' have the underlying forms /_ur-/, /_iit-/, and /_am-/ respectively. In the continuant grade, an epenthetic [] is expected to surface in the empty slot because it is (a) velar, (b) default continuant, and (c) voiced. Because this segment is not allowed in Pulaar, however, Paradis argues that in the continuant grade, this segment cannot surface and instead, there is feature spreading from the vowel onto the empty slot.55 Underlying /a/ cannot spread onto the empty slot because it does not contain the 'charm' or particle I or U and therefore the slot remains empty.56 In the stop grade, on the other hand, [g] is expected to surface because it is (a) velar, (b) voiced, and (c) [-continuant]. In the noncontinuant forms, velar epenthesis is required to fill the empty slot (Paradis 1987b: 334). The major problem with this analysis is that it rests on the assumption that velar is the default place in Pulaar. Abundant evidence exists, however, for the coronals as the unmarked segments. I will now provide a brief excursus on the unmarked status of coronals in Pulaar. Unmarked Place in Pulaar Paradis & Prunet 1989a&b provide four examples in which vowels can spread across coronal consonants. The first type of evidence comes from the imperfect paradigm in (103). I1-4 represent four sub aspects within the imperfect. Given the symmetry of this paradigm, Paradis & Prunet suggest that the active, middle, and passive suffixes as well as the four imperfect suffixes can be reduced to those in (104). The active marker is simply a segment. The Middle and Passive markers are vowels prelinked to a segmental slot. I2 and I4 are only slots, while I3 has a segment and a slot that are not linked. I1 is null.

55 56

Though Pardis 1992 and 1987b use charm/particle phonology, the result is the same. This slot could also remain empty due to a ban on low glides. In either case, the result is the same.

80 (103)Imperfect Paradigm I1 I2 I3 I4 Active --a -at -ata Middle o | X Voice Middle -o -oo -oto -otoo Passive -e -ee -ete -etee I1 I2 X I3 t X I4 X Imperfect

(104) Active a

Passive e | X

In order to get an imperfect marker, all of the preceding imperfect suffixes are added together. For instance, I4 is formed by all of I1, I2, I3, and I4. To create I4-active [-ata], consider (105). Here, the active, I1, I2, I3, and I4 are all added together to create the full I4-active form. In order to create the output, association occurs one-to-one and left-to-right. The vowel spreads to the first slot and the consonant to the second. Finally, the vowel is allowed to spread onto the remaining X-slot because the intervening consonant is coronal and underspecified.57 (105) a Ø XI2 t XI3 XI4 a Ø X t X X

The second piece of evidence for the unmarked status of coronals in Pulaar comes from rV/dV markers. As shown in the previous section, these markers are analyzed as having an underlying /r/ and an extra slot. In addition to the variants [rV] and [dV], this class of markers also has VrV variants where the first vowel is a copy of the second. Paradis & Prunet (1989a) assume the underlying representation in (106). (106) X X X | | r V


Paradis & Prunet 1989a consider coronals to be underspecified by assuming that they do not have a place node. I am not taking a stance about the validity of underspecification at this point. Here, it suffices to see that coronals have a special status as unspecified or default segments. See Steriade 1995 for arguments against underspecification.

81 When the stem ends in a coronal obstruent, these markers surface as VrV (see (107)). Both of the alternatives [dV] and [rV] violate the requirement that sonority must fall across a syllable boundary. When the stem ends in a cluster, the VrV variant is also used. Paradis and Prunet argue that the vowel can spread across the underlying /r/ because it is coronal. (107) (Paradis & Prunet 1989a: 330) w-r *w-r *w-d woot-uru *woot-ru *woot-du kes-iri *kes-ri *kes-di

'hare' 'unique' 'new'

Both the rV/dV and the V markers were analyzed as having an extra slot in Paradis 1992. The extra slot at the beginning of V markers explained their apparent immunity to the shortening rule. Despite this extra slot, V markers never surface as VV. When a stem ends in two consonants and the addition of the V marker would result in an unsyllabifiable sequence of three consonants, we do not find VV variants. Paradis & Prunet argue that the lack of VV variants is due to the fact that the vowel cannot spread over a labial segment exactly because it is not coronal. Instead, stems which end in two consonants and take V markers require epenthesis. (108) /wajl-/ /lf-/ /awl-/ wajl-u- olf-u- awl-u- *wajl- *lf- *awl- 'blacksmith' 'Wolof' 'griot'

The third piece of evidence comes from epenthesis. Pulaar has two epenthetic vowels, [u] and [i]. [u] surfaces at the end of a verbal domain, while [i] surfaces before a verbal suffix (Paradis & Prunet 1989a). (109) a. Domain final: [u] (Paradis & Prunet 1989a: 335) o-u *o-i 'cough!' ekk-u *ekk-i 'teach!' b. Before verbal suffix: [i] o-i-n *o-n-u o-i-t *o-t-u

*o-u-n *o-u-t

'make cough' 'cough again'

In (109a), only the domain-final [u] can be epenthesized because there is no verbal suffix. These forms contrast with those in (109b) which require epenthesis of [i] before the verbal suffix. It is impossible to use domain-final [u] in these cases because they would

82 result in an illegal sequence of three consonants (e.g. onu). When either epenthetic vowel could be used, [u] is selected, showing a preference for [u] over [i]. (110) /wr-/ /as-t/ /aam-t/ wor--u as-t-u aam-t-u *wor-i- *as-i-t *aam-i-t 'become infected' 'dig again' 'eat again'

In spite of the general preference for [u] epenthesis, when two epenthetic vowels must occur, there is a preference for those vowels to be the same. The explanation offered by Paradis & Prunet (1989a) is that spreading an epenthetic vowel across a coronal is less costly than inserting an entirely new segment. (111) /utt--t/ utt-i--i-t 'become fat again!'58

The final piece of evidence for the unmarked status of coronals is vowel assimilation across coronals. In both pronouns and imperfect plus object clitic sequences, vowel assimilation occurs across the coronals [t], [n], and [] (Paradis & Prunet 1989b: 333). Unfortunately, no examples are provide in Paradis & Prunet that show that assimilation is blocked elsewhere. (112) a. Pronouns59 a-en een 'we (incl.)' a-on oon 'you (pl)' on-en onon 'you (independent pl.)' b.-Imperfect - object clitics -at-en eten '1 pl. inclusive' -at-on oton '2 pl' -at-e ete '2 sg.' In fact, in Paradis's analysis, coronals they are actually underspecified for a place node. Thus, coronals are clearly the unmarked segments in Pulaar and would be expected in epenthesis situations. Summary This section has provided the basic data involved in consonant gradation in the western dialect of Pulaar. It was shown that Paradis's analysis rests on the notion that velars are

58 59

Another possible output is [utt-i--t-u], but is less preferred. In Paradis & Prunet 1989b, all mid vowels are transcribed with <e> and <o> symbols for ease of typing. These data make no claim as to the ATR status of mid vowels in the language.

83 the unmarked segment in Pulaar. However, abundant data shows that coronals behave as unmarked. Assuming that both velars and coronals are somehow unmarked or default is flawed. An additional problem with her analysis is why the empty slot should be specified as voiced when cross-linguistically obstruents, especially velars, are voiceless (Maddieson 1984). One part of Paradis's analysis, however, is very insightful. I will adopt her notion that the class of words that have [g]-alternations contain an empty slot in the underlying representation. My full analysis will be presented in §5.3.4. Eastern dialects Other analyses of consonant gradation in Fula exist, but most have dealt with the eastern dialects of Gombe and Adamawa. In these dialects, there are cases where [w] alternates with [g] or [b] before [a] as in (113). (113) war-/(bar-)/mbarwar-/(gar-)/gar'kill' (Anderson 1976a: 118) 'come'

Anderson (1976a&b) argues that the possible differences are linked to the fact that [w] is a multiply articulated segment. He argues that multiply articulated consonants, such as labio-velars, behave differently cross-linguistically, some acting as part of the labial series, others as part of the velar series. He argues that in Fula, some of the [w]s are primarily labial and therefore alternate with [b], while others are primarily velar and therefore alternate with [g]. This stance is also espoused in Sagey's 1986 account of consonant gradation in Fula. She draws on the data and analysis of Anderson and argues that the two different [w]s are the result of different major and minor articulators. She treats both of these glides as underlying segments. Sagey argues in favor of this analysis, and against an analysis that brings in historical [], pointing out that in these dialects alternations with [j] have regularized, but the labio-velar glide has not, exactly because it is multiply articulated. Halle 1995 also sketches an analysis of the eastern Fula dialects. He assumes that there are two sources for the surface [w]s. In one case, the glide is underlying and therefore labial and in the other it is derived and therefore dorsal. Halle claims that vowels are underlyingly dorsal, while /j/ is coronal and /w/ is labial. Summary The analysis of consonant gradation that I will propose in §5.3.4 will only deal with the Pulaar data. My analysis will draw on Paradis's insight that there is an empty segmental slot at the beginning of some stems. It will also explain why the segment that surfaces is (a) voiced and (b) velar, but will not rely on the assumption that velars are default. My analysis will also rely on the notion that some glides in Pulaar are underlying while others are derived, similar to Halle's account of the Eastern dialects. My analysis differs from Halle's in providing evidence for the different status of the two surface glides.


5.3.2 Vowel Harmony Vowel harmony in Pulaar is analyzed by Paradis 1992 as ATR harmony.60 The harmony spreads [+ATR] from suffix vowels leftward to the preceding suffixes and stems. In general, harmony is triggered by [+ATR] high vowels. There are also two classes of markers with [+ATR] mid vowels that cause harmony, /-gel/ and /-gol/.61 (114) shows examples of harmony triggered by suffixal [+ATR] vowels. Forms without harmony are included for comparison. (114) [+ATR] sof-ru ser-du peec-i dog-oo-ru [-ATR] cf-n cr-kn pc-n dg--w-n Gloss 'chick' 'rifle butt' 'slits' 'runner'

(Paradis 1992: 87)

The forms in (115) confirm that harmony does not propagate rightwards. (115) Directionality hel-ir-d *hel-ir-de dill-r *dill-ere *fu-ere fu-r

'to break with' 'riot' 'pimple'

Harmony in Pulaar is blocked by the low vowel /a/. If the low vowel intervenes between a mid and high vowel, the mid vowel does not surface as [+ATR]. (116) bt-aa-ri ndd-aa-li *boot-aa-ri *nodd-aa-li 'lunch' (Paradis 1992: 88) 'call'

Furthermore, the epenthetic vowels [i] and [u] trigger the harmony. (117) /dkk-n-d/ /lrl-d/ [dokk-i-n-d] 'to make s.o. give' [lorl-u-d] 'to torture' (Paradis 1992: 18)62 (Paradis 1992: 128)


Though Paradis analyzes Pulaar harmony as ATR, an analysis that follows Parkinson 1996 might better explain the data. According to Parkinson, the blocking effects of [a] cross-linguistically are due to a constraint that bans height features on placeless (central) vowels. 61 Paradis shows that these suffixes are only two of five morphemes which contain underlying [+ATR] mid vowels. 62 /n/=causative, /d/=class marker

85 Glides neither trigger nor block ATR harmony. In (118a), the glide does not cause the preceding vowels to surface as [+ATR]. In (118b), the glide does not block the harmony from the suffix vowel. (Here the underlying marker is /-gel/ and can cause harmony.) (118) a. Non-trigger dg--w- cj-kn hw hw dw-al cj- t-j-n lt-w-n b. Non-blocker oto-j-el looto-w-el kejn-iraa-gel

*dog-oo-w- *cooj-kn *hoow *how *dew-al *cooj- *oto-j-n *looto-w-n

'one who runs' 'spinster (dim. pl.)' 'to copulate' ' to enclose' 'fat woman' 'bachelor spinster' 'small cars' 'small washers'

(Paradis 1992: 35) (Paradis 1992: 90) (Paradis 1992: 93) (Paradis 1992: 93) (Paradis 1992: 27) (Paradis 1992: 29) (Paradis 1992: 1) (Paradis 1992: 1)

*t-j-el 'small car' (Paradis 1992: 1) *lt-w-el 'small washer' (Paradis 1992: 1) *kjn-iraa-gel 'brother in law (dim)' (Paradis 1992: 100)

Recall that Paradis 1992 argues that some VR sequences (where R is a sonorant) are treated as nuclear diphthongs. Even in these nuclear diphthongs, glides do not trigger harmony, nor do they block it. According to Paradis 1992, these all contain nuclear diphthongs since they do not have the shortened class marker. If her analysis of nuclear sonorants is correct, the data in (119) show that harmony in Pulaar is not a nucleus to nucleus relation, but purely feature spreading. (119) a. Non-trigger kj- *koj- *peew- pw- b. Non-blocker ow-i *w-i cew-i *cw-i

'feet' 'person'

(Paradis 1992: 173) (Paradis 1992: 173)

'mosquitoes' 'panthers'

(Paradis 1992: 173) (Paradis 1992: 173)

If glides in Pulaar were all derived from vowels, then we would expect them to participate in the harmony in some way since they would be featurally identical to vowels. Even if one were to appeal to a model of harmony where only nuclear elements

86 participate, the fact that the so-called nuclear diphthongs do not participate is problematic. Recall from Chapter 4 on Turkish that the analysis of harmony as nucleusto-nucleus spreading was dismissed. If on the other hand, the glides in Pulaar are underlying and therefore different from vowels, their transparency can be explained. 5.3.3 Vowel Epenthesis Most of the basics of vowel epenthesis were discussed in § What is relevant to the present discussion is the fact that the inchoative suffix /-w/ causes epenthesis.63 If there were truly no distinction between underlying glides and vowels, as Paradis 1992 assumes, then there would be no impetus to epenthesize a vowel since the inchoative suffix would be /u/ and basic syllabification could create the structure in (120). The actual output, however, is [al-wu-d] with an epenthetic vowel. (120) Predicted output, assuming underlying glides do not exist /al-u-d/ [al-u-d] 'black-inchoative-marker' However, if the inchoative suffix is actually an underlying glide, then the need for epenthesis is explained. Recall that the syllable structure in Pulaar bans consonant clusters and requires epenthesis (see (121)). The forms in (122) with the inchoative suffix are analogous to (121) and require epenthesis to rescue the unsyllabifiable cluster. It is not possible to simply vocalize the glide. (121) /lrl-d/ /wrs-d/ (122) /al-w-d/ /ran-w-d/ /naj-w-d/ /tun-w-d/ lorl-u-d wors-u-d al-w-u-d ran-w-u-d naj-w-u-d tun-w-u-d *lorl-d *wors-d *al-u-d *ran-u-d *naj-u- *tun-u-d 'to torture' 'to knot' 'to blacken' 'to whiten' 'to grow old' 'to dirty' (Paradis 1992: 128) (Paradis 1992: 128) (Paradis 1992: 196) (Paradis 1987a: 131) (Paradis 1987a: 131) (Paradis 1987a: 131)

5.3.4 Gemination This section shows the possible geminates in Pulaar. In particular, we will see that w and j geminate as bb and . Geminates in Pulaar are always [-continuant] (Paradis 1988: 34).64 Continuant geminates (*ff, *ss, *rr, *ww, *jj, *hh) are completely absent.65 In


Paradis 1992 actually assumes that the inchoative suffix is /-/, but this does not affect the discussion because it is still a glide. This is the only place where she assumes this segment underlyingly. 64 Allowable geminates are: pp, bb, tt, dd, cc, , gg, , , , mm, nn, , ll. Here, the lateral patterns with the [-continuant] segments, as it does with consonant alternations.

87 Pulaar, V markers cause gemination of the preceding segment. When the stem-final segment is [+continuant], it changes to a [-continuant] segment as the result of gemination. The examples in (123) show continuants alternating with noncontinuant segments when geminated. (123) Other marker caw-el dew-el wuj- kof-el ks-am -V (shortened) cabb-i dbb- gu- kopp-i kcc- Gloss (Paradis 1992: 172) 'stick' 'woman' 'thief' 'knee' 'milk'

When the preceding vowel is long, the segment does not geminate, but it does occlusify.66 (124) Other marker iiw-a maaj- ls- V (shortened) iib-i *iibb-i maa- *maa- lc- *lcc- Gloss (Paradis 1992: 173) 'elephant' 'river' 'bed'

In gemination contexts, the glides [j] and [w] only alternate with [b] and [] respectively. They never alternate with [g], as they do in word initial position (see §5.3.1). The analysis I will adopt accounts for the lack of w/j~g alternations in final position. I will show that the initial alternations are due to vowel spreading. In final position, there are no empty slots and therefore no need to spread. Instead, all non-initial glides are true underlying glides and therefore alternate with b/. 5.4 Comparison of the representations The two characteristics of Pulaar phonology that test the four models are consonant gradation and vowel harmony. In particular, the fact that some glides alternate with velars, while others alternate with labials or coronals requires explanation. In harmony,


Niang 1997 argues that there are continuant geminates in Pulaar. However, the words that he uses as examples either have different segments in Paradis's 1992 lexicon or do not have geminates. I must therefore conclude that though the dialects described by Niang and Paradis are clearly similar in some ways, they differ crucially in others. For example, Niang lists [llajjo] 'sacrifice!' and [awwu] 'catch fire!' but in Paradis 1992, the root for 'fire' is /ajn/ ('sacrifice' is absent). In the entry for this word, there is no geminate glide. 66 Long vowels do occur in closed syllables, but Diop 1999 argues that CVVC is never derived (pg. 123).

88 transparency of the glides and blocking of /a/ must be explained. A summary of the relevant cases is given in (125). (125) a. Consonant alternations /w/ wukk-uru ~ bukk-i 'pompom' /j/ jim-r ~ im-el 'poem' /a/ af- ~ af- 'eldest' b. /a/-blocking bt-aa-ri *boot-aa-ri

/_u/ wur- ~ gur- /_i/ ji- ~ gi- /_a/ am-r ~ gam-

'village' 'friend' 'turtle'


Epenthesis and other syllable-based processes can be explained in all four approaches equally because all that is necessary is to separate underlying glides from underlying vowels. 5.4.1 LEX The representation of underlying glides is problematic in LEX for both consonant gradation and harmony in Pulaar. The representations of two sample forms are provided below. In LEX the only difference between underlying vowels and glides is in the suprasegmental structure, where underlying glides are not allowed to surface as nuclear. The features of the initial glides in (126) and (127) are identical since they are both derived from /u/. The form in (126a) would begin with two featurally identical segments. The first segment is barred from surfacing as nuclear and therefore surfaces in the onset. The form in (126b) contains an empty slot that is filled by the features of the following vowel. (126) Underlying representation in LEX N * a. / u u kk-/ b. /_ur-/ In both cases, the features of the initial segments are the same following spreading in (126). Since their features are identical, it is not possible to explain why the first form alternates with [b] while the second with [g]. One could argue that it is something about the way in which the segments are linked. That is, in the first form, the surface glide has its own set of features, whereas in the second, the features of the glide are derived from spreading features from the vowel as shown in (127). There is, however, no motivation for saying that multiply linked features must alternate with velars, while singly linked ones must not.

89 (127) Surface form N · · · | | | u u kka. [wukk-] ~ [bukk] · N · · | | u r­ b. [wur-] ~ [gur]

In terms of the harmony, it is also difficult to explain why the epenthetic high vowel in [lorl-u-d] causes the initial vowel to surface as [+ATR] while the glide in [dw-al] does not since the underlying glide would featurally be /u/. If the only difference between these segments is where they are associated in the syllable and if their features are identical, then they should act together in the harmony, given our earlier argument that all non-stress dependent harmonies are feature-driven and do not involve nucleus-to-nucleus spreading. 5.4.2 [CONS] In the [CONS] model, the only difference between underlying vowels and glides is in their respective values of the feature [consonantal]. The place features, however, remain the same. For the consonant gradation, we are still left to wonder why a [+consonantal] vocoid (glide) should alternate with labials or coronals, while a [-consonantal] vocoid (vowel) should alternate with velars, since both segments would otherwise have the same features in [CONS]. As far as harmony goes, it is possible for an SPE-style rule using stacked features to differentiate between vowels and glides. In (128), mid vowels become [+ATR] before [+ATR] vowels. The blocking of /a/ is implicit in the rules because it states that there may be consonants (including glides) which intervene between the target and trigger, but not vowels. (128) ­cons [+ATR] / ______C 0 -cons -high +ATR -low 5.4.3 Vowel-Place Theory Explaining the full range of consonant gradation in Vowel-Place Theory (VPT) is a challenge. Underlying /w/ has [labial] and [dorsal] features under C-place, while /j/ has a [coronal] specification under C-place. It is not surprising that these segments alternate with a labial and coronal [-continuant] in the consonant gradation environment.

90 (129) a. /w/ Root | C-place labial dorsal b. /j/ (Simplified trees) Root | C-place | coronal

The problem arises when features from the front vowels are spread to the empty slot. In VPT, front vowels are considered coronal. (130) a. /u/ Root | C-place | Vocalic | V-place labial dorsal b. /i/ (Simplified trees) Root | C-place | Vocalic | V-place | coronal

In the [+continuant] grade, we expect to find a glide in this position, as in fact we do. In the [-continuant] grade, on the other hand, we expect a [coronal] non-continuant to surface. Instead, the velar [g] surfaces. Why the velar alternates with a supposedly [coronal] front vowel is not easy to explain. Without altering the representation of vowels in VPT, it is not possible to account for this alternation. The ATR or raising harmony in Pulaar can be explained using VPT with a few stipulations. In the geometry outlined in Clements & Hume 1995, the feature [ATR] is not present. Instead, all height features are assumed to be multiple occurrences of the feature [open] under the aperture node. Parkinson 1996 argues that the feature under the aperture node should be multiple occurrences of the feature [closed]. Adopting Parkinson's model of height, the aperture nodes for the vowels of Pulaar would be as in (131). According to Parkinson, a lower level of [closed] spreads to the mid vowels and raises them to the upper-mid or [+ATR] counterparts.

91 (131) i/u ... Aperture | [closed] | [closed] | [closed] e/o ... Aperture | [closed] | [closed] / ... Aperture | [closed] a67 ... Aperture

The harmony would progress as in (132). In (132a) the lowest [closed] feature spreads to the mid vowel, which then surfaces with two [closed] features as an upper mid vowel. Similarly, in (132b), the lowest [closed] spreads from the upper mid vowel to the preceding lower mid vowel. (132) a. Aperture | [closed] ... i Aperture | [closed] | [closed] | [closed] [e...i] b. Aperture | [closed] ... e Aperture | [closed] | [closed] [e...e]

Finally, it is necessary to explain the blocking effect of the low vowel. The analysis here is based on the insights in Parkinson 1996. In particular, we must explain why the vowel /a/ does not surface as [] or [] in the raising environment. Furthermore, it blocks the spread of raising across it. It is necessary to explain why the situation in (133) is banned. It is not possible for a later vowel to spread to the /a/ and then to a preceding mid vowel as in (133a) because /a/ does not undergo ATR harmony. Nor is it possible for the low vowel to be skipped as in (133b) because it violates locality.

One thing that is not discussed explicitly in Parkinson 1996 is whether the representation of height is on underlying or surface vowels, though there is a brief discussion in his §3.3.2. In Pulaar, if we assume only the three heights, omitting the [+ATR] mid vowels, then it is hard to explain why / alternate with e/o and not with i/u. I will not deal with this issue here.


92 (133) a.* Aperture | [closed] a Aperture i Aperture | [closed] | [closed] | [closed] i Aperture | [closed] | [closed] | [closed] *[e.../...i]

b. * Aperture | [closed]

a Aperture


Parkinson 1996 explains the cross-linguistic blocking of low vowels in an OT framework. A harmony constraint, such as ALIGN, requires that [closed] spread leftwards. To capture the blocking of /a/, he formulates the constraints HEIGHT-PLACE, NOGAP, and IDENT-PLACE, which are ranked above the alignment constraint. (134) a. HEIGHT-PLACE: if a vowel is specified for an occurrence of [closed], it must also be a front or back vowel (Parkinson 1996:33) b. IDENT-PLACE: The place (or lack thereof) of a vowel should be the same in the input and output. c. NOGAP: Spreading cannot skip a possible target. If these three constraints are higher ranked, they will rule out the outputs in (133), and the actual harmony paradigm will have blocking by low vowels. In VPT, it is possible to explain the ATR/raising harmony by using multiple occurrences of [closed]. This explains the blocking of /a/ and the general cases of harmony. VPT can also explain the transparency of underlying glides in harmony. Because underlying glides do not have any vocalic features, they do not have any features under the Aperture node and are therefore completely transparent to the harmony (see (135)). Just as the other consonants are transparent due to the lack of Aperture features, so are the underlying glides. Unfortunately, there are no cases of derived glides interacting with harmony since the derived glides only occur stem-initially to fill an empty slot.

93 (135) ... C-place | Vocalic | Aperture | [closed] j/w ... C-place etc. i ... C-place | Vocalic | Aperture | [closed] | [closed] | [closed] [e...j/w...i]

In sum, VPT cannot explain the two types of consonant gradation of surface glides, since the features of the two surface glides, though in different levels of the hierarchy, are the same. It can, however, explain the transparency of glides in ATR/raising harmony. 5.4.4 Revised Articulator Theory In Revised Articulator Theory (RAT), the behavior of glides in both consonant gradation and vowel harmony are easily explained. In RAT, the underlying glides /j/ and /w/ have the designated articulators [coronal] and [labial] respectively. (136) a. /w/ Root | Place Body | [lab] [+rd] [+bk] Lips b. /j/ (Simplified trees) Root | Place | Blade | [cor]

Thus, when these segments alternate with a [-continuant] segment, it is not surprising that the segments are coronal [] and labial [b]. There is another set of forms which requires explanation, those vowel-initial forms that begin with an empty or extra slot. In these forms, the following vowel spreads to the empty slot and the stem-initial vocoid surfaces in the onset. Since vowels are [+continuant], they are compatible with the continuancy requirement in this grade. In (137a&b), the Root node from the following vowel spreads to the empty slot. In (137c), however, there is no segment that is [+continuant] and [+low] that is allowed to surface.

94 This is equivalent to saying that there is no low glide allowed in Pulaar. The glides which surface in (137a&b) are derived and are [dorsal]. (137) Continuant grade a. __ i/ [ji/] | Root ... Place Lips [-rd] Body [dor] [-bk] [-lo]

b. __ u/ [wu/] | Root ... Place Lips [+rd] Body [dor] [+bk] [-lo]

c. ___ a [__a] | Root ... Place Lips [-rd] Body [dor] [+bk] [+lo]

In the non-continuant grade, all stems with an empty slot surface with stem-initial [g]. The vowel features that spread to the empty slot are incompatible with the [-continuant] requirement on this grade. The features from the vowel spread to the empty slot as in the continuant grade. Because these features are spread from underlying vowels, they have the designated articulator [dorsal]. The only [-continuant] dorsal segments that are allowed in Pulaar are [, g, k, g]. The segment that shares the most features with the spread vowel features is [g], since vowels are [-nasal] and [+voice]. (138) Non-continuant grade a. __ i/ [gi/] | Root ... Place Soft Palate | Lips Body [-nasal] [-rd] [dor] [-bk] [-lo]

b. __ u/ [gu/] | Root ... Place Soft Palate | Lips Body [-nasal] [+rd] [dor] [+bk] [-lo]

c. ___ a [ga] | Root ... Place Soft Palate | Lips Body [-nasal] [-rd] [dor] [+bk] [+lo]

In Paradis's analysis, the voicing of the [-continuant] segment was stipulative. Even if the default segment were velar (as she argues, but which I argued against), it does not explain why it should surface as [+voice]. Since voiceless obstruents are less marked than voiced obstruents, the voicing is suspicious. Under my analysis, since the empty slot is filled by features from the following vowel, it explains why the [-continuant] segment is both [dorsal] and [+voice]. Paradis needed to stipulate that the default segment was velar and that it was voiced. In RAT, on the other hand, both the place (velar) and the voicing are predicted since the features are spread from a vowel.

95 In terms of harmony, RAT, which adopts Calabresian markedness theory, can also explain the general harmony pattern: blocking of /a/, and transparency of underlying glides. Whether the five vowels /i,,a,,u/ or the seven vowels /i, e,,,o, u/ are assumed underlyingly, [ATR] is still not contrastive on the triggering high vowels /i/ and /u/ or the blocking low vowel /a/. Since the low vowel /a/ must block harmony, the harmony rule must be sensitive to all features. This is illustrated in (139). (139) a. b. View all features. [-cons] | Place Body [-cons]

Root = [dor][-high] [-low] [-ATR]

Root [+ATR]

The rule in (139) will cause a [-ATR] mid vowel to raise before a [+ATR] high vowel [i] or [u] and also before a [+ATR] mid vowel [e] or [o]. The rule also accounts for the blocking of the low vowel /a/. In (140a), [+ATR] cannot spread from the final high vowel /i/ to the initial mid vowel // because the intervening /a/ causes a linecrossing violation. Since all features are visible (visible features in bold), the intervening low vowel blocks the spread of [+ATR]. The alternative in (140b) is also disallowed. The rule in (139) explicitly states that the target must be a mid vowel, which /a/ is not. (140) Blocking of /a/ a. * a ... ... Root Root [-ATR] b. * ... Root [-ATR] [-ATR] a ... Root [-ATR]

i ... Root [+ATR] i ... Root [+ATR]

Finally, both consonants and underlying glides are transparent to the ATR harmony because they do not have ATR features. Since these features do not exist, they will not cause line-crossing violations.

96 5.4.5 Summary Neither LEX nor [CONS] can explain the full range of phenomena in Pulaar. VPT runs into problems with the consonant alternations. Since front vowels are coronal in VPT, it is not clear how the [-continuant] counterpart of the derived glide [j] could ever be dorsal. RAT, on the other hand, explains why the derived segments are velar. Since vowels, which are dorsal in RAT, spread to the empty slot, it makes sense that the [-continuant] grade will be a [dorsal] segment. Whereas Paradis's analysis required two stipulations (default velar and a [+voice] empty slot), RAT requires no such stipulations. Instead, both the velarity and the voicing of the [-continuant] grade are explained by the fact that these features spread from a following vowel which is [+voice] and dorsal. Both VPT and RAT can explain the transparency of glides to [ATR] harmony by the fact that underlying glides do not have the relevant feature. In VPT, glides have no height features and therefore cannot interact with the harmony. Similarly, in RAT, glides lack ATR features and do not participate.


6 Yawelmani/Yowlumne

6.1 Introduction Yawelmani68 is a Yokuts dialect spoken primarily along the Tule River south of Porterville in south-central California. Yawelmani is relevant to the present study because of the behavior of the two glides /j/ and /w/ in vowel harmony. In addition to harmony, these glides pattern with consonants with respect to epenthesis and the glottal series. The data in this section are primarily based on the work of Newman 1944, though some more recent work exists (e.g. Hansson 1999). This chapter will show that Yawelmani is a Type V language. The underlying vowels are not allowed to surface as glides (evidence from hiatus), and the underlying glides are not allowed to vocalize (evidence from epenthesis). (141) /i/ | [i] /j/ | [j] /u/ | [u] /w/ | [w]

The consonant inventory of Yokuts is given in (142).


According to Latta 1977 and Hansson 1999, the correct name of the language is Yowlumne. In this dissertation, I will use the name Yawelmani despite the fact that it is actually the diminutive of Yowlumne since this is the term that is most commonly used in the literature. "At this point in the account of the Yowlumne Indians it is necessary to make some explanation concerning their tribal name. It appears in other printed literature as Yauelmani, the i pronounced as long e. I used this form until I had interviewed Wahumchah over a period of more than twenty years. One afternoon, when we were returning from a program before the Standard School in Oildale, he had the following to say: "You know, Mr. Latta, you make one bad mistake when you talk about my people. You call their tribe Yauelmani. That is wrong. The correct name is Yowlumne. Yauelmani means a little group of Yowlumne. I think sometime some old-timer white man met a few of my people. He asked them who they were and they said 'Yauelmani.' That was right. But it was not their tribe name. I know that from all my old people."" (Latta 1977: 280-1)

98 (142) Consonant inventory of Yokuts (Newman 1944) Labial Dental Alveolar Palato-alveolar Stops b p p' d t t' d t t' Affricates dz ts ts' d t t' Fricatives s s Nasals m m' n n' Liquids l l' Glides w w' j j' Other Velar g k k' x ' h Glottal

According to Newman 1944, the Yokuts dialect of Yawelmani lacks the palatoalveolar affricates, but does have a set of alveolar affricates. In present-day Yawelmani, what is transcribed as alveolar obstruents <d, t, t', s> are actually affricates and fricatives, varying from apico-retroflex [t, t, t', ] to lamino-post-alveolar [t, t, t', ] (Hansson 1999). Here, the stop symbols <d, t, t'> will be used in order to be consistent with Newman who transcribes and describes them as stops. In the stop and affricate series, the 'voiced' symbols (e.g. b, g) actually represent voiceless unaspirated segments. Those written with the plain voiceless symbols actually represent voiceless aspirated stops and affricates. Finally, those segments written with an apostrophe <'> are glottalized. The transcriptions in this chapter have been changed from previous texts in order to conform more closely to standard IPA transcriptions. These correspondences are given in (143). (143) Here j dz ts ts' d t t' d t t' s d t t' s Newman 1944 ' y c c' z c c' d t t' s d t t' s

The underlying vowels of Yawelmani are given in (144). The vowels [e] and [e] exist on the surface, but only as the result of lowering (and shortening) of /i/. (144) i i a a u u o o

99 All data in this chapter will be listed with the appropriate author and page number. 'N' refers to Newman 1944, 'A84/A91' to Archangeli 1984/1991, 'S' to Steriade 1986, 'K' to Kisseberth 1969, and 'Z' to Zoll 1993. 6.2 Basic Yawelmani phonology 6.2.1 Basic Processes Yawelmani has found its way into the linguistic literature because of several processes that involve rule interactions. In recent years, Yawelmani has also been relevant to Optimality Theory because of widespread opacity that results from the processes discussed below. In this section, the four processes of vowel epenthesis, vowel harmony, vowel lowering, and closed syllable shortening will be discussed in turn. These processes will be discussed in general terms and do not explicitly provide evidence for underlying glides. They are, however, necessary to understand the vowel changes that occur in the data. In addition to these four processes, this section will discuss hiatus resolution. The data on hiatus resolution shows that underlying vowels are banned from surfacing as glides. In this section, the processes will be discussed in terms of rules, but this is for ease of discussion and does not imply any particular theoretical affinity. In terms of rule ordering, the above four processes are ordered as follows: epenthesis, harmony, lowering, shortening (Archangeli 1984 and elsewhere). Epenthesis Vowel epenthesis is widespread in Yawelmani because all tautosyllabic consonant clusters are banned. Verb roots in Yawelmani generally consist of two or three consonants and one vowel. In roots that have three consonants and have either the light syllable or heavy syllable template (see §6.2.2), the two final consonants are not allowed to surface as a cluster. Instead, vowel epenthesis works to rescue the banned sequence. The epenthetic vowel is [i].69 In (145), these vowels are bolded. (145) /idl/ /t'ojx-k'a/ /sonl-mi/ [idil] [t'o.jix.k'a] [sonilmi] 'getting hungry' (N145) 'cure!' (N118) 'having back-packed' (N135) Vowel Harmony Epenthetic and non-epenthetic vowels are subject to rounding harmony. The rounding harmony only affects vowels with the same value for the feature [high]. Thus, /i/ becomes [u] after vowels that are [+high] (and of course [+round]) and /a/ becomes [o] after vowels that are [-high]. Harmony in Yawelmani can be analyzed as spreading only the positive value [+round]. Under such an analysis, rounding spreads from /u o/ and


We will see in section 6.3.2 that the epenthetic vowel can also be [a] in some nouns. The choice of [i] or [a] is determined by the noun.

100 onto the following vowels of the same height. Almost all suffixes can be analyzed as having the unrounded vowels /i a/ that are subject to harmony.70 (146) a. Harmony /gop-al/ /hud-nit/ [gop-ol] [hud-nut] 'take care of an infant-dubitative' (K3) 'know about, recognize-future passive' (K2)

b. No harmony /gop-nit/ [gop-nit] *[gop-nut] 'take care of an infant-fut. passive' (K2) /hud-al/ [hud-al] *[hud-ol] 'know about, recognize-dubitative' (K3) Lowering Underlyingly long high vowels are subject to a lowering process, whereby they surface as [-high]. One of the ways to determine the underlying height of the lowered vowels is to see whether they trigger harmony among high or non-high vowels. For example, the first word in (147a) must contain an underlying high vowel because the high epenthetic vowel surfaces as round. Similarly, the second word in (147a) must also contain an underlying

I was only able to find two suffixes that did not fit this pattern. The first is the durative suffix /-xo/. Newman points out that this suffix does not "observe the rules of suffix harmony; the suffix vowel is immutable" (Newman 1944:104). The examples with this suffix did not have any vowels following the suffix, so it is not possible to determine whether this suffix would in fact spread its [round] feature. It is clear, as Newman states, that this vowel is not subject to unrounding, as seen in [hadza-xo-m] 'finally' (N104). This suffix also changes the stress placement from the usual penultimate position, to the antepenultimate. It looks suspiciously like the verb /xo-/ 'live' (A84:259) and might be a clitic and therefore outside the domain of harmony. The only other suffix that contains a rounded vowel is /-ijo/ or /-iju/ 'prioritive'. If this suffix were /-ijo/, then it is only unusual because it contains a rounded vowel, but would not be exceptional with respect to harmony. If, on the other hand, this suffix were /-iju/, then it is exceptional in terms of harmony because both vowels are the same height but do not have the same value for the feature [round]. However, if only [+round] spreads, then this suffix is perhaps not as exceptional for the [­round] of /i/ would not be allowed to spread onto the following [+round] /u/. This suffix is also exceptional in that it seems to have variation in triggering harmony. That is, /ip-ijV-hin/ 'first he swam' surfaces as both [ep-ijo-hun] with harmony and then lowering, implying that the vowel is underlyingly high, and as [ep-ijo-hin] without harmony, implying that the vowel is of a different height than the final suffix vowel. Newman writes, "no other suffix in Yokuts is marked by this curious ambiguity of vowel content" (Newman 1944: 16). The only other place where disharmonic forms could be found is in the nouns, which do not have the same regular templates as the verbs. In the nouns, however, it is common for the vowels to alternate between [+high] and [-high] (Archangeli 1991: 258), thereby obscuring any possibility of having true disharmonic roots. Though there is often an alternation in height, the rounding of the vowels is not predictable /antuw-/ 'shaman sg.' (N205), /bimamguts'-/ 'humming bird' (N191), /uplalli/ 'dove sg.' (N125). I have found one noun that contains two high vowels with different values for [round], but again, the left vowel is [-round] which is predicted to not spread /p'islu-/ 'mouse' (N179).


101 high vowel because it does not cause rounding of the [-high] suffix. Since only vowels of like height can cause harmony, the vowel in the initial syllable must be underlyingly high. This contrasts with the forms in (147b) where the surface mid vowels correspond to underlying mid vowels. In this case, they do not trigger rounding of high vowels, but do trigger rounding of [-high] vowels. (147) a. Underlying high vowels /ut'-t/ [ot'-ut] *[ot'-it] 'steal-passive aorist' (K11) /ut'-as/ [ot'-as] *[ot'-os] 'steal-precative gerundial' (K11) /mik'-t/ [mek'-it] 'swallow-passive aorist' (K11) b. Underlying non-high vowels /dos-t/ [dos-it] *[dos-ut] /dos-as/ [dos-os] /tan-t/ [tan-it]

'report-passive aorist' (K11) 'report-precative gerundial' (K11) 'go-passive aorist' (K11) Shortening The final process that further obscures the true identity of underlying vowels is closed syllable shortening. In the first word of (147a), it is clear that the lowered long vowel could be derived from a high vowel because it is long. However, in closed syllables, the long vowels shorten, causing further opacity. In other words, in a closed syllable it is difficult to determine whether a mid vowel is derived from a long high vowel, a long mid vowel, or a short mid vowel. As with lowering in a non-alternating root, the way to determine the height of the vowel is to see whether it triggers harmony. In (148b), the high suffix vowel is round, implying that the root vowel is also a high vowel. In (148c), on the other hand, the suffix vowel is not round, implying that the root vowel is actually mid. (148) a. /mik'-nit/ b. /ut'-nit/ c. /dos-nit/ [mek'-nit] [ot-nut] [dos-nit] 'swallow-future passive' (K11) 'steal-future passive' (K11) 'report-future passive' (K11) Hiatus resolution One last relevant aspect of Yawelmani phonology is its complete ban on onsetless syllables. No words begin with a vowel and there are never sequences of vowels on the surface. There seem to be two ways to avoid vowel hiatus. First, [] can be inserted as a 'hiatus-filler' between vowels that should "morphologically follow each other" (Newman 1944: 17).

102 (149) a. -insertion /sasa-in/ [sasa--in] /loxo-its'/ [loxo--its'] /pana-its'/ [pana--its']

'eye-possessive' (N17) 'one who is pouring (it) (lit. pour-agentive)' (N27) 'one who is arriving' (N27)

b. Vowel deletion /wilil-da-ihni/ [wilal-d_-ihni] 'one who is always preparing to depart' (N109)71 /paxju-ila-hin/ [paxj_-ula-hin] 'caused to scatter' (N92) The second method employed to avoid vowel hiatus in Yawelmani is vowel deletion (see (149b)). According to Archangeli 1984, when a vowel initial suffix is added to a vowel final word, the left vowel deletes. In (149b), the first form shows a suffix vowel deleting next to another suffix vowel, while the second form shows deletion of the root vowel. In the second form in (149b), the second vowel in the surface form must come from the suffix (after harmony) because it is not lowered and because it is short in an open syllable. If this vowel were actually the root vowel, then it would surface as [o] because it is long and in an open syllable. There are, however, some examples where the right vowel deletes when two vowels are adjacent. For example, the prioritive suffix in (150) loses its first vowel when the root ends in a vowel. (150) /tana-ijo-hin/ [tana-_jo-hin] 'first took (lit. take-prioritive-aorist)' (N116)

The first vowel of the prioritive suffix /-ijo/ cannot be considered to be epenthetic in forms such as /wuj-ijo-k'a/ [woj-ujo-k'] 'sleep-prioritive-imperative' because it also occurs in the form /xatxat-ijo-k'a/ [xatxat-ijo-k'] 'eat repeatedly-prioritive-imperative'. In this latter example, the suffix vowel [i] occurs in the context VC_CV where an epenthetic vowel does not occur. Recall that an epenthetic vowel only occurs to rescue an unsyllabifiable consonant. In this example, both of the consonants could be syllabified and there would be no need for an epenthetic vowel. Therefore, the initial vowel of the suffix must be underlying. However, this suffix is exceptional in Yawelmani (see fn. 70) and the rightward deletion of this suffix might not be part of a general process.


There are a few other processes that occur in this form that are not relevant to this thesis, though they should be noted. First, it is common for suffixes with /a/ to replace the final vowel of the root, which explains [wilal-] and not [wilil-]. Also, this vowel is shortened because it is in a closed syllable. The other process in this form that is not discussed in this thesis is that word-final vowels are closed with a glottal stop and then undergo vowel shortening.

103 The relevant point about the ban on hiatus in Yawelmani is that gliding is not a possible repair for the banned vowel-vowel sequences. Thus, the form in (150) does not surface as *[tana-jjo-hin] with the glide (and then shortening). This form is not banned of the geminate glide. In fact, it is not a problem for two identical consonants or glides to surface next to each other (e.g. /giw-wis-k'a/ [gewwisk'a] 'meet-reciprocal-imperative' (N18), /mik'-k'a/ [mek'-k'a] 'swallow-imperative' (N18)). Nor does the second example in (149b) surface with a glide *[paxjo-jla-hin] (and with lowering and shortening of the preceding vowel). This form is not banned because of syllable phonotactics. It is perfectly acceptable for a syllable to be closed with a glide, as in [oj.nit] 'move the place of residence-future passive' (K12). Furthermore, in (149a), it is not possible for the vowel to surface as a glide and then trigger epenthesis *[sasa-jin]. (The epenthetic vowel is bolded.) In Yawelmani, the way to deal with vowel hiatus is either by glottal stop insertion or by vowel deletion, but crucially not by gliding. 6.2.2 Templates Common among the analyses of Yawelmani is the need for a template in the verb root, but not the suffixes. Most verbs in Yawelmani fit into one of three templates, given in (151). Archangeli 1984 uses a CV-template, while more recent work uses a prosodic morphology approach (Archangeli 1991, Zoll 1993). It is important to note that the syllables in the templatic representation are different from the syllables in the surface representation. In the template, moras are only allowed to dominate vowels (Zoll 1993). In the surface syllables, on the other hand, moras can dominate coda consonants (witness the closed syllable shortening). As with the processes in the previous section, the discussion of templates is simply to familiarize the reader with possible differences in surface forms of a single root. Each verb root has a default template associated with it which surfaces when a 'neutral' suffix, such as /-hin/ 'aorist', is added. Some suffixes, however, can impose a specific template onto the root regardless of its default template, such as /-wsil/ which forces the root into the iambic template. The three templates that exist in Yawelmani are given in (151). (151) a. Light syllable (short vowel) b. Heavy syllable (long vowel) c. Iambic (short-long) CVC(C) / µ CVC(C) / µµ CVCV(C) / µµµ

Most verbs contain one vowel and two to three consonants. In the iambic forms, the vowel in both syllables is the same. The tables in (152) show how the three templates surface with three different neutral suffixes. These three suffixes are selected to show how epenthesis interacts with these processes. The paradigm in (152) is taken from Archangeli 1984 (pg. 252). The leftmost column indicates the default template of each

104 root. The vowels that have been altered in some way (e.g. are epenthetic, shortened, rounded, or lowered) are in bold. (152) a. Biconsonantal roots with neutral suffixes Underlying -hin 'aorist' -al 'dubitative' Template Light /tsaw-hin/ /tsaw-al/ syllable [tsawhin] [tsawal] Heavy syllable Iambic /tsum-hin/ [tsomhun] /hojo-hin/ [hojohin] /tsum-al/ [tsomal] /hojo-al/ [hojol]72 Gloss 'shout' 'destroy 'name'

-t 'passive aorist' /tsaw-t/ [tsawit] /tsum-t/ [tsomut] /hojo-t/ [hojot]

b. Triconsonantal roots with neutral suffixes Underlying -hin 'aorist' -al 'dubitative' Template Light /amts-al/ /amts-hin/ syllable [amitshin] [amtsal] Heavy syllable Iambic /dijl-hin/ [dejilhin] /binit-hin/ [binethin] /dijl-al/ [dejlal] /binit-al/ [binetal]

-t 'passive aorist' /amts-t/ [amtsit] /dijl-t/ [dejlit] /binit-t/ [binetit]

Gloss 'be near' 'guard' 'ask'

As stated above, some suffixes are not neutral, but are template-inducing. The two roots in (153) are first given with the neutral future suffix /-in/. It is clear that this is a neutral suffix because the roots for 'sing' and 'walk' surface as light and iambic respectively. Only neutral suffixes will allow this variation in the template. The durative suffix /-an/ and the gerundial suffix /-()in'aj/, on the other hand, force the root into a specific template. The durative suffix forces the iambic template on the light root in 'is singing', while the gerundial suffix forces the light template on the iambic root in 'while walking'. The root templates and the suffix inducing templates are drawn from the same set of three templates in (151).


In this form, it is not clear if the bolded vowel is from the root or the suffix. In either case, it would surface as [o] as the result of rounding harmony and shortening.

105 (153) a. ilk'sing' Neutral: /ilk-in/ Iambic suff: /ilik-an/

[ilk-en] [ilekan]

'will sing' (N26) 'is singing' (N49)

b. hiwt'walk' Neutral: /hiwit-in/ [hiwet-en] Light suff: /hiwt-()in'aj/ [hiw't-in'aj]

'will walk' (N101) 'while walking' (N136)73

6.2.3 Tiers/Planes The presence of templates brings forth the question of the number of tiers or planes that are necessary to represent the language. In Arabic, for example, three tiers are used (Steriade 1984 and references therein). The consonantal tier can be different and provides the basic semantic content of the word, as in (154a). These two forms have the same vowels and the same template CVCVC but the consonants are different. Similarly, the vowels can change while the template and the consonants remain the same, (154b). The passive has the same basic template CVCVC, but requires the vowels to be [u...i]. Finally, the consonants and vowels can be unchanged, while the template is altered as in (154c). These examples show that consonants, vowels, and templates can each be altered and that this difference causes a change in meaning.74 (154) a. katab-at qara-at b. katab-at kutib-a c. katab-at a-ktab-a 'he wrote'(past) 'he read' (past) 'he wrote' (past) 'it was written' (passive) 'he wrote' (past) 'he causes s.o. to write' (causative) CVCVC CVCVC CVCVC CVCVC, u/i CVCVC a-CCVC

The need for three tiers in the Yokuts languages is less clear.75 The discussion of tiers/planes is directly relevant to the evidence for underlying glides in Yawelmani. If consonants and vowels were truly on different planes, then the transparency of glides in


The glottal stop in parentheses is a 'ghost' consonant. Ghost consonants only appear when the root is biconsonantal (as though there were room for the ghost consonant to fill the position of the third consonant). When the ghost consonant is a glottal stop and when it does not appear as a separate position, it surfaces as glottalization on the second consonant if it is a sonorant. 74 This is a simplification, but provides a basic picture of the Arabic facts. The data are from Mike Brame, p.c. 75 " particular the segregation of vowel and consonant segments on separate tiers, is not independently supported by the morphology of Yokuts..." (Schein & Steriade 1986: 739).

106 harmony (see §6.3.1) could be explained by the fact that glides are on the consonantal plane. Assuming different tiers for consonants and vowels could also explain the fact that clusters which include glides also force epenthesis. It is therefore necessary to consider whether there is evidence in Yawelmani for three tiers (C, V, and template) or simply two (melody and template). In this section I will follow Steriade 1986 and argue that Yawelmani is characterized by only two tiers. Archangeli (1984; 1991) argues for three distinct tiers parallel to the Arabic tiers. Under her analysis, the vowel and consonant melodies are on separate tiers, as is the template tier. Steriade 1986, however, argues against separate vowel and consonant tiers. Her argument is two-fold. First, the difference between languages like Arabic and those like Yokuts is rooted in the morphology. In Arabic, the vowel and consonant melodies can be changed separately to bring different meanings. As shown above, the same template with [a] is active while [u...i] is passive. In Yokuts, on the other hand, the vowel and consonant melodies act together in bearing the semantic content of the root. That is, it is not the case that the consonantal melody brings one piece of morphological information, while the vocalic melody brings another. There are no examples where roots with the vowel [a] are all part of a particular semantic set, while those with [u] are part of another. Thus, Steriade's first argument against separate melody tiers relies on the importance of maintaining the notion that distinct planes are only necessary when there are distinct morphemes. The second type of evidence against the separation of melodic tiers in Yokuts comes from the process of Secondary Lowering that is found in related dialects, though not in Yawelmani. Secondary Lowering is a process whereby a short vowel in the same root also lowers when the second syllable contains a long vowel. The example in (155) from the Gashowu dialect shows that the first vowel also lowers when the second vowel does. (155) /uduk'-it'/ [odok'-ut'] *[udok'-u'] no gloss (S135)

Archangeli 1984 argues that Secondary Lowering is the result of long distance geminates. Under such an analysis, all three slots for the vowel in the iambic template are linked to one melody. Thus, when the long vowel of the second syllable lowers, so must the vowel of the first. Under this analysis, long vowel lowering and 'Secondary Lowering' are actually the result of the same process (multiply linked vowels must lower). (156) C V C V V Cuo Steriade argues that Secondary Lowering is not the result of long distance geminates on three accounts. First, nouns, which do not have evidence for a templatic

107 tier, also show secondary lowering. In verbs, only a single vowel quality appears regardless of the template. Nouns, on the other hand, can have multiple vowels in the root. In (157), 'bottle' and 'stone-loc.' show that the vowel melody is not predictable, as it is in the verbs. Thus, there is little evidence for a predictable template and vowel melody in nouns. Though there is no evidence for templates in nouns, they still exhibit the pattern of Secondary Lowering. Since the nouns do not have templates, the Secondary Lowering cannot be the result of long distance geminates. (157) a. Choynimni: nepej 'brother's wife' lemeda 'bottle' hogojus 'sibling of opposite sex' boton 'man' b. Chukchansi and Gashowu selela-w 'stone-loc.' nebet' 'older brother' c. Gashowu gojolum k'etet'

'young woman' 'short'

Second, the direction of Secondary Lowering is always leftward. If Archangeli's long distance geminates analysis were correct, then CVCVV and CVVCV would act identically and lower the vowels in both syllables. However, the secondary lowering only occurs when the short vowel is followed by a long vowel, not when it is preceded by one. In (158b), the short vowels in the second syllables are not epenthetic because they occur after an open syllable (not a location where an epenthetic vowel occurs). Therefore, these short vowels must be part of the root. In these forms, the short vowel does not lower. (158) a. Choynimni /husu-i/ /ussut-it'/ b. Chukchansi [k'enit'-in] [wewila-w] [hosuu] [ossutut']

'coyote' 'steal-agentive'

'red ant-poss' *[k'enet'-in] 'tree limb-loc' *[wewela-w]

108 Finally, Secondary Lowering does not occur in suffixes. For example in Chukchansi /-iwi/ 'prioritive', the long vowel lowers, but the first vowel does not. (159) [tux-uwo-k'] [belen-iwe-k'] 'tie-prioritive-imperative' 'wrap up-prioritive-imperative' *[tux-owo-k'] *[belen-ewe-k']

The exact nature of Steriade's analysis is not what is important, though a small sketch is provided below. What is important here, is the ability to account for the Yokuts data without reference to separate vowel and consonant tiers. Instead, Steriade's analysis allows for only two tiers, the melodic and the templatic. Steriade accounts for the lowering facts by a rule of vowel copy. In her analysis, the underlying verb roots have segments in the order /CCV(C)/ and the vowel is copied from the second syllable to the first. What differentiates the dialects with Secondary Lowering from those without it, under Steriade's account, is the order of the copy rule and the simple lowering rule. If lowering occurs before copy, then Secondary Lowering occurs. If copy occurs before lowering, then no Secondary Lowering occurs. This copying explains the directionality of Secondary Lowering as well. In the verb roots that are not exceptional, the long vowel of the second syllable is copied to the first syllable (CCVCCVCVC). The reverse iambic forms (CVCVC), on the other hand, do not have the underlying form CCVC because they are not part of the regular class of verbs. The reverse iambic forms simply have the underlying form /CVCVC/ and therefore do not require a rule of vowel copy at all. With no need for vowel copy, secondary lowering does not apply. Conversely, nouns, which do not have evidence for a template, still require vowel copy. For example, 'bottle' in (157a) would have the underlying form /lmida/. In this dialect with secondary lowering, lowering is ordered before copy, so the surface [lemeda] is explained. The copying analysis also explains the lack of lowering in the suffixes because they do not have a template and there is no need to copy the vowel. 6.3 Evidence for underlying glides This section discusses the evidence for underlying glides in Yawelmani. In particular, glides behave differently from vowels, but similarly to consonants with respect to vowel harmony, epenthesis, and other distributional facts. 6.3.1 Vowel Harmony Glides in Yawelmani do not behave like the vowels with respect to rounding harmony. Whereas high vowels are triggers, targets, and blockers of rounding harmony, the glides are not. Since only [+round] spreads, the examples will be of the form where a high round vocoid is followed by an unrounded one. Thus, the potential triggers are /u/ and /w/ and the potential targets are /i/ and /j/. Blocking effects are only possible between vowels of different heights, thus the blocking data will be of the form V[-high]...V[+high]...V[-high], where the middle high vocoid is either a vowel or a glide.

109 Vowels, not surprisingly, do interact with the vowel harmony. They act as triggers, targets, and blockers. In (160a), the high round root vowel acts as a trigger in causing the following suffix vowels (or epenthetic vowels) to become round. In (160b), high vowels act as blockers for non-high harmony. Notice that even the epenthetic vowels prohibit rounding from docking on the suffix vowel. Finally, in (160c) both underlying vowels and epenthetic vowels are the targets of rounding harmony. (160) a. Vowels as triggers /dub-inaj/ [dub-unaj] b. Vowels as blockers /t'ojx-k'a/ [t'ojix-k'a] /moxl-as/ [moxil-as] c. Vowels as targets /dub-inaj/ [dub-unaj] [tsum-u-t] /tsum-t/


'leading by the hand' (A84:37) (N136)

*[t'ojix-k'o] 'cure-imperative' (N118) *[moxil-os] grow old-precative gerundial' (K32)

*[dub-inaj] *[tsum-i-t]

'leading by the hand' (A84:37) (N136) 'destroy-passive aorist' (A84:252)

The glide data are presented below in (161). The data in (161a) show that the rounded glide /w/ does not trigger rounding harmony on the following high vowel. If the surface glide were actually an underlying vowel /u/, then it would be possible for its features to spread to the following vowel of like height. In (161b), the glide does not block harmony from a non-high vowel, as would be expected of a high vowel. Finally, (161c) shows that glides (/j/) are not targets of harmony even after a high vowel. Again, if these glides were actually underlying vowels, then they would be predicted to participate in the harmony either as a trigger, a blocker, or a target. The fact that they do not, and that their behavior is parallel to other consonants, implies that these segments are indeed underlying glides. (161) a. Glide as a non-trigger [tsaw-hin] *[tsaw-hun] 'shout-aorist' (A84:262) /tsaw-/ /bown-/ [bowin-hin] *[bowun-hun] 'trap aorist' (A84:263) /bown-/ [bown-it] *[bown-ut] 'trap passive aorist' (A84:263) /lowon-wsil/ [lowon-i-wsel] *[-wsol] 'attend a feast-reciprocal' (N166) /luk'ul-wsil-w/ [luk'ol-u-wsol-a-w] *[-wsol-] 'bury-recip.-loc.' (N167) /jolow-/ [jolow-it] *[jolow-ut] 'assemble-passive aorist' (K24) /jolow-/ [jolow-hin] *[jolow-hun] 'assemble-aorist' (K24)

110 b. Glide as a non-blocker /jolow-/ [jolow-ol] /logw-/ [logw-ol] /jolow-k'a/ [jolow-k'o]

*[jolow-al] *[logw-al] *[jolow-k'a]

'assemble-dubitative' (K23) 'pulverize-dubitative' (K32) 'assemble-imperative' (N118)

c. Glide as a non-target /t'unuj-/ [t'unoj-hun] *[t'unow-hun] 'scorch-aorist' (K24) /wuj-/ [woj-ujo-k'] *[wow-ujo-k'] 'go to sleep-prior.-imper.' (K59) /duj-dj-hin/ [duj-duj-hun] *[duw-duw-hun] 'sting-red.-aorist' (N122) This section has shown that glides behave differently form both underlying and epenthetic vowels with respect to rounding harmony. Naturally, the vowels participate in harmony. The surface glides, however, do not interact in any way with the harmony, implying that they are indeed true underlying glides. 6.3.2 Epenthesis The second type of evidence for the existence of underlying glides comes from data involving vowel epenthesis. Underlying glides, like consonants, trigger epenthesis when they cannot be syllabified. This is true both of glides in suffixes and in roots. When the final root glide is forced into a cluster that cannot be syllabified, the glide does not vocalize, but requires epenthesis. The forms in the first two rows of (162) have a glide in root-final position. When a consonant initial suffix is added, the resulting three consonant cluster is banned. (Yawelmani does not allow any tautosyllabic consonant clusters.) (162) Root glides that require epenthesis UR passive aorist /-t/ aorist /-hin/ utj logw wuwl pat utj-ut logw-it wowl-ut pat-it utuj-hun logiw-hin wowul-hun pait-hin imperative /-k'a/ utuj-k'a logiw-k'a wowul-k'a pait-k'a 'fall' 'pulverize' 'stand up' 'fight'

If these segments were in fact underlying vowels, then there would be no need to epenthesize, and the segments would have surfaced as vowels *[uti-t], *[logu-t]. The two final rows show examples where the third root consonant is not a glide. These serve to illustrate that the glide final forms pattern exactly the same way as the non-glide final forms. The reason for the non-vocalization in (162) is not the result of a ban on verbs ending a vowel. First, words in Yawelmani can end in either a vowel or a consonant (Newman 1944: 25). Second, all regular verb roots with two consonants will end in a

111 vowel in the iambic template CVCVV. Third, irregular verb roots can end in a vowel, as in [wastu-] 'hurt' (S131) or [hoj'li-] 'hunt' (N153). Thus, there is no general ban on ending in a vowel and this cannot be the reason for the non-vocalization. One might perhaps argue that the glides in (162) are in fact underlying vowels but that the template requires them to surface as glides (that is, in a syllable peripheral position). This is not a possible analysis for verbs. The forms in (162) have the light template (µ) which only requires the first syllable to have a short vowel. It says nothing of the final segment of the verb root. Since some roots have only two consonants, the presence or absence of the final segment in light templates is irrelevant. It is plausible that the realization of the final segment is also irrelevant. If that final segment were an underlying vowel, then it would be expected to surface with the best syllabic parse. Though assuming that glides are really underlying vowels is potentially a possible analysis for roots, it is not a possible analysis when considering the glides in suffixes, which have no templatic requirements. The two suffixes that have glides and require epenthesis are the reciprocal suffix /-wsil/ and the locative suffix /-w/.76 When the roots end in a vowel, these suffixes are simply added with no epenthesis or deletion. When the stem ends in a consonant, an unsyllabifiable cluster is created, requiring epenthesis. (163) Reciprocal Suffix /-wsil/ /lowon-wsil/ [lowoniwsel] /luk'ul-wsil-w/ [luk'oluwsolaw] /jawal-wsel/ [jawaliwsel] s [amatsiwsel] /amat -wsil/ /dijil-wsil/ [dijeliwsel] /binit-wsil/ [binetiwsel]

'attend a feast-recip' (N166) 'cemetery' (N167) no gloss (A91:272) no gloss (A84:274) no gloss (A84:274) no gloss (A84:274)

If the reciprocal suffix were actually /-usil/ then there would be no need to epenthesize and the forms in (163) would all look like [...C-usel]. Actually, if the suffix-initial segment were an underlying vowel, it would cause the long vowel /i/ to become round due to harmony, which is also not the case. Furthermore, if this suffix began with an underlying vowel, then it would cause deletion of a preceding vowel, which it does not (see §6.3.3). The locative suffix also requires epenthesis. In these forms, the epenthetic vowel is either [i] or [a] and is determined by the noun itself. In most of the forms with the


Yawelmani has several suffixes that contain 'ghost' consonants. These consonants do not surface in some circumstances. Several of the suffixes with ghost consonants begin with a cluster like the reciprocal does. As far as I can tell, the reciprocal suffix is the only one that begins with a cluster where the initial segment is not a ghost consonant.

112 locative, the chosen epenthetic vowel is the opposite height of the preceding vowel. However, the last form in (164) shows that this is not true in all cases.77 The epenthetic vowels are underlined. (164) Locative suffix /-w/ /lagaij-w/ [lagaej-a-w] 'a place for staying overnight-loc.' (N195) /pil-w/ [pil-a-w] 'road loc.' (K62) 'sweathouse loc.' (K62) /mus-w/ [mos-a-w] 'narrative loc.' (K62) /hol-w/ [hol-i-w] s s /t apt'-w/ [t apt'-i-w] 'gambling tray loc.' (K62) /lomt-w/ [lomt-o-w] 'mountain-loc.' (N203) As with the reciprocal, it is not possible to analyze the locative suffix as containing an underlying vowel /-u/. If it were simply an /-u/ then it would behave like the other vowel suffixes and would surface as a vowel in final position (for example the objective case marker /-i/ or /-a/ [polm-a] 'husband], [nusos-u] 'paternal aunt', [tok'-o] 'kill' (N198)). That this suffix requires epenthesis indicates that it is indeed an underlying glide. Some might wonder if it is not possible to analyze the epenthetic vowels in (164) as simply being part of the root. If this were the case, then the locative suffix might be expected to surface as a glide. The second form in (163) for 'cemetery', however, shows that epenthesis is also required after the consonant-final suffix /-wsil/. Assuming that the epenthetic vowels in (164) are actually part of the root would be odd. There is no reason to assume that root vowels should be deleted in all other circumstances except before the locative suffix. Furthermore, other cases of would-be hiatus are never resolved by gliding, but instead by deletion. It is therefore not possible to analyze the nominal roots in (164) as ending in a vowel. 6.3.3 Other parallels with consonants There are several other ways in which underlying glides pattern with the consonants, to the exclusion of vowels. First, all consonants and glides contrast for glottalization, whereas vowels do not. The first two examples in (165) show that forms can minimally contrast for glottalized versus plain glides.78

"The quality of the vowel which fills...the slot epenthesized before w, is determined by the noun itself, and is either /i/ or /a/" (Archangeli 1984: 311). 78 Glottal sonorants, including w' and j', are limited to the second consonant position in the root, whereas glottal obstruents are allowed in all three consonantal root positions.


113 (165) t'ojix t'oj'ix aj'k k'an' t'aw' gij' t'aw' 'applying medicine 'getting rusty' 'throw a glance' 'touch a goal' 'throw' 'touch' 'aim' (N19) (N19) (A84:285) (A84:285) (A84:285) (A84:285) (A84:285)

Second, glides and consonants do not delete when adjacent to a vowel, whereas vowels do delete. Several examples were given in §6.2.1 where vowels in hiatus were repaired by vowel deletion. One example is repeated below in (166b). When a glide and a vowel are adjacent, however, the glide does not delete, implying that it is in fact an underlying glide and not an underlying vowel. The patterns are true for glides that are morpheme internal as well as those created across a morpheme boundary. (166) a. Glides: No deletion /tunu-wsil/ [tuno-wsol] /hojo-wsil/ [hojo-wsel] /mokj-a/ [mokj-o] /hiwt-()in'aj/ [hiw't-in'aj] b. Vowels: Deletion /paxju-ila-hin/ [paxj_-ula-hin]

'door obj.' (N167) no gloss (A84:274) 'wife-objective' (N198)79 'while walking' (N136)

'caused to scatter' (N92)

It is not possible to assume that the glides in (166a) are the realization of underlying vowels. Section 6.2.1 already showed that gliding of underlying vowels is not a possible repair for of hiatus. The data in (149) and (150) showed that gliding is not possible, even if the syllabification would allow it. Gliding as a process is not banned because of illicit syllable structures. It is simply not part of the grammar of Yawelmani to allow an underlying vowel to surface as a glide. Thus, the surface glides in (166) cannot be underlying vowels, but must be underlying glides. 6.3.4 Summary The preceding sections showed that underlying glides do exist in Yawelmani. Glides pattern with consonants, to the exclusion of vowels, with respect to harmony, epenthesis, vowel deletion, and glottalization. Yawelmani does not allow underlying glides to surface as vowels, nor does it allow underlying vowels to surface as glides.


This form is from the Chawchila dialect.

114 6.4 Comparison of the representations This section will discuss the ability of the four models to represent the facts of Yawelmani vocoids. Though several types of evidence for the existence of underlying glides were discussed in §6.3 (e.g. syllabification and epenthesis, glottal series), only the harmony process will be discussed in this section. The reason for this is that it is the only segmental process that occurs on a featural level and thus can distinguish among the models. The fact that underlying glides trigger epenthesis in clusters is not a featural process. The relevant harmonic data is summarized in (167). Underlying vowels act as triggers, blockers, and targets in the harmony, whereas glides do not. It is their lack of interaction that requires an explanation. (167) Summary of Relevant Data Underlying vowel Trigger? Yes /dub-inaj/ [dub-unaj] Blocker? Yes /moxil-as/ [moxil-as] Target? Yes /dub-inaj/ [dub-unaj]

Undelrying glide No s / t aw-hin/ [tsaw-hin] No /jolow-al/ [jolow-ol] No /wuj-/ [woj-ujo-k']

6.4.1 LEX Representing the difference between underlying vowels and glides with merely a structural difference does not explain their different behaviors in vowel harmony. If there were no featural difference between vowels and glides, then the glides would be predicted to interact with the harmony. Their representation in LEX is given in (168). The underlying glides are those segments that cannot be linked to the nucleus. (168) Representation of vowels and glides in LEX N N * * /i/ /i/ /u/ /u/ (features are identical) A simple demonstration of harmony in this model is provided below. The expectation is that underlying vocoids with the same value for the feature [high] will agree in rounding. A sample rule is provided in (169). If underlying glides are present in a word, they are wrongly expected to interact with the harmony as in (170). The vocoid that surfaces incorrectly in these forms is bolded. (169) ­cons [ round] / -cons high round high C 0 _____

115 (170) a. Predicted trigger *[tsaw-hun] b. Predicted blocker *[jolow-al] c. Predicted target *[wow-ujo-k'] /wuj-/ Though the rule above was formulated with SPE-style stacked features, the same incorrect results would hold if we were to assume a standard feature hierarchy, like that of Vowel-Place Theory. In this case, the only difference between an underlying glide and vowel would be structural and both would have place features under the V-Place node. If the underlying glides had V-Place features, they would be predicted to interact with the harmony. Any representation based on LEX cannot predict the correct results. 6.4.2 [CONS] Using the feature [consonantal] to distinguish between underlying vowels and glides only explains the data in Yawelmani if an SPE-style representation is also assumed. The rule would be the same as the previous section, and is repeated below. Since the underlying glide is [+consonantal] in this approach, it will be completely transparent to the harmony. However, using [consonantal] in this way is essentially replacing [syllabic] since it is separating the class of sounds into two, those that form syllabic nuclei and those that do not. Using the rule in (171) predicts the transparency of underlying glides; they are neither targets nor triggers since they are not [-consonantal]. Furthermore, they will not block because the rule allows for the existence of intervening consonants (i.e. [+consonantal] segments). (171) ­cons [ round] / -cons high round high C 0 _____

If a feature geometric representation is assumed instead of an SPE-style one, then the question arises as to the location and type of the features for underlying glides. If the assumption is that they are under C-place, then this is identical to using the representation in §6.4.3. If, on the other hand, glides have place features under V-place, then they are incorrectly predicted to interact with the harmony, since their features will cause line crossing violations. 6.4.3 Vowel-Place Theory It is not surprising that Vowel-Place Theory (VPT) can account for vowel harmony in Yawelmani, as it is these types of cases which it was developed to explain. In Yawelmani, only vowels (and their vocalic features) interact in the harmony. The

116 spreading is illustrated in (173). If two adjacent aperture nodes have the value of the feature high, then the V-place node spreads rightward. An example is given in (174). Recall that the four underlying vowel qualities are /i, a, o, u/ which will be assumed to have the features in (172). (172) Assumed underlying vowel features of Yawelmani in VPT labial coronal dorsal high80 i + u + o a (173) Spreading rule in VPT Root ... vocalic V-place [lab] [dor] (174) Spreading u ... Vocalic V-place [lab] [dor] aperture [ high] Root ... vocalic = V-Place aperture [ high]

e.g. /dub-inaj/ [dub-unaj] C0 i ... ... Vocalic = aperture V-place aperture [coronal] [+high]


If a vowel of a different height intervenes between the two vowels of like height as in (175), V-place is banned from spreading to the final /a/. The V-place node of /o/ cannot spread across the intervening high vowel /i/ because this would cause a line crossing violation. This node will not spread to the intervening /i/ either, since the two vowels do not have the same height.


Of course, in VPT, [closed] (Parkinson 1996) or [open] (Clements & Hume 1995) would be used.

117 (175) Blocking * o ... Vocalic V-place [lab] [dor] e.g. /moxil-as/ [moxil-as] *[moxil-os] i a ... ... Vocalic Vocalic = aperture V-place aperture V-place aperture [-high] [cor] [+high] [dor] [-high]

In the version of VPT outlined in Chapter 3, glides are represented with only Cplace features. In Yawelmani then, the glides /j/ and /w/ (and their glottalized counterparts) lack features under the vocalic node and therefore behave like the consonants in being transparent to the harmony. An example is illustrated in (176). Here, the height features of the two non-high vowels are identical, allowing the spreading of the V-place node. (176) Non-interaction of underlying glides e.g. [jolow-ol] o w ... ... C-place C-place Vocalic V-place [lab] [dor] aperture [-high] [labial]

a ... C-place Vocalic = V-place [dor] aperture [-high]

The essence of Yawelmani harmony in VPT is based on the fact that the height features between two harmonic vowels must be the same. In this way labial harmony in Yawelmani is parasitic on height. Glides, which lack any vocalic features in VPT, are of course completely inert with respect to harmony. Thus, VPT can account for the behavior of glides in Yawelmani vowel harmony. 6.4.4 Revised Articulator Theory Revised Articulator Theory (RAT) relies on a different mechanism to explain the inertness of glides in Yawelmani harmony. Having only a single layer of place features, RAT relies on the visibility of features based on whether they contrast (see Chapter 3 and Calabrese 1995 for further discussion). Relying on the strict Calabresean approach to features, however, causes problems in the representation of harmony in Yawelmani. According to Calabrese 1995, languages that look like standard vowel systems (e.g. five or seven vowels) but that lack one of these vowels should be treated like the standard set with respect to contrastiveness. Therefore, the absence of the fifth vowel /e/ should be

118 attributed to an accidental gap in the language. When considering which of the marking statements are active or inactive, the five vowel system is taken as the base. In (177), the contrastive features are boxed. (177) Yawelmani vowel features based on Calabrese 1995 i e a o high low back round + + + + + u + + +

The problem with assuming the system in (177) is that /a/ is contrastive neither for [high] nor for [back]. Furthermore, based on the way Calabrese's marking system is established, the feature [round] is always subordinate to the feature [back]. That is, backness is considered the primary difference between /i/ and /u/, and not rounding. This is problematic. If [back] is the feature that spreads in harmony, it can explain a change of i~u, but cannot explain a~o. If [round] is the feature that spreads, as is intuitively plausible, then the spreading rule would have to be sensitive to all features since round is not contrastive in (177). The problem with a spreading rule that is sensitive to all features is that underlying glides would be expected to interact since they also contain the spreading feature. In fact, the full specification of RAT means that consonants are also specified as [-round] and would be predicted to block if all features were visible to the spreading. Though the marking statements and the insights of Calabrese's markedness theory are helpful to understanding how assimilation/spreading occurs in many languages, it may be too narrow in its focus. That is, assuming that the four vowel system of Yawelmani is really a defective five vowel system misses the point and is unable to account for the data. A rule sensitive only to contrastive features will predict that /a/ will be transparent, for none of [high], [back], or [round] are contrastive for /a/. As a result, it is expected that any features could spread across this segment. Thus, seeing only contrastive features is too limited and not sensitive enough because the features of /a/ will not be visible. Being sensitive to all features, on the other hand, runs into a similar problem. In this case, though the glide /w/ and the consonants are not contrastive for [round] or [high], these features will be visible to a rule which sees all features. If, however, we assume an underlying four vowel system (as actually exists) as in (178), the relevant contrastive features are clear. The intuition in Calabrese is that identifying the contrastive features is dependent upon what segments are in the inventory in the first place. When considering the actual inventory of vowels in (178), it is clear that [low] and [back] do not partition the set of vowels into two equal sets. Only the two features [high] and [round] fully describe and partition the set of vowels. That is not to

119 say that the other features are underspecified. On the contrary, Calabresean markedness assumes full specification. Instead, we say that only [high] and [round] are contrastive, indicated in (178) with the boxes. The other features are present in the vowels, but are not contrastive. (178) Revised features of Yawelmani i a o high low back round + + + + + u + + +

Now that the contrastive features are established, the spreading rule is simple. In (179), [+round] spreads from one [-consonantal] element to a following [-consonantal] segment if they have the same value for the feature [high]. As with VPT, RAT captures the notion that rounding harmony is parasitic on height. (179) a. View contrastive features. b. Spread contrastive [+round] [-cons] ... Place Lips [+round] Body [high]

[-cons] ... Place Body [high]

Lips = [round]

In this representation, any segment that has a contrastive [round] feature will block the spread of harmony. Consonants as well as glides are not contrastive for [round] and are therefore inert. Intervening vowels of a different height, on the other hand, block the harmony because they are also contrastive for that feature. In (180), [+round] from /o/ cannot spread to the final vowel because it is blocked by the intervening [-round] of /i/. In (180), the contrastive features are given in bold and the non-contrastive features are grey and small.

120 (180) Blocking * o ... Place Lips Body


e.g. /moxil-as/ [moxil-as] *[moxil-os] i a ... ... Place Place Lips [-round] Body [+hi] [-bk] [dor] [-lo] Lips Body = [-round] [-hi] [+bk]

[+round] [-hi] [+bk]


The difference between (180) and (181) is that the glide in (181) is not contrastive for [round] or [high]. Therefore, in this example, the [+round] can spread across /w/ and onto the final vowel because there is no visible intervening feature which blocks it. (181) Non-interaction of underlying glides e.g. [jolow-ol] o w ... ... Place Place Lips Body


a ... Place Lips Body = [-round] [-hi] [+bk]


[+rd] [lab]



[+round] [-hi] [+bk]


By using a single tier of features and the notion of invisibility of certain features, RAT can account for the transparency of the glide in Yawelmani harmony. Because the glides are not contrastive for the relevant feature and because the spreading rule is sensitive to only contrastive features, their transparency is explained. 6.5 Discussion To summarize, LEX cannot account for the different behavior of glides and vowels in harmony because it maintains that their features are identical. [CONS] can simultaneously explain the transparency of glides and the non-transparency of vowels only if a stacked SPE-style analysis is adopted. VPT can also model the transparency of underlying glides because they are not specified for features on the plane where harmony occurs. In VPT, the V-place node spreads to other vowels which have the same height feature. VPT must also assume that the feature [high] be shared between the two vowels which harmonize. The blocking of intervening vowels is therefore due to an inability to merge the feature [high], rather than the fact that the intervening vowel can block the spreading feature. Similarly, RAT represents the blocking of rounding harmony by vowels of a different height by the fact that the intervening vowel contains the feature that is being spread. Both RAT and VPT must encode the fact that the harmony is parasitic on height, meaning that the two vowels must agree in height (broadly) in order for rounding to spread.



7 Tahltan

7.1 Introduction Tahltan81 is a northern Athabaskan language spoken in Iskut and Telegraph Creek, British Columbia, Canada. Tahltan is known for its consonant harmony, which occurs not only with sibilants, but also with the interdentals. What is of particular interest here is the fact that the palatal glide participates in the consonant harmony. The primary sources of data for this chapter come from Hardwick 1984 and Shaw 1991. The data in these articles were collected in Telegraph Creek by Shaw in 1982. Because consonant harmony is a recent innovation within Tahltan, most of the data will be cited from these two sources. However, in order to illustrate other aspects of the language, data will on some occasions be drawn from other sources as well.82 Tahltan is a Type V language, where the underlying vowels and glides do not show alternations. (182) Type V /i/ /j/ [i] [j]

The consonant inventory of Tahltan, and of Athabaskan languages in general, is usually described in terms of series. In these terms, Tahltan has five series of coronals (D (d, t, t'), D (d, t, t', , ), DL (dl, t, t', l, ), DZ (dz, ts, ts', z, s), D (d, t, t', /ç, /j)) as well as a velar and marginally, a uvular series. In the phoneme inventory in (183), I have chosen to list the consonants in the standard format based on place of articulation.


I would like to extend special thanks to Sharon Hargus for helping me understand several aspects of this Athabaskan language and for helping me locate the proto-forms for several of the words that are of special interest in this chapter. 82 In this chapter, data sources will be abbreviated to KL=Krauss & Leer 1981, H=Hardwick 1984, L87=Leer 1987, N=Nater 1989, S=Shaw 1991, B=Bob 1999, A=Alderete (in press), R=Rice 1997.

123 (183) Consonant Phoneme Inventory (based on Hardwick 1984) Lab. Interdental Dental Palatal Velar Stop b d t t' g k k' gw kw kw' Affricate dz ts ts' d t t' d t t' dl t t' Fricative zs çj x l () w xw Nasal m n n' n n' Glide w Laryngeal


There are three series of stops and affricates: voiceless unaspirated, voiceless aspirated, and voiceless ejective. The voiceless unaspirated stops are notated with the 'voiced' symbol and the voiceless aspirated with the 'voiceless' symbol. The chart in (183) differs from that in Hardwick 1984 in several respects. First, Hardwick uses the symbol <y> for the palatal glide. Second, she lists the uvular consonants [ q q' ], but notes that they are not contrastive and only exist in the speech of a few older speakers (Hardwick 1984: 10). Therefore, they have been omitted from the inventory. Finally, the inventories from Hardwick 1984 and Shaw 1991, as well as Nater 1989 and Bob 1999, include and either as separate phonemes or as variants of ç and j. In searching through Hardwick 1984, Shaw 1991, and Nater 1989, I found only one form listed with the voiced [], [-tu] 'to handle a sheet, fold' (Nater 1989: 40, Hardwick 1984: 16). According to Nater, [] is "marginal ... insofar as it has not been recorded word-initially" (Nater 1989: 40). In addition to the form above, Nater also found [] in the two forms [ah-ane] 'mature ram' and [e-ee] 'my brother's son'. However, he further notes that [ane] is a variant of [-jane], as in [iba-jane] 'mature billy goat', and that [e-ee] varies freely with [e-jee]. Nater further remarks that [] is formed through progressive assimilation targeting [z] and [j] and triggered by a consonant in the D series. This is different from the regressive assimilation that will be dealt with in §7.3. The alternations [-jane] ~ [-ane] and [-jee] ~ [-ee] illustrate that the glide can be the target of the assimilation, and especially that it does not need to be directly adjacent since in [ah-ane] it is not adjacent. The segment [] has a wider distribution than [], though I was only able to find it in the roots [to] 'paddle' or 'feathers' (H3, 32), [du]/[du] 'cat' (B31/H44), [ta] 'set line'/[t 'a] 'nightline' (B65/N28). [] also surfaces in some prefixes due to consonant harmony. Since and are marginal in roots, I have omitted [] from the inventory in (183) and put [] in parentheses.

124 The vowel phonemes of Tahltan are described differently in Hardwick 1984 and Alderete & Bob (in press). Since we will be concerned primarily with the consonants and their interaction in consonant harmony, I will simply provide the two vowel inventories below, without making a decision as to the correct one. (184) Vowel Phonemes of Tahltan Hardwick 1984 Alderete & Bob (in press) i u i i u u e o a oi e e o o a a 7.2 Some basic phonology of Tahltan This section will deal with a few aspects of Tahltan phonology, namely syllable structure, the d-effect, and the distribution of stem initial and final consonants. The section on syllable structure is included for completeness. The discussion of the d-effect is necessary in order to understand alternations that affect stem-initial consonants. Some discussion of the stem-initial and final consonants is necessary to establish the validity of the inventory in (183) which shows the glide /j/ as phonemic. The distribution of consonants in Tahltan, and other Athabaskan languages, is quite complex, showing different patterns in stem-initial versus final positions, as well as different distributions in nouns and verbs. 7.2.1 Syllabification Syllables in Tahltan are generally open. According to Bob 1999, there are two places where closed syllables can occur: immediately preceding a stem and word-finally. No tautosyllabic consonant clusters are allowed (Nater 1989).83 Some examples are given in (185). The consonants which close a syllable are underlined. (185) bs dadnhdos h dahbd esbd 'knife' 'he's boiling Obj.' 'he's tanning Obj.' 'our food' 'my food' (B29) (B29) (B29) (B29) (B29)


The only exceptions in Nater's data are [skadi] 'crazy', [sohde] 'soapberries', and [k'unts] 'potatoes'.

125 7.2.2 D-effect The d-effect is a process whereby a /d/ from the d-classifier (see §7.2.3) or the first person plural prefix /id-/ fuses with the following consonant. In the d-effect, an input of d+fricative surfaces via coalescence as either an affricate or a stop at the same place of articulation. The resulting segment is voiceless unaspirated, notated here with the 'voiced' symbol. Since the d-effect only occurs on verb stems, and the voicing of steminitial fricatives is predictable, the representation of d-x g or d- g, for example, is the same. In addition to the combinations with fricatives, d- surfaces as [t']. Before all other consonants, the /d/ simply deletes. The paradigm is given in (186). The forms in (187a) take the l-classifier, while those in (187b) take the h-classifier. This explains why the stem-initial fricatives in (187a) surface as voiced in the absence of the d-effect and why those in (187b) surface as voiceless. In (187), the segments that alternate are underlined. (186) Input d-/l d-/ d-s/z d-ç/j d-x/g d- d-C Surface (modified from Hardwick 1984:57) dl d dz d g t' C (C=other consonants)

(187) D-effect from the /id-/ a. i-da 'we got sick' c.f. -ja 'I got sick' 'we hollered' kasi-dz c.f. kas-z 'I hollered' ni-dit 'we got hot' c.f. ni-it 'I got hot' b. i-d 'we're tanning O' 'I'm tanning O' 'we're throwing O' 'I'm throwing O'

(B47) (B46) (B46)

(B40) (B40)

c.f. - di-gu c.f. ds-xu

126 7.2.3 Stem-initial consonants The inventory of stem-initial consonants is given in (188). According to Hardwick, [] and [] do not occur in stem-initial position, "rather the palatal [ç, j] are the sole fricative reflexes for *G" (Hardwick 1984: 10). Stem-initial [j] is derived either from ProtoAthabaskan *j or from a proto "front-velar" (palatal) fricative. (188) Stem-initial segments (Hardwick 1984: 9)84 Lab. Interdental Dent. Pal. Stops b d t t' Affricates Fricatives Sonorants mw d t t' dz ts ts' dl t t' zs l n 'goes on foot' 'louse' 'sky' (KL29) (KL23) (KL23) d t t' çj x w xw h Vel. g k k' gw kw kw' Laryngeal

(189) a. Reflex of Proto glide 'sg. go away' (H20) -ja 'louse' (H20, A10) ja ja 'sky' (H20, A10)

*ja *ja' *ja

b. Reflex of Proto front-velar fricative d-jin 'sing' (H13) *d-n 'sister in law' (H13) -je *e 'grow' (H13) -jan *xa

'sing shamanistically'

'sister in law' 'grow'

(KL24) (H13) (L87:295)

In nouns, stem-initial fricatives show two different voicing patterns. The general pattern is for underlying voiceless fricatives to become voiced when preceded by another syllable. That is, they only surface as voiceless when in absolute initial position. Several examples are provided in (190). Notice that [ç] alternates with [j] in this voicing pattern.


As in (183), this inventory has been rearranged by place.

127 (190) Independent d sg çin çina xos My _____ (Voicing Alternation) 85 sld 'smoke' (B59, H91) es 'belt' (B59, H91) szeg 'spit' (B59, H91) sjin 'song' (B59) sjina 'slave' (B59) soz 'thorn' (B59, H91)

Bob 1999 writes, " The voiced counterparts of the lateral consonant, [], i.e. [l], and the palatal consonant, [ç], i.e. [j], are not fricatives phonetically. However, in Athapaskan languages the lateral consonant [l] and the palatal consonant [j] form a natural class with the fricative consonants..." There are also forms which always surface with a voiced initial that are assumed by Hardwick to be underlyingly voiced. Some examples are provided in (191). While more conservative speakers exhibit the pattern in (190), innovative speakers seem to have lost the voicing rule responsible for the alternation in (190) and simply show a pattern of either all voiceless or all voiced regardless of their environment. Thus, in nouns, voicing of stem-initial fricatives is contrastive. (191) Independent a nd ja ja My ___ (Always Voiced) a 'snow' snd 'geese' 'sky' 'louse'

(B60) (B60, H91) (H91) (H91, A10)

In verbs, voicing of initial fricatives is entirely predictable based on the classifier or other prefix that precedes the stem. All verbs are lexically specified to take one of the following four classifiers. The h-classifier is always followed by a voiceless fricative. The h-classifier actually surfaces in most cases and is therefore easy to identify. The lclassifier, sometimes analyzed as a floating [voice] feature (Hardwick 1984), is always followed by a voiced fricative.86 Unlike the h-classifier, the l-classifier itself never surfaces. The only evidence for its existence is the voicing on the following fricative. After the Ø-classifier, an initial fricative agrees in voicing with the preceding segment. Finally, after the d-classifier, the so-called Athabaskan d-effect causes alternations of the


The data here does not mark consonant harmony. The data out of Hardwick clearly states that these forms given are from earlier sources and therefore do not show the harmony, as it is an innovative process. I assume that the data from Bob is also drawn from these sources, especially since harmony is not the focus of her study. 86 Here, 'voiced fricative' refers to members of the 'voiced fricative class', including [l] and [j].

128 initial fricative (see §7.2.2). Some examples of the h- and l-classifiers are given in (192) and (193). In these examples, the prefixal subject markers ('1p. sg.' = /s/, '2p. sg.' = /n/, '3p. sg.' = Ø, '1p. pl.' = /id/, '2p. pl.' = /ah/, and '3p. pl.' = null or /he/) are underlined. The first person markers undergo consonant harmony (see §7.3) and their fricatives surface as [s], [], or []. The root-initial segments which show predictable voicing patterns are found immediately after the dash. (192) h-classifier (B37) a. h- 'he's tanning O' 'they're tanning O' hh- b. dh-xu hidh-xu h-xws heh-xws 'he's throwing O' 'they're throwing O' 'he's tickling O' 'they're tickling O'


(193) l-classifier (H56) a. kas-ze 'I hollered' 'you hollered' kan-ze ka-ze 'he hollered' kasi-dze 'we hollered' kadah-ze 'you pl. hollered' 'they hollered' kahe-ze b. ni-it nin-it ni-it nei-dit heni-it e-jae en-jae e-jae ei-dae ah-jae ehe-jae 'I get hot' 'you get hot' 'he got hot' 'we got hot' 'they got hot' 'I'm sick' 'you're sick' 'he's sick' 'we're sick' d-effect 'you pl. are sick' 'they are sick'


129 Since the realization of stem-initial fricatives in verbs is entirely predictable, it is possible to analyze them as being either voiced or voiceless underlyingly. 7.2.4 Stem-final consonants Stem-final segments are given in (194). (194) Stem-final segments (Hardwick 1984: 16, 21-23) Stop dt gk gw kw Affricate dz ts d t d t dl t Fricative zs çj x l () () xw Sonorant m n n' n n'


As with stem-initial fricatives in nouns, there are two voicing patterns in the stemfinal obstruents. Some stems exhibit voicing alternations, while others surface only as voiceless. I follow Hardwick 1984 and Bob 1999 in assuming that the alternating obstruents are underlyingly voiced (see (195)), while those that always surface as voiceless are underlyingly voiceless (see (196)). The underlyingly voiced segments devoice in word-final position. As usual, the data come from earlier sources, thus the consonant harmony is not necessarily present or indicated in these forms. An important aspect that emerges from the data in (195) is that [j] alternates with [ç] in the verb 'weep' and the noun 'grease'. Furthermore, since these are cases where the final segment alternates in voicing, they are underlyingly voiced, ending in /j/. (195) Underlyingly voiced final obstruents (alternations) a. t'u -t'ule 'rope' (H18) mi smil '(my) snare' (B64) ese '(my) snow' (H43) e -ei 'itch' (H18) bes esbeze '(my) knife' (H44, B64) kux eskue '(my) rice' (H43) et esede '(my) shirt' (H43) ut esude '(my) whistle' (H43)

130 b. -tsiç teç -tsije esteje tsije k'aje xaje e-aje 'weep' '(my) grease' 'grandfather' 'birch/willow' 'winter' 'root' (H18) (H43, B64) (H22) (H14) (H22) (H46)87

(196) Underlyingly voiceless stem-final obstruents (no alternation) se -see 'hook/gaff' (H18, B65) w w - 'moose thigh' (H18) te estee '(my) cane' (H44, B65) ge esgee '(my) king salmon' (H44) kodgis -kodgise 'poker' (H18) sts 'my arrow' (B65) du esdue '(my) cat' (H44, B65) t'a 'my set line' (B65) skux 'my rice' (B65) Hardwick explicitly lists both [ ] and [ç j] in the chart of stem-final obstruents, but a search of the sources shows that perhaps [ ] should be eliminated from the list. As mentioned above, according to Nater 1989, the only stem with final [] is [-tu] 'fold'.88 In searching all of my sources (see fn. 82); this was also the only form I found with a stem-final []. Therefore, I will not consider this segment as a possible underlying stemfinal segment. As stated in §7.1, the only forms with stem-final [] are the three forms [to] 'paddle' or 'feathers' (H3, 32), [du]/[du] 'cat' (B31/H44), and [ta] 'set line'/[t'a] 'nightline' (B65/N28). Furthermore, both 'cat' and 'nightline' do not show voicing alternations, and are therefore analyzed as underlyingly voiceless. Because of their limited distribution, I have put them in parentheses in (194). As far as I can tell, [] and [ç] are in complementary distribution, [] occurring after back vowels and [ç] after front vowels, though this is in part based on the three forms above. The glide [j] does


These last four forms are listed with stem-final j in Hardwick 1984. It is not clear to me if these forms ever surface without the final [e] or if the glide surfaces as voiceless if it does. The glide in these last four forms is the reflex of *j ("*y" in most sources). 88 Nater 1989 actually assumes the opposite voicing situation as Hardwick and Bob. He assumes that the segments are underlyingly voiceless and voice intervocalically. He argues that stem-final [] is generally "immune" to voicing, except in this one form. The result of both analyses are the same, namely that this word is the only one with final [].

131 occur after back vowels (see (195b)). It would be relevant to this argument to see whether they surfaced as [ç], [], or even [j]. Unfortunately this data is lacking. Stem-final sonorants also occur. Unfortunately, I was only able to find data for one sonorant that was in both absolute final position and medial position. This particular example does not show any voicing alternation (see (197a)). The other forms in (197b) show that voicing among sonorants is contrastive in this position. (197) a. gun esgune b. lin -tin nen' ten' -gane tsme -jan '(my) gold' (H44) 'be (bloody)' (H22) 'lie down' (H22) 'land' (H23) 'meat' (H22) 'arm' (H22) 'son' (H23) 'grow/be wise' (H22)

7.3 Evidence for underlying glides: Consonant Harmony The data in this section comes exclusively from Hardwick 1984 and Shaw 1991. According to Hardwick, consonant harmony (CH) in Tahltan prefixes is an innovative process. It is not just CH that distinguishes innovative from conservative speakers. The innovative speakers also exhibit the innovative voicing pattern (losing the rule of voicing after another syllable) and lack uvulars and stem-final [kw] (Hardwick 1984). As discussed in Shaw 1991, Tahltan is particularly interesting because it has three series of coronal consonants which interact with consonant harmony, as well as two that are transparent. Whereas other examples of coronal harmony tend to involve only alternations between the sibilants [s] and [] (e.g. Navaho and Chumash), Tahltan includes the interdentals in the harmony process. Including these segments means that an analysis based simply on anteriority or stridency is not available. Another relevant aspect of Tahltan CH is the fact that the palatal glide participates in the harmony. Consonant harmony in prefixes in Tahltan is typically triggered by a stem consonant from the D, DZ, or D series and spreads leftward to certain prefixes that contain segments from one of these three series, but it can also be triggered by a prefix and spread to another prefix. Only place of articulation is affected in the harmony; the voicing and manner (fricative vs. affricate) remain unchanged. Members of the D and DL series are transparent to the harmony, acting as neither triggers nor targets of the harmony. Furthermore, all vowels, including /i/, are transparent. Some examples illustrating the transparency of these segments are given in (198). Targets of harmony are given in bold. Segments that do not undergo harmony are underlined. The dash indicates the boundary between stem and prefixes.

132 (198) a. D-series, not a blocker 'I whipped myself' /s/ dd-du 'I'm cutting the hair off' /s/ xa-t'a dsi-t's 'we are walking' // 'we are breast feeding' // mi-t'ot b. D-series, not a target dd-du 'I whipped myself' ta-ta 'I'm dying' dsi-t's 'we are walking' nisi-t'ats 'we got up' c. D-series, not a trigger dedei-dih 'we waved'

(S146) (S146) (S146) (S146)

/s/ /s/ // //

(S146) (S146) (S146) (S146)



d. DL-series, not a blocker/target nod-di 'I melted it over and over' /s/ (S146) ja-t't 'I splashed it' /s/ (S146) e. DL-series, not a trigger da-s-be 'I'm floating'



The data in (198a&d) illustrate the transparency of members of the D and DL series. In these examples, [d, t', , t'] do not block the spread of harmony. For example, in [dsit's] 'we are walking', the first person plural prefix /id-/ surfaces with an [s] as the result of harmonizing with the stem-final consonant. The intervening [t'], though coronal, is not visible to the harmony and allows harmony to pass over it. Similarly, in (198b&d), members of the D and DL series are not targets of the harmony. For example, in [noddi] 'I melted it over and over', the stem initial // does not surface as [] as would be expected if all coronals were targets. Members of these two series also do not trigger harmony, as shown in (198c&e). In these examples, the prefix does not show an alternation. There are five prefixes that undergo harmony. The data that follow in (199)-(211) represent virtually all of the data I was able to compile from Shaw 1991 and Hardwick 1984 for each of the five prefixes. Part of the reason for including so much data is that Hardwick 1984 is not readily available and I would like to make the data available to others who are interested in this process. Patterns with only one or two examples are due to the fact that further data is lacking and not because I have omitted any forms. The first

133 is the first person singular subject pronoun /s/. It surfaces as [] before members of the D-series and as [] before members of the D-series. The data in (199) show that consonants from the DL- and D-series do not cause harmony. They further show that non-coronals have no affect on the prefix. The data in (200) and (201) show that D and D series consonants cause harmony whether or not they are strictly adjacent. Also, the last form in (201a) shows that the glide triggers harmony of an adjacent pronoun. In all of (199)-(211), a dash is placed between the stem and the prefixes, as well as on either side of the prefix that is being discussed. (199) No harmony, [s] idi-s-in a-s-t'e e-s-dan da-s-be nae-s-a e-s-k'a nadd-s-bat s-s-x n-s-t

'I'm angry' 'I am' 'I am drinking' 'I'm floating' 'I'm gambling' 'I'm gutting fish' 'I hung myself' 'I'm going to kill it' 'I'm sleepy'

(H99) (H99) (H99, S144) (H99) (H99) (H99, S144) (S144) (S144) (S144)

(200) a. [] from D-series, adjacent ni--it 'I got hot' naadede--ta 'I dried myself' e--tet 'I'm eating' hodeedi--t'et 'I fell down' ne--ten 'I'm fat' di--t'k 'I heard' e--t 'I'm hot' e--e 'I'm itchy' tede--t'et 'I'm tired' edeni--dil 'I warmed myself' -- 'I'm hot' na--t't 'I fell off (horse)'

(H99) (H99) (H99) (H99) (H99) (H99) (H99) (H99) (H99) (H99) (S144) (S144)

134 b. [] from D-series, not adjacent di--k 'I coughed' (H99) edede--t' 'I cut myself' (H99) hot'a--det 'I rolled myself...' (H99) e--du 'I whipped him' (H99, S144) w de--k 'I cough' (S144) m-- 'I'm wearing (on feet)' (S144) (201) a. [] from D-series, adjacent ene--t 'I'm folding it' di--ta 'I love you' ninane--d 'I moved camp' je--dn 'I sung' hudi--ta 'I love them' --dni 'I'm singing' n--tu 'I'm folding it' e--jae 'I'm sick' n--j 'I'm growing'

(H99) (H99) (H99) (H99) (S144) (S144) (S144) (H99) (S144)

b. [] from D-series, not adjacent nedene--dit 'I moved...a body part' (H99) ja--t'et 'I splashed it' (H99) The second prefix which undergoes harmony is the first person plural pronoun, /id-/. The initial segment surfaces as [s] before members of the DZ and D series and as [] before members of the D series. The final /d/ undergoes the d-effect. Because of the intervening vowel of the prefix, all of the cases of harmony are non-adjacent. This data also shows that [i] is transparent to the harmony, though the palatal glide is not. Data where this glide triggers harmony is given in (204). Also these data show that /i/ is neither a trigger ([dedei-dih], *[dedei-dih]) nor a blocker ([lasi-t's], *[lai-t's]). Presumably, if [i] and [j] were derived from one source, they would pattern together in the harmony.

135 (202) No harmony, // [] dede-i-dih 'we waved' mene-i-dle 'we wanted it' t'an-i-t'a 'we turned around' de-i-gt 'we threw it' ne-i-bin 'we were swimming' (203) [s], all non-adjacent dadene-si-ds la-si-t's t'ande-si-t's d-si-t's na-si-t'ats ne-si-tets edide-si-t'ats tine-si-t'ats de-si-dzel xa-si-dis xa-si-dts edede-si-ts't si-gse hogaa-si-t'as d-si-dzl ni-si-t'ats de-si-dih mee-si-t'ot e-si-tt ene-si-tu

(H100) (H100) (H100) (H100) (H100)

'we're boiling it' 'we 2 came' 'we 2 are walking around' 'we are walking' 'we 2 went on a trip' 'we pl went on a trip' 'we're going hunting' 'we 2 are going away 'we 2 hollered' 'we're plucking it' 'we plucked it' 'we scratched ourselves' 'we're tickling him' 'we're working' 'we shouted' 'we got up' 'we're breathing' 'we're breast feeding' 'we tied it' 'we folded it'

(H100) (H100) (H100) (S145, H100) (H100) (H100) (H100) (H100) (H100) (H100) (S145) (H100) (H100) (H100) (S145) (S145) (H100) (H100) (H100) (H100)

136 (204) [], all non-adjacent a. i-i-çot u-i-de89 teedene-i-dut meae-i-t i-tut ede-i-ta e-i-dae i-dn ka-i-d i-i-tot b. gade-i-je e-i-jae

'we blew' 'we're called' 'we chased it' 'we combed it' 'we grabbed it' 'we loved each other' 'we were sick' 'we sang' 'we pl took pl obj. out' 'we blew it up' 'we 2 are running' 'we are sick'

(H100) (H100, S145) (H100, S145) (H100) (H100) (H100) (H100) (H100) (H100) (S145) (H100) (H100) 90

The third prefix is the conjugation marker /e-/. It surfaces as [s] before the DZ and D series, and as [] before the D series. The data in (206b) show that harmony can be triggered by a prefix, illustrating that the harmony is not simply root controlled, but that it is directional. (205) No harmony (H101) i-un i-lih i-ke meana-i-tin ne-i-n-bin i-t'

'I shot it' 'I tasted it' 'I was sitting down' 'he dreamed' 'you swam' 'I cut it'


These forms can be compared with the morphologically related froms that do not show the d-effect on the stem-initial segment: [u-je] 'one is called' (H75), [teedene-jut] 'I chased O'. 90 Though not explicitly stated in any of the sources I consulted, I believe that the difference between the form 'we are sick' and 'we were sick' has to do with the presence/absence of a classifier, which would explain the fact that one exhibits the d-effect while the other does not.

137 (206) a. [s] triggered by root consonant (H101) si-n-dan 'you're old' e-si-tt 'I tied it' si-h-tsan 'I smelled it' si-bets 'I'm stretching' b. [s] triggered by another prefix (H101) 'I froze up' se-s-ten 'I killed myself' ede-se-es-gin se-s-da 'I was sitting down' ne-se-s-bin 'I swam' tan-se-s-a 'I turned around' medaga-se-s-da 'I waited a long time for' (207) [] triggered by root consonant (H101) d-e-tu 'it's hanging on a pole' The fourth prefix is the unspecified subject pronoun /ts'e-/. It surfaces as [t'e] before D series segments and as [t'e] before D series segments. Again, the glide indirectly triggers harmony in (210). (208) No consonant harmony (H101) naa-ts'e-a 'Tahltan gambling game' nah-ts'i-bat 'one is hanging it' w 'one is scraping it' ts'e-de i kase-ts'i-ts'et 'someone scratched me' mekada-ts'e-ots 'one is shouting at him' (209) [t'e] (H102) t'e-t'a ta-t'e-de-t'a

'one cuts' 'one fleshes'

(210) [t'e] (H102) tedahu-t'e-dene-jut 'one chased them away' The final prefix is the first person singular possessive pronoun /es-/. It is the only prefix discussed in the Tahltan harmony literature that occurs on nouns. Hardwick notes that in this prefix, there is "clear strident harmony" but less evidence of anterior

138 harmony.91 In other words, this prefix surfaces regularly as [e] but that both [es] and [e] are found before stems with D consonants. As with the other prefixes discussed above, this prefix surfaces with [] before segments of the D series, which includes the palatal glide. Nater 1989 also states that it surfaces as [] before j and "3C" (i.e. D) consonants. I have only provided some forms where [] occurs. Again, the glide in 'my rainbow' triggers harmony. (211) e-t'de e-jat'ule e-tiede 'my gristle' (H102) 'my rainbow' (H102) 'my pillow' (H102)

In addition to the dynamic harmony, there is evidence for some static harmony in roots. That is, corresponding words in Proto-Athabaskan (PA) and other Athabaskan languages can have roots which contain segments from different series. In Tahltan, however, some of these stems have undergone alternation of the segments so that only a single series is represented in a root. The static harmony in stems appears to be bidirectional (Hardwick 1984: 105). Of particular interest here is the fact that the word 'snow' shows that the initial palatal glide can be the target of the static harmony. Nater and Hardwick provide the forms in (212c) but give no counterparts in other languages though the implication is that they are also derived from what would be disharmonic roots. (212) a. Right-Left Tahltan 'snow' (H105) 'black bear' ss (H105) 'paddle' eto (N28) PA *jxs (R427) *xs (H105) 92 *t'U (L87:289) Other Lgs. ja (Slave) (H105) tso (Southern Carrier) (N28)


There is also some evidence of an adjacency condition. For example, Hardwick 1984 provides the forms [esed] 'my dress' and [esnde] 'my goose'. 92 Hardwick notes that by regular rules from Proto-Athabaskan to Tahltan, the expected form of 'bear' would be [çs] and therefore ç acts as a target, resulting in [s].

139 b. Left-Right Tahltan 'fold' -tu (H105) Other Lgs. tu (Chipewyan) (H105) -tsoz (Chiricahua Apache) (N28) -tuz (Central Carrier) (N28) -tod (Slave) (H105) -tos (Ft. Ware Sekani) (S. Hargus p.c.)


-to (H105)

c. Other roots in Tahltan 'eat -tet 'edible type of root' ta 'firewood' tsets 'nightline' t'a 'awl' et't 'berries' dde 'rustling noise' zas

(H105) (N28) (N28) (N28) (N28) (N28) (N28)

7.4 Previous Analyses Hardwick 1984, Shaw 1991, and Halle 1995 sketch analyses of Tahltan consonant harmony. Any analysis of the consonant harmony in Tahltan must account for the following facts: (1) members of the D, DZ, and D series are both triggers and targets of the harmony, (2) members of the D and DL series are neither triggers, targets, nor blockers of the harmony, (3) glides act as members of the D series, (4) all other consonants are completely transparent to the harmony, and (5) vowels, including /i/, are completely transparent to the harmony. Hardwick assumes that the segments which participate in the harmony are [+coronal, -lateral] (Hardwick 1984: 103). In her analysis, the features [strident] and [anterior] spread right-to-left from members of the D, DZ, and D series. The problem, however, is that this class of segments includes members of the D-series which are also [+coronal, -lateral]) but do not participate in the harmony. Shaw 1991 argues for the representations in (213). Under this analysis, the D and DL series do not have a Coronal Node in the feature tree and instead are only specified as [-continuant] and [-continuant, +lateral] respectively. Because she adopts radical underspecification, "for consonants [in Tahltan-SVL] the only lexically available feature value is [-continuant]." Therefore, fricatives are unspecified for the feature [continuant]. A further assumption in her analysis is that vowels have features only under the Dorsal Node. Notice that in (213) three different features are used to characterize the three coronal series that interact in the harmony.

140 (213) Root Node [continuant] [lateral] Coronal Node [distributed] [strident] [anterior] [d] · [dl] · + [d] · · + [dz] · · + [d] · ·

Crucial to Shaw's analysis is the notion of underspecification. She assumes that the level of scansion for harmony is the Coronal Node. It could not be the Place Node (the next higher level) because then other segments with place nodes (i.e. vowels and other consonants) would be visible to the harmony. Because her analysis rests on the fact that members of the D and DL series have no Coronal Node, they are transparent to the harmony (see (214)). The other three series (D, DZ, D) do have a Coronal Node and therefore participate. If all coronals were specified with a Coronal Node, then they would all be predicted to interact with the harmony, producing incorrect outputs (see (215)). (214) Shaw's analysis: coronal underspecification / d d s d u / [ d d d u ] (S149) Place · · · · · · 'I whipped myself' = Cor · · [+strid] [+distr] (215) No underspecification / d d s d u / * [ d d d u ] Place · · · · · · · · · = = == Cor · · · · · [+strid] [+distr] A final analysis of the Tahltan data is found in Halle 1995. He analyzes the data using Revised Articulator Theory (RAT).93 Halle's analysis does not require


His brief analysis forms the basis of the analysis I provide in §7.5.4. Our analyses differ in the representation of the affricates, as well as a few other points. Halle's analysis does not consider the interaction of the glide in harmony.

141 underspecification. Halle assumes the representations in (216). Here, the values in angle brackets are noncontrastive. (216) continuant anterior distributed lateral D <+> <-> DL <+> <-> + D + + + <-> DZ + + <-> D + + <-> (Halle 1995:24)

Under Halle's analysis, only contrastive values are visible to the harmony in Tahltan. Harmony spreads all dependents of the coronal node that are visible. An example is given in (217). The medial segment /dl/ has the features [+anterior, -distributed], which are not contrastive, represented in (217) with a lighter grey and smaller font. Because they are not contrastive, spreading across them does not incur line crossing violations. (217) / dz [+cons] [+cont] Place Coronal = = [+ant] [-distr] ... dl [+cons] ... d / (based on Halle 1995: 25) [+cons]

[-cont] Place Coronal

[+cont] Place Coronal



[-ant] [+distr]


Comparison of the representations

The data that need to be accounted for in any explanation of the Tahltan facts are (1) the fact that the high front glide and vowel do not pattern together with respect to harmony, the latter being transparent, (2) the transparency of two 'series' of coronals, (3) the nontransparency of three 'series' of coronals, (4) the transparency of other consonants, and (5) the transparency of vowels. 7.5.1 LEX The major problem with LEX is, as we have seen in the other cases, an inability to distinguish /i/ from /j/. Because the features of these two segments are identical, they cannot be distinguished by the harmony. That is, the segments in (218) share the set of features [G] and should either both allow spreading or both disallow spreading since the features are identical. However, the spreading in (218a) from the vowel is banned, while that in (218b) is allowed. The only way to distinguish the two forms would be to make

142 reference to the syllable structure, but harmony, at least that not dependent upon stress, should interact as an across the board process to any segment that meets the featural requirements. Furthermore, this is a consonant harmony, for which no appeal to prosodic prominence is possible. Because LEX cannot featurally distinguish these segments from each other, it fails to explain a central fact of Tahltan. (218) a.* ... [F] i [G] b. [F] ... j [G]

7.5.2 [CONS] In [CONS] the difference between underlying glides and vowels lies in the feature [consonantal]. All other features are identical, as illustrated in (219). The harmony would have to consider the features (220). Those segments that participate in the harmony are boxed. (219) a. /i/ -cons +son -back -round +high -low b. /j/ +cons +son -back -round +high -low DZ + + + -

(220) Features for consonant harmony in [CONS] D DL D consonantal + + + coronal + + + anterior + + + distributed + lateral + -

D + + + -

The problem with [CONS] is that the underlying glide must have the features in (219b), none of which overlap with those in (220), save the feature [consonantal]. The only way to force the glide to interact would be to posit a separate harmony process that would target the glide, in addition to the harmony process that would target the boxed segments in (220). Creating an additional rule just to account for the glide is unreasonable. Thus, [CONS] is unable to explain a fundamental distinction in Tahltan harmony.

143 7.5.3 Vowel-Place Theory The Vowel-Place Theory (VPT) outlined in this section will be reminiscent of Shaw's proposal (described briefly in §7.4), though it crucially differs from hers in allowing a different level for vowel features and in treating front vowels as coronal. A partial representation of the five series is presented in (221). Only the three series that participate in the harmony have a coronal node. The other two series (D and DL) are underspecified and lack the coronal node. The D series, which consists only of /d t t'/, is specified as [-continuant] only. The coronal nasals /n n' n n'/ must also be represented as members of the D series in order to account for their transparency. The DL series is simply specified as [+lateral]. The differences among the various members within one series are primarily in the laryngeal features and the feature [continuant]. That is, /s z ts ts' dz/ differ in continuancy and the laryngeal settings. Crucially, the glide, which forms part of the D series, is [+sonorant] and [+continuant] but still has the coronal node and the features [-anterior, +distributed], which forces it to interact in the harmony. (221) Representation of coronals in VPT DZ (s, z, dz, ts, ts') D (, , d, t, t') ... ... Oral Cavity Oral Cavity | | C-place C-place | | coronal coronal [+ant] [+distr] [+ant] [-distr]

D (ç, j, d, t, t') ... Oral Cavity | C-place | coronal [-ant] [+distr]

D (d, t, t', n, n') ... Oral Cavity [-cont] C-place

DL (l, , dl, t, t') [+lateral] Oral Cavity | C-place

In order to explain all of the facts of coronal consonant harmony in Tahltan, the coronal node must spread between two segments with a coronal node (see (222)). The original coronal node is delinked from the left segment and the two segments surface with the same values for the features under the coronal node. Segments which are not specified for a coronal node can be neither triggers nor targets, as they lack the coronal node. This rule also predicts that the manner remains unchanged. For example, fricatives remain fricatives even if the trigger is an affricate, since the manner features are not under the coronal node.

144 (222) Characterization of consonant harmony in VPT C-place C-place = coronal coronal [ ant] [ distr] [ ant] [ distr]

Spreading the coronal node accounts for the five facts of Tahltan harmony. The D and DL series neither block nor trigger harmony because they do not contain a coronal node. An example is given in (223). The fact that D and DL series are not targets is addressed by the formulation of the rule.94 Only segments that contain a coronal node can be targets of the spreading. (223) Transparency of D and DL series x a s t' a C-place C-place = coronal [+ant] [-distr] [xa-t'a] C-place coronal [+distr]


Just as the D and DL series are transparent, so too are the non-coronal consonants. The features [dorsal] and [labial] are present in these segments, but they are on different planes and are therefore not visible to the spreading. Similarly, vowels are transparent to the harmony. The C-place of vowels links to the vocalic node, not the coronal node, and is therefore on a different plane. As with the non-coronal consonants, they do not block the spread of the coronal node because their features are on a different plane. (224) Transparency of vowels d ... C-place = coronal (etc) i ... C-place Vocalic V-place Aperture coronal


-dz ... C-place coronal (etc)


If the resulting structure in (223) contains gaps, as it appears to, then the representations in (221) can be amended so that the D and DL series contain no C-place nodes, analogous to Shaw 1991.

145 The non-transparency of the other three series of coronals is also explained. Each of these three series has a coronal node with [anterior] and [distributed] dependants. The spreading rule explicitly targets segments that contain the coronal node. Therefore, any segment with a coronal node, namely any element from these three series will act as a target or trigger. The final aspect of Tahltan harmony that needs to be explained is the different behavior of /i/ and /j/. In VPT, the difference between an underlying glide and an underlying vowel is in the location of the place features. The underlying vowel has place features under V-place. As a result, it is transparent to harmony, as discussed above and shown in (224). The glide, on the other hand, has place features under C-place. The glide has the same representation as the D series in (221). It has a coronal node as well as the two dependents [-anterior] and [+distributed]. Because it has this representation, it interacts with the harmony just as the other members of the DZ, D, and D series do. The spreading rule and representation of the five coronal series does account for the five aspects of Tahltan harmony. This characterization in VPT, however, requires the stipulation that some segments be underspecified, a stance that has fallen out of favor (see Steriade 1995). Only the DL and D series are underspecified for coronal and we are left wondering why the remaining dental segments (i.e. the "DZ" affricates and fricatives) are not similarly underspecified. These segments are at the same place of articulation and should pattern together yet they behave as though they were part of a different series. Also, the motivation for the underspecification is the harmony.95 If the harmony is what motivates the underspecification, then it is not surprising that the underspecification analysis can account for the harmony. Ideally, the motivation for the representation would come from an independent fact about the language, thereby not being circular. 7.5.4 Revised Articulator Theory Revised Articulator Theory (RAT) can also account for the facts of Tahltan harmony, without the stipulations needed for VPT. In particular, there is no need to stipulate that the dental stops, fricatives, and affricates are different 'series'. Instead, their different behavior will fall out from the presence of other features in the segment. The analysis in this section is based on that of Halle 1995 (see §7.4), but goes into significantly more detail. The representation of the coronal segments is of course relevant to the discussion of Tahltan coronal harmony. The representations below in (225) only contain the relevant features. The harmonizing fricatives are represented in (225a). They have one root node and are [+continuant]. The harmonizing affricates in (225b) have two root


Shaw 1991 mentions an earlier 1989 manuscript where she argues that the representations with underspecification are motivated by language-internal factors. Unfortunately, [continuant] is the only feature whose underspecification is discussed in the 1991 article. I was unable to obtain a copy of the 1989 manuscript that presumably discusses the other reasons for assuming underspecification.

146 nodes which contain the manner (articulator free) features.96 In the affricates the first half is [-continuant] and the second half is [+continuant]. The root nodes merge in the place node. The remaining coronals are represented in (225c). The coronal stop /d/ is [-continuant] and contains only one root node. The nasal stops are also similarly represented. The lateral affricate has two root nodes, the first of which is [-lateral, -continuant] and the second of which is [+lateral, -continuant]. The simple lateral is [+lateral, -continuant]. (225) Representation of coronals in RAT z a.



Place [+cont] | Blade [cor] [+ant] [+distr]



Place [+cont] | Blade [cor] [-ant] [+distr]

Place [+cont] | Blade [cor] [+ant] [-distr]

This representation of affricates is taken from Clements & Hume 1995. Hualde 1988 proposes an alternative representation that assumes an unordered set of [±continuant] features for affricates. His representations do not include any feature trees. Lombardi 1990 assumes a single root node with two unordered manner features [stop] and [continuant]. Steriade 1993 and 1994 use three aperture values A0, Amax, and Africative. All of these alternatives, with the possible exception of Steriade's representation, can account for the data. The only thing that is necessary is for the affricates to share some feature with fricatives that is not present in the lateral affricates.


147 b. d dz








[-cont] Place [+cont] | Blade [cor] [+ant] [+distr] c. d

[-cont] Place [+cont] | Blade [cor] [+ant] [-distr] dl [-lat]

[-cont] Place [+cont] | Blade [cor] [-ant] [+distr] l


Place [-cont] | Blade [cor] [+ant] [-distr]






[-cont] Place [-cont] | Blade [cor] [+ant] [-distr]

Place [-cont] | Blade [cor] [+ant] [-distr]

Given the full specification of RAT, it is not possible for the harmony rule to be sensitive to all features because then all coronals would interact. The marking statements which I propose are given in (226). These marking statements yield the contrastive features (boxed) in (227). Because the affricates in (225b) have [+continuant] features, they fall under the description of the marking statements (226b&c). As a result, their [anterior] and [distributed] features are contrastive. The lateral affricate, having no [+continuant] portion, is not contrastive for these two features.97 (226) Deactivated Marking statements a. [-continuant, +lateral] (i.e. [±lateral] is contrastive in [-cont] segments) b. [+continuant, -anterior] (i.e. [±anterior] is contrastive in [+cont] segments) c. [+continuant, +distributed] (i.e. [±distr] is contrastive in [+cont] segments)


Halle 1995 has a similar display of the contrastive features (see (216)) but does not provide any marking statements. He also does not provide representations of the segments that are involved, particularly the affricates which he assumes to be [+continuant]. As for the laterals, he also assumes that they are [-continuant].

148 (227) Feature specification of Tahltan coronals dz z d/n d continuant lateral anterior distributed -+ + + + + + + -+ + + + ç + + j + +

dl - + + -

l + + -

d -+ +

The transparency of the D and DL 'series' is evident from the feature specifications and the marking statements. These segments, bearing no [+continuant] feature, are not contrastive for the two coronal features [anterior] and [distributed]. The rule of coronal harmony is formulated to only be sensitive to contrastive features (see (228)). The inertness of /d, n, dl, l/ is explained by a separate mechanism, and is therefore not circular as was the case with VPT. Their lack of interaction is due to the fact that their features are not contrastive, which is in turn due to the structure of the inventory. We are not required to stipulate a special series where /d/ behaves differently from /z/ and from /dz/. Its behavior is different because it has no [+continuant] portion. The harmony rule is provided in (228). As stated, it is only sensitive to contrastive features. It spreads contrastive dependents of the Tongue Blade node to preceding segments which have contrastive features under the Tongue Blade node. These features then are delinked. This rule also ensures that only the place and not the manner of the target is changed. (228) a. View contrastive features. b. Place Place Blade = = [cor] [ ant] [ distr] Blade [cor] [ ant] [ distr]

The rule in (228) accounts for the five aspects of Tahltan consonant harmony. The transparency of non-coronal consonants and of vowels is explained by their lack of Tongue Blade features. Since each place node is on a different plane, non-coronal consonants and vowels (which are dorsal in RAT) cannot interact with the harmony. The transparency of the [-continuant] coronals in Tahltan is due to the fact that their Tongue Blade dependents are not contrastive and therefore not visible to harmony. As illustrated in (229), the features of medial /t'/ are invisible, allowing the features of // to spread across them. The fact that coronals with a [+continuant] feature are targets and triggers

149 of the rule is also illustrated by the form in (229). Because their Tongue Blade features are contrastive, they participate in the harmony. (229) Transparency of some coronals x a s t' [+cont] Place Blade = = [+ant] [-distr] [-cont] Place Blade [+cont] Place Blade




[coronal] [+ant] [-distr]


[-ant] [+distr]

The final aspect of Tahltan harmony that must be accounted for is the different behavior of the underlying glide /j/ and the vowel /i/, which is not discussed in Halle 1995. Vowels, having Tongue Body features, do not participate in coronal harmony because they lack the harmonic features. The underlying glide /j/, on the other hand, has the representation in (225). Because it is [+continuant] and its coronal features are contrastive. It interacts in the harmony just as the other coronals do. The different behavior of these two vocoids is captured by their different designated articulator: [dorsal] for the vowel and [coronal] for the glide. By using the representations in (225) and the marking statements in (226), it is not necessary to characterize Tahltan as having five series of coronals. Instead, the split in the behavior is due to whether a given segment has a [+continuant] portion or not. The stop /t/ should, for all practical purposes, form a class with /s/ and /ts/ if place were the relevant factor. Consonant inventories in Tahltan are generally listed with series, a practice I depart from in (183) by putting in a traditional place-based inventory. The fact that some coronals participate in the harmony while others do not is due to other features. I believe that eliminating the need for series in Tahltan is an asset because it explains the harmony in independent terms. That is, the inventory itself dictates which features are contrastive. The contrastive features in turn specify which segments may participate in other phonological processes. Thus, RAT, in contrast to VPT, is not circular. 7.6 Discussion Of the four representations, LEX and [CONS] are unable to account for a key fact of Tahltan coronal harmony, namely the different behavior of /i/ and /j/. The other two representations, VPT and RAT, are able to accommodate the five facts of harmony. VPT relies on underspecification to explain the transparency of the D and DL series. RAT, on the other hand, accounts for their transparency by utilizing the system of marking statements and contrastive feature specification outlined in Calabrese 1995. The major difference between the two representations is that in RAT, the invisibility of the D and

150 DL series is due to the fact that the structure of the inventory prohibits them from being contrastive. In VPT, on the other hand, the underspecification is primarily motivated by the harmony; thus explaining the harmony with the underspecification is circular.


8 Karuk

8.1 Introduction Karuk is a Hokan language spoken in northwestern California (Bright 1957).98 Karuk is also discussed in Herman 1994 and Hume 1995 because of the different behavior of two labial glides. In this chapter, I will argue that Karuk has two surface labial glides, one derived from an underlying vowel, and a second derived from an underlying glide. Karuk also has a surface palatal glide, which patterns as a derived glide and is not phonemic. Karuk vocoids exhibit two different patterns. The labial vocoids show a Type VI distribution, while the palatal glide shows a Type III pattern. (230) Type VI /u/ /w/ [u] [w] Type III /i/ [i] [j]

Section 8.2 will focus on syllabification in Karuk and the distribution of labial and palatal glides and high vowels in order to prove the complementary distribution and predictability of these two segments. Section 8.3 will examine the evidence that an underlying labial glide exists in Karuk. The palatal glide and vowel will also be discussed in Chapter 11. The consonant phoneme chart is given in (231), adapted from Bright 1957. The data in this section are all taken from Bright 1957. (231) Consonant phonemes of Karuk (adapted from Bright 1957)99 Labial Dental/Alveolar Palatal Velar Obstruents p t k t f x s Sonorants m n w r




In previous literature, the language is also referred to as Karok. According to Herman 1994 (note 1), Karuk is preferred by native speakers. 99 Bright 1957 also lists <> as a phoneme of the language. However, he points out that the output [] is completely predictable by a general rule that palatalizes /s/ after front vowels and [j], except in a few forms which are mostly borrowings (e.g. [murasis] 'molasses'). I have therefore omitted it from the phoneme chart.

152 The palatal glide is not listed in the phoneme inventory in (231) because in virtually all cases, the distribution of the glide and the vowel (j vs. i) is completely predictable based on syllabification. The only place where the distribution of glides and vowels is not predictable is in word initial position. This will be discussed further in §8.2. Bright, however, does believe the palatal glide is an underlying phoneme of the language (Bright 1957; personal communication). If the glide is not a phoneme, then it is the case that many words in Karuk contain underlying vowel sequences, though no surface vowel sequences exist. In the theory espoused in this dissertation, whether or not a surface vocoid can be predicted based on environment determines its phonemic status. Therefore, it is possible that underlying vowel sequences exist, but are not present on the surface as a result of gliding. The fact that the surface distribution of j~i is predictable in Karuk indicates that these two segments are in fact derived from one source (see §8.2 on syllabification in Karuk). For this reason, the palatal glide is omitted from the phoneme inventory. The transcription of the labial glide in the phoneme inventory in (231) departs from that used in Bright 1957, but follows Herman 1994 and Hume 1995 in using the symbol <w>. I will use the symbol <w> for the underlying glide and for both surface glides. Bright uses the symbol <v> for both surface "glides" and writes that they are bilabial spirants (IPA []) (Bright 1957; personal communication). Indeed on the surface, they are fricated. However, the underlying glide patterns as a sonorant phonologically. Thus, the symbol <w> will be used for this phoneme. Evidence that the glide w is a sonorant comes from three areas. The first piece of evidence comes from distribution. If the glide were truly a fricative, then it would be the only voiced fricative in the inventory (and also the only voiced obstruent). According to Maddieson 1984 when languages include a voicing contrast, it is generally used across all fricatives, and most likely to include the voiced counterpart of /s/ (that is, /z/). The absence of other voiced fricatives, and in particular, the absence of /z/, makes it unlikely that this segment is an underlying fricative. Second, Karuk differentiates length among consonants. One degree of length is the result of an allophonic process triggered by the preceding or following vowel. In this case, the consonant is slightly longer than a simple singleton consonant and is marked C. The second degree of length is even longer and is phonemic (e.g. [nukuh] 'deep', [sukkux] 'a woman's name'). The segments [, h, w, r, j] are "nongeminable", and never occur with either degree of lengthening (Bright 1957: 17). These segments form the natural class of non-nasal sonorants. If the labial glide were truly a fricative, then its presence in this group would be anomalous.

Bright 1957 classifies the two laryngeals as obstruents. I have elected to place them with the sonorants in the inventory since they pattern with the non-nasal sonorants in not being allowed to surface as geminates..


153 Finally, surface [r] (from /r/) and surface [w] (from /w/ in this analysis) become nasal stops when they occur before another consonant at a morpheme boundary. These form the class of underlying supralaryngeal non-nasal sonorants.101 That is, when these segments occur in coda at a morpheme boundary, they surface as nasal [n] and [m] respectively. This will be discussed in more detail in §8.3.1. The vowel phonemes of Karuk are given in (232). (232) Vowel phonemes of Karuk (adapted from Bright 1957) i i u u e o a a Two types of accent are transcribed on vowels. The acute accent / / is a "combination of strong stress and a high level pitch" (Bright 1957: 11). Before a pause, it is realized as strong stress, falling pitch, and partial glottalization on the vowel. The circumflex accent / / is realized as strong stress and falling pitch. 8.2 Syllabification in Karuk Syllabification of sequences of high vowels is completely predictable in Karuk. I will show that word-internal sequences of high vowels consistently surface as GV and not *VG even though both should be allowed based on syllable shape. I will also show that word-final sequences of high vowels consistently surface as VG and not *GV. This type of directional syllabification can be accounted for in a non-OT approach (Mester & Padgett 1994). Though a standard directional syllabification algorithm can also account for the data, I have elected to illustrate syllabification in Karuk using OT. Before turning to the actual OT analysis of Karuk syllabification, a few facts concerning vowel and consonant clusters must be discussed. First, no tautosyllabic consonant clusters occur on the surface. That is, no words begin or end in a sequence of two consonants. Bright states that syllables in Karuk are of the shape CV(C). Word medially, sequences of two consonants do occur. Because no more than two consonants can occur medially, I assume that all medial clusters are syllabified between the two consonants, with one in the coda of the preceding syllable, and one in the onset of the following syllable. Glides must occupy syllable peripheral positions as well, since both

The absence of the other non-nasal sonorants /h, / from this process may be related to their ability to be nasalized in the first place. Since the acoustic difference between oral and nasal [] is null during the segment, it is possible that it is actually nasalized as well. Similarly, oral and nasal [h] is a difficult contrast to maintain, though apparently it exists in a few languages (Walker & Pullum 1999).


154 *VCCGV and *VGCCV never occur.102 If glides could surface as nuclear diphthongs, then these sequences would not be problematic. Second, heterosyllabic vowel sequences are also not observed on the surface, implying a ban on vowel hiatus. In the analysis presented here, underlying sequences of vowels do occur, but they are avoided on the surface through either deletion or coalescence (across a morpheme boundary) or through gliding (within a morpheme if one vowel is high). When both of the vowels are non-high, the presence of a sequence of vowels is only observable at morpheme boundaries as in (233a). Since non-high vowels are not allowed to surface as glides in Karuk, any underlying tautomorphemic vowel sequences in Karuk will always undergo deletion or coalescence, providing no evidence of their existence in the first place. When two vowels come together at morpheme boundaries, however, they undergo deletion in (233a) or coalescence in (233b). (233) a. Deletion [aho] [-a] to walk ­ deverbative b. Coalescence [jêpa] [-iip] good ones ­ best


'the act of walking'


'best ones'

The remainder of this section will focus on vowel sequences that contain at least one high vowel. A survey of the 1706 forms glossed in the Karuk-English dictionary of Bright 1957 shows that, with the exception of word initial high vowels or glides,103 all

My search of Bright's glossary found no cases of more than two consonants (including glides) intervocalically. 103 The word initial sequences of glides and vowels are slightly problematic, but are intertwined with another problem of the glottal stop. Words can begin with GV or with VG sequences, for example, [ijtup] 'to knead' and [jikihi] 'to be seriously sick'. Clearly these two examples are problematic for an account that assumes a predictable syllabification. If these two words begin with /iiCV/ then there is no way to predict whether the first or the second segment surfaces as a glide. The status of the glottal stop might shed some light on this question. A glottal stop is predictably present word initially in 'vowel-initial' words. This means that the examples actually begin with [ijC] vs. [jiC]. It is difficult to determine whether these words are vowel initial and have a predictable glottal stop or whether these words actually have an initial glottal stop underlyingly. Bright 1957 considers a phoneme, though his view has changed (pc). If the is underlying, then the words with [VG] are completely predictable because gliding the first vowel is prohibited by *COMPLEX. With this assumption, has the distribution of other word initial consonants. Syllable final is not predictable, occurring in forms such as /ma/ 'uphill', /sa/ 'downhill', /ka/ 'upriver', but here the syllable-final contrasts with h and Ø. Essentially, this means that the status of is not clear. If the words are V-initial and is simply inserted on the surface, then these forms are problematic for


155 vocoid sequences can be explained with simple syllabification constraints (e.g. avoiding clusters, avoiding hiatus, the inability of long vowels to surface as glides104, avoidance of non-high glide, etc.). From the first 200 entries, I found the following medial sequences containing a surface glide: 10 next to a long vowel VG or GV, 4 with a non high vowel in VG or GV, 2 in forms with a high vowel adjacent to a glide where gliding the other vowel would result in a banned cluster (e.g. VCCVGCV), and 2 forms with an intervocalic glide where gliding a different vowel would result in hiatus. In final vowel sequences in the first 200 entries, I found 10 final sequences of a glide next to a non-high vowel, 2 examples of a glide next to a long vowel that was high, 2 examples with adjacent high vocoids where gliding the other would have resulted in a banned cluster, and 1 with a sequence of three high vowels which surface as VCGVG where the alternative VCVGV would have an additional syllable and therefore an additional *STRUC violation. Out of 200 entries, there were 38 entries containing sequences of vowels and glides. 33/38 were predictable based on general syllable tendencies. The remaining 5 examples were predictable based on the directional syllabification (see the discussion after (240)). It is clear that surface glides (even [j]) are not simply transitional or epenthesized to avoid hiatus because they occur in the context VGCV and VCGV (e.g. [ijtup] 'to knead', [imjat'] 'fur'). A simple analysis of the forms in Karuk using Optimality Theory will account for all of the types of data. When an underlying vowel sequence contains a high and non-high pair, there is only one repair strategy; the high vowel must surface as a glide. Some data are provided in (234). (234) High vowels adjacent to non-high vowels a. /takai/ [takaj] 'to be disgusted by b. /tupiau/ [tupjaw] 'to sort' c. /taiau/ [tajaw] 'to choose' For this type of data, only the basic constraints given in (235) are needed. The constraints *COMPLEX, ONSET, {A}=V, and DEPSEG are never violated in the language and are therefore undominated. In monomorphemic forms, segmental deletion is banned, thus MAXSEG is highly ranked.

predictable syllabification. If they are -initial words, then the words with initial [GV] are predicted in order to fulfill ONSET. 104 In calculating these distributions I assumed that long vowels are not allowed to surface as glides. This is perhaps due to the fact that /i/ [j] incurs two violations of MAXµ while /i/ [j] incurs only one. The data from Sanskrit in Chapter 11 show that it is possible for long vowels to surface as glides.

156 (235) a. MAXSEG: Segments in the input must have a corresponding segment in the output. (Do not delete.) b. DEPSEG: Segments in the output must have a corresponding segment in the input. (Do not epenthesize.)105 c. *COMPLEX: Complex onsets and codas are disallowed. d. *STRUC: All syllabic structure is banned (one violation is incurred for each syllable). e. ONSET: Syllables must have onsets. f. {A}=V: Non-high vowels must be parsed as syllabic (Rosenthall 1994). g. VOWEL-MORA (V-µ): For every vocalic root node, there exists a mora (Rosenthall 1994:26). (Underlying vowels should not surface as glides.) Tableau (236) provides an example of an underlying vowel sequence where one of the vowels is non-high. The constraint *V#, defined in (245) is included here to ease comparisons with future tableaux. The alignment constraint which will be used for directional syllabification (ALIGN-R) has been omitted, as it is the lowest ranked constraint and is not relevant to Tableau (236). At this point, no directional syllabification is necessary because there is only one possible output of the vowel sequences. The vowel sequence in (236) is not allowed to surface unchanged. Forms where all of the underlying vowels surface in separate syllables are banned because of the highly ranked ONSET constraint (see (236b&c)). Deletion of one or more of the vowels is also not a possible outcome, due to MAXSEG. Thus, the only possible solution to the input vowel sequence is to allow the high vowels to surface as glides. The winning candidate violates the low ranking V-µ constraint in allowing the underlying vowel to surface as a glide. Since this constraint is so low-ranked, it will not be included in subsequent tableau. (236) /tupiau/ [tupjaw] 'to sort' ONSET *COMPL {A}=V a. tup.jaw b. tu.pi.a.u c. tu.p_a.u d. tu.p_a_ e. *!* *! *!



*V# * * *

* *!*

*STRU ** *** ** ** ***

V-µ *


There are some underlying sequences of high vowels whose realization can be accounted for by the few constraints defined in (235). Some examples are given in (237). In these examples, the competing outputs where the other vowel surfaces as a glide (i.e.


I am using MAXSEG and DEPSEG as cover constraints for both MAXC, MAXV, DEPC, and DEPV.

157 *[ju], *[taxuwk], *[kwirak]) are not possible because they all violate *COMPLEX and are therefore ruled out. As stated above, glides do not form nuclear diphthongs in Karuk. If a they could, sequences like *[ju] and *[VCCGV] would be allowed. Diphthongs are ruled out by a high ranking constraint *DIPHTHONG (constraint from Rosenthall 1994). (237) Sequences of high vowels a. /iu/ [iw] b. /taxuuk/ [taxwuk] c. /kuirak/ [kujrak] (238) /kuirak/ [kujrak] ONSET a. kuj.rak *! b. ku.i.rak c. kwi.rak d. ku.ji.rak e. k_i.rak f. ku.wi.rak106

'to be, to lie' 'to hook' 'three' *[kwirak] DEP MAX

'three' *COMPLEX


*! *! *!

*STRUC ** *** ** *** ** ***!

Examples like those in (237) are not the only forms with sequences of high vowels. Basic syllabification accounts for all of the forms in (234) and (237), where outputs with fewer syllables are preferred over those with more syllables, where hiatus is disallowed, where deletion is banned, and where tautosyllabic consonant clusters are barred from surfacing. However, two types of outputs cause problems for this simplistic approach. They are forms with medial high vowels surrounded by only one consonant on either side /...VCUUCV.../ and forms with a final sequence of high vowels preceded by only one consonant /...VCUU#/. Here, /U/ designates a high vowel. In each of these forms, our analysis therefore predicts that two outputs are possible. That is, either the first high vowel could surface as a glide ([...VC.wu.CV...] and [...VC.wu#]) or the second could ([...V.Cuw.CV...] and [...V.Cuw#]). Notice that both of these possible outputs are licit, given only the constraints given in (235). That is, neither incurs violations of the highly ranked ONEST, *COMPLEX, DEPSEG, or MAXSEG. Furthermore, they incur the same number of *STRUC violations. Given only these constraints, there is no way to distinguish these two possible outputs.

In this form, I assume that the [w] is not a separate slot/segment, but is the result of the linking the initial vowel to the onset of the following syllable. This explains the lack of a DEP violation.


158 (239) /...VCUUCV.../ ? *COMPLEX a. [...VC.wu.CV...] b. [...V.Cuw.CV... (240) /...VCUU/ ? *COMPLEX a. b. [...VC.wu] [...V.Cuw] ONSET DEPSEG MAXSEG




A closer look at the distribution of words with medial, underlying sequences like those in (239) shows that there is an overwhelming preference for the first high vowel to surface as a glide in Karuk (i.e. VC.wu.CV). 19/24 (79%) of the words with this form have a glide in the first position. The remaining 5/24 (21%) of the forms have a glide in the second position. Crucially, in these five forms, the second glide is always [w]. Thus, these forms do not actually cause a problem, because they can be analyzed with an underlying glide as the second high vocoid. (Recall that Karuk has /w/ but no /j/.) Thus, it would be impossible for the other adjacent high vocoid to surface as a glide (e.g. /CVCiwCV/ *[CVCjwCV]) because the output would incur multiple *COMPLEX violations. Some examples with medial sequences of high vowels are given in (241). (241) a. First vowel surfaces as a glide /imuira/ [imwira] 'fishery' /imiuha/ [imjuha] 'soap plant' /suniiih/ [sunjiih] 'nut of a giant chinquapin' b. Second vowel surfaces as a glide /uiriwsaw/ [wiriwaw] 'to bequeath to' If the form in (241b) were /uiriusau/ then it should surface as [wirjusaw] which it does not. The forms in (241a) show the need for right-to-left directional syllabification. In OT, the addition of an alignment constraint serves this purpose (Mester & Padgett 1994). This constraint in (242) ensures that the first vowel surfaces as a glide. (242) ALIGN(, R, PrWd, R): Align the right edge of each syllable with the right edge of the prosodic word. One violation is incurred for each segment that intervenes between the right edge of the syllable and the right edge of the prosodic word. Tableau (243) provides an example where the alignment constraint distinguishes between gliding the first and gliding the second high vowel. Candidates (a) and (b) both incur three violations of *STRUC because they both have three syllables. However,

159 candidate (a) only incurs six violations of ALIGNR, while candidate (b) incurs a total of seven violations of ALIGNR. The crucial distinction between these two candidates is that (a) incurs only four violations of ALIGNR from the first syllable, while (b) incurs five. Thus, the gliding of the first vowel forces the preceding consonant into the coda, while gliding the second does not. (243) /imiuha/ [imjuha] 'soap plant' *[imiwha] *COMPLEX ONSET DEP MAX *STRUC a. im.ju.ha b. i.miw.ha c. i.mi_.ha d. i.mi.u.ha *** *** *** ****

*! *!

ALIGNR 1 2 3 **** ** ***** **! **** ** ***** *** **

Forms such as those in (241a) require ALIGNR, but those in (241b) do not because there is no sequence of high vowels. There is only one way to realize the bolded /i/ in (244) and that is as a vowel. (244) /uiriwsaw/ [wiriwaw] *COMPLEX a. b. wirj.waw c. d. wir._waw 'to bequeath to' 107 ONSET DEP MAX

*STRUC *** ** *** **

*!* *! *! *

ALIGNR 1 2 ****** *** **** ***** *** ****

Forms with a sequence of high vowels word finally (/...VCUU#/) also require an additional constraint. As mentioned above, there are two plausible outputs of these forms, [...VC.Gu] where the first vowel glides and [...V.CuG] where the second vowel glides. Though either seems possible, only the second occurs (11/11 forms).108 Some examples are provided in (246).

The constraint that rules out vocalization of the underlying glide /w/ is *NUCLEUS/GLIDE, which disallows the underlying glide from surfacing as a syllabic nucleus. (This is the same as the former PEAK constraints.) The *NUC constraints are based on the sonority hierarchy and are universally ordered as *NUC/STOP >> *NUC/FRIC ... *NUC/GLIDE >> *NUC/VOWEL (McCarthy 2002: 22). In Karuk, the constraint *NUC/GLIDE is simply highly ranked since underlying glides are never allowed to surface as syllabic. The candidate [wir.jw.saw] incurs a violation of the high ranking *NUC/GLIDE. 108 In theory, we might expect to find /...VCwU/ [...VC.wu] with an underlying glide. I cannot explain the absence of such forms and suggest that they must be accidental gaps.


160 (245) *V#: Words should not end in a vowel. The absence of hypothetical forms such as [tatju] requires the additional constraint *V#. This constraint disallows final open syllables; thus forms in which the final vowel surfaces as a glide are preferred.109 Some examples are given in (246). Because MAXSEG is ranked above *V#, words with a final non-high vowel will never opt for an output with a deleted final vowel. Thus, a final non-high vowel always surfaces. A final non-high vowel will not surface as a glide because of the highly ranked {A}=V. (246) /afiu/ [afiw] /tatui/ [tatuj] /iuui/ [iwuj] /kuiiu/ [kujiw] 'to make the bottom of a basket' 'to sweep' 'to cost' 'place name' *[afju] *[tatwi] *[iuwi] *[kujju]110

Tableaux (247) and (248) provide an example of a final sequence of high vowels. Notice that the addition of *V# is necessary in order to eliminate candidate (247b) where the first vowel surfaces as a glide. Indeed, candidate (247b) is the predicted form given ALIGNR and would be expected as it is parallel to the medial sequences of high vowels discussed above. However, the fact that the second vowel surfaces as a glide requires the *V# constraint. Thus, *V# outranks ALIGNR. (247) /tatui/ [tatuj] 'to sweep' ONSET DEP MAX a. ta.tuj b. tat.wi c. ta.tu.i d. ta.tu_ *[tatwi] *V# *STRUC ** ** *** ** 1 *** ** *** **



*! *!

*! * *


(248) /kuiiu/ [kujiw] *COMPLEX

'place name' ONSET MAX




Chitoran 1997 uses a similar constraint for Romanian: *Pi/u#: Word-final high vowels must not be parsed as syllable peaks (Chitoran 1997: 192). The constraint in (245) is more general. The fact that only high vowels are affected by this constraint is due to its interaction with {A}=V. 110 This form is not ruled out because it has a long vowel in a closed syllable; witness the word 'to be afraid' [a.wa].


161 a. ku.jiw b. kuj.ju c. kwi.ju d. ku.ji.u e. ku.j_u ** ** ** ** *** ** ** **

*! *! *!

*! * *

The presence of the constraint *V# does not result in bizarre outputs when there is only a single high vowel preceded by an unambiguous consonant in word final position. Because *COMPLEX, DEP, and MAX all dominate *V#, the only reasonable output is one in which the final vowel simply surfaces as a vowel in final position. As tableau (249) shows, the only possible output is one with a final vowel. (249) /kari/ [kari] 'still' *COMPLEX a. ka.ri b. kar_ c. ka.rit d. karj



*V# *

*STRUC ** * ** *

ALIGNR 1 2 ** ***

*! *! *!

Now that the basic syllabification in Karuk has been outlined, it is clear that the surface form of an underlying high vowel is fully predictable. When it is adjacent to a non-high vowel, it surfaces as a glide. When it is surrounded by consonants, it surfaces as a vowel. Syllabification in medial position favors a right-to-left syllabification, captured by ALIGNR. In final position, however, the right-to-left syllabification of a sequence of high vowels produces the wrong output with a final vowel. The existence of a constraint which disprefers final open syllables accounts for this fact. In this way, underlying high vowels surface as vowels or glides in predictable ways, such that the surface form has the optimal syllabification. 8.3 Evidence for underlying glides Surface glides in Karuk are derived from three different sources. One set of surface labial glides [w] is derived from /w/ and shows similarities with consonants in the language. The other two glides [w] and [j] are derived from /u/ and /i/ and pattern with vowels. Though many of the surface glides can be derived from underlying vowels by general syllabification (as seen in the previous section), there are some that cannot. Example (241b) was already one example of this kind, where the second high vocoid surfaced unexpectedly as a glide word medially [...VCuGCV...]. The few forms that fit this pattern were analyzed as having an underlying glide and were therefore not subject to the same sort of syllabification.

162 In addition to this distributional fact, there are several differences between the surface glides that are derived from vowels and those which are derived from glides (Bright 1957, Herman 1994, Hume 1995). Where possible, the same stem will be used to illustrate the combined evidence for underlying glides (e.g. 'to fall' in (253a) and (257b), 'to seek' in (250a) and (259b). Bright 1957 clearly points out the difference between the two surface labial glides with respect to Sonorant Nasalization and their behavior before vowel-initial suffixes. These facts are also used by Herman 1994, Hume 1995, and here in this dissertation. Bright mentions the vowel harmonic suffixes, but does not explicitly mention the difference between the two surface glides. He is, however, consistent in marking the two surface glides with different diacritics so it is clear that he interprets the vowel harmonic suffixes as behaving differently when adjacent to the different glides. Again, the harmony is used as evidence by Herman 1994, Hume 1995, and here. Suffix allomorphy of the future tense is mentioned in Bright 1957, but not in Herman 1994 or Hume 1995. None of Bright 1957, Herman 1994, nor Hume 1995 discusses the predictability of syllabification. The examples in the following sections are taken from Bright 1957. Some of them are taken directly from his discussion of the particular phenomenon, while others were found by reading the glossary and the remainder of his grammar. 8.3.1 Sonorant Nasalization Underlying supralaryngeal non-nasal sonorants (glides and r) become nasals before any morpheme that begins with a consonant (Bright 1957). In this process, /r/ and /w/ surface as [n, m] (see (250), while underlying high vowels that surface as glides do not become nasals (see (251)). (250) a. /apiw/ [apiw] [apim-tih] /asiw/ [asiw] [asim-tak] /iw/ [iw] [im-kara] /u-aw/ [u-aw] [am-kira] b. /kur/ [kûr] [kûn-taku] /sir/ [sir] [sin-kara] 'to seek' 'to be seeking' 'to sleep' 'to close one's eyes' 'to die' 'to drown' 'he eats' 'place for eating' 'to sit' 'to sit on' 'to disappear' 'to swallow'

*[apiw-tih] *[asiw-tak] *[iw-kara] *[aw-kira]

*[kûr-taku] *[sir-kara]

163 (251) a. /asuu/ [asuw] [asuw-tih] /ikriu/ [ikriw] [ikriw-tak] b. /axai/ [axaj] [axaj-tak-i(rih)] 'to grumble' 'to be grumbling' 'to sit' 'to sit in the way' 'to take from' 'to seize'

*[asum-tih] *[ikrim-tak]


Thus, the underlying glide patterns with the other non-nasal sonorant. The two derived glides [w] and [j] pattern together in not alternating with a nasal. This shows a difference between the underlying glide /w/ which does become a nasal, on the one hand, and the derived glides [w] and [j] which do not become nasals on the other. The segments that alternate in (250) must be /w/ and /r/ and not /m/ and /n/. Stem-final /r/ and /w/ contrast with final nasals in word final position. (252) [patum] [sam] [arin] [in] 'to put one's mouth or head' 'to remain, be left' 'to sit, stand' 'glands in the throat'

8.3.2 Vowel-initial suffixes The second difference between the two labial glides is their behavior before vowel initial suffixes. The underlying glide /w/ patterns with /r/ (and the other consonants) in remaining unchanged on the surface before a vowel-initial suffix. Examples with final underlying glides and consonants are given in (253). (253) a. /ikiiw/ [ikjiw] [ikjiw-i(rih)] [ikjiw-ûr] /iw/ [iw] [iw-apuh] (c.f. [im-kara] 'to drown') b. /istur-ahi/ [itur-ahi] /parih-a/ [parih-a] 'to fall' 'to fall down' 'to fall for a long time' 'to die' 'dead' 'to be cracked' 'rain (noun)'

Whereas the morpheme-final glides are allowed to surface before vowel-initial suffixes, final high vowels that surface as glides pattern with other vocalic segments and are deleted before vowel initial suffixes. Final vowels are deleted before vowel-initial suffixes, as shown in (254a).

164 (254) Final /V/ [V] /tuphi-a/ [tuph_-a] 'speech' /ina-is(rih)/ [in_-i(rih)] 'to kneel down' Even when the final vowel is one that could and does surface as a glide when unsuffixed, it is still deleted before a vowel initial suffix. These forms are given in (255a&b). The vowels that remain after the deletion coalesce (e.g. ikriu-at ikri-at ikret). (255) a. Final /u/ [w] /ikriu/ [ikriw] /ikriu-is(rih)/ [ikrî(rih)] /ikriu-at/ [ikrêt] /ihiiu/ [ihjiw] /ihiiu-unis/ [ihjûni]

'to sit' 'to sit down' 'lived' 'to shout' 'to shout at'

b. Final /i/ [j] /iiuai/ [ijwaj] 'to pour' /iiuai-is(rih)/ [ijwê(rih)] 'to pour down' As was the case in Sonorant Nasalization, the derived labial and palatal glides pattern together to the exclusion of the underlying labial glide. Whereas the true final glides (i.e. underlying) surface intervocalically at a morpheme boundary, derived glides behave like vowels. 8.3.3 Vowel harmony in suffixes The third difference between the underlying and derived glides is their effect on vowel harmonic suffixes. Karuk has several suffixes in which the suffix-initial vowel harmonizes with the preceding vowel of the stem to which it attaches. If the stem final vowel is front, the harmonizing vowel is [i]. If it is back and not low, the vowel is [u]. If the stem final vowel is low, then the harmonizing vowel is [a] (see (256)). (256) Final V in stem: i, e u, o a Harmonic/copy vowel: i u a

As with the data in the preceding sections, the underlying glide patterns with the rest of the consonants. With these harmonic suffixes, the stem final underlying glide or

165 consonant is transparent to the harmony. (257a) shows the harmony in forms that end in a consonant, while (257b) has forms with a final underlying glide. The harmonic vowel is symbolized by <V>.111 Notice that the final underlying glide is transparent to the harmony and does not require the harmonic vowel to surface as [u]. (257) a. /taknih-Vuna/ [taknih-iuna] 'to roll around' /ikfuk-Vwra/ [ikfuk-uwra] 'to climb over' b. /ikiiw-Vwra/ [ikjiw-iwra] /pikiiw-Vwra/ [pikjiw-iwra] 'to fall into a sweathouse (sg.)' 'to fall backwards'

Alternatively, if the stem ends in a vowel, then deletion and coalescence occur. Every example attested in Bright's data of a vowel-final stem with a harmonic suffix is included in (258). (258) a. /axau-Vwruk/ [axawruk] (c.f. [axaw] 'to collapse') b. /aha-Vuna/ [ahauna] /iwrara-Vwra/ [iwrarawra] /aha-Vwrik/ [ahawrik] c. /u-Vwrin/ [uwrin] 'to slide down a bank' 'to drive around' 'to fall into a sweathouse (pl.)' 'to head off' 'to turn around'

In (258a), the stem-final vowel surfaces as a glide when no suffix is attached. However, when the vowel-initial harmonic suffix is attached, the final vowel is deleted and the remaining vowels coalesce (axau-Vwruk axa_-Vwruk axa-wruk). In (258b), there is some ambiguity. I assume that the same process of deleting the final vowel takes place in these examples. Then the harmonic vowel surfaces as a copy of the preceding vowel. In all of the cases with root-final [a] plus a harmonic suffix in Bright 1957, the final two vowels of the root are identical, so it is impossible to determine if this analysis is correct. It is, however, consistent with the data and the process of vowel deletion (e.g. aha-Vuna ah_-Vuna ah-auna). The example in (258c) also exhibits the expected pattern. When the final vowel of the stem is long, it is then the second vowel that deletes (c.f. aho-ar aho-r 'to go travel'). Therefore, in this example, the harmonic suffix vowel deletes and the long vowel of the root remains.112


In Bright 1957, the harmonic vowel is notated by 0, but I follow Herman 1994 and Hume 1995 in using

V. There are only six examples of vowel-final stems with harmonic suffixes in Bright 1957. Five are given (258). The remaining example /on-wa-Vuna/ [onwauna] 'to take (people) around' is an exception


166 The number of stems ending in [j] are not numerous and their behavior with harmonic suffixes has not been established (Bright, pc). I would predict that they would pattern like the form in (258a), triggering harmony. 8.3.4 Allomorphy The final evidence for the difference between underlying /w/ and derived [w] is allomorph selection. The future suffix has two allomorphs [=awi] and [=e].113 Stems that end with a derived glide (i.e. underlying vowel) or a vowel take the former allomorph, while consonant and glide final stems take the latter.114 Consonant final and /w/-final forms are given in (259). The forms in (259b) clearly end in an underlying glide, as evidenced by its alternation with a nasal and with its inability to delete before vowel-initial suffix. Again, both consonant final and glide final forms take the [=e] allomorph. (259) a. Verb final C [u-if=e] [kunituk=e] [uwaxrah=e]

'It will grow' 'They will pick' 'It will dry'

b. Verb final /w/ [u-aw=e] 'He will eat' (c.f. [am-kira] 'place for eating=table') (c.f. [aw-iruw] 'to each too much') (c.f. [u-aw] 'He eats') [nup-apiw=e] 'We will seek' (c.f. [apiw] 'to look for') (c.f. [apim-tih] 'to be looking for')

to the basic process of vowel deletion. The expected output would have deleted the final /a/ (on-wa-Vuna on-w_-Vuna) and then shown a high copy of the stem vowel, as [onwuuna]. This, however, is not the correct output. This form differs from the other examples in that it is morphologically complex, with the suffix /-wa-/ 'plural action'. A complete analysis of what happens between vowels at morpheme boundaries is clearly warranted, but unfortunately cannot be undertaken in this dissertation. 113 The double hyphen is used for 'syntactic affixes' (Bright 1957: 29). 114 There are some vowel-final forms listed in Bright's grammar that appear to take the [-e] allomorph. However, I believe that this is due to regular processes. All of the vowel final stems that take this allomorph end in [a]. This is perhaps analyzed as /...a-auis/ a-uis (merging of the two [a]s) a-_is (via regular process that deletes u/i at a morpheme boundary) [e] (coalescence). Unfortunately a full analysis of these two phenomena is absent in Bright 1957.

167 Stems that end in an underlying vowel take the other allomorph of the future [=awi]. This includes stems that end in a derived glide when not suffixed or before consonant-initial suffixes. In all of these forms, one of the vowels at the morpheme boundary deletes and the remaining vowels coalesce. For example, in /ihruu-awis/ the medial /u/ deletes and the remaining /u/ and /a/ coalesce to form a long [o]. In particular, the forms in (260) show again that the derived [j], the derived [w], and vowels pattern together to the exclusion of the underlying glide. (260) a. Verb final V [niaho=wi] 'I will walk' (c.f. [niaho] 'I walk') b. Verb final /u/ [w] /u-ihruu-awis/ [uhrô=wi] (c.f. [uhruw] 'he uses') (c.f. [ihruw] 'to use') /u-ikjau-awis/ [ukjâwi] (c.f. [ukjaw] 'He makes'

'He will use'

'He will make'

c. Verb final /i/ [j] /ni-axai-awis/ [niaxawi] 'I will take' (c.f. [axaj-taki(rih)] 'to grab, seize') (c.f. [niaxaj] 'I take') 8.3.5 [s]~[] allophony Palatalization of /s/ is one other piece of evidence that shows that the derived palatal glide [j] and the vowel [i] pattern together. According to Bright 1957 [] occurs after [j] and front vowels even if there is an intervening consonant.115 For palatalization to occur, no vowel can intervene between the trigger (/i/, /e/) and the target (/s/). In this way, other vowels act as blockers to this palatalization process. Consonants, on the other hand, can intervene and are transparent. The palatalization is unidirectional, not occurring if the vowel follows /s/, as in [apsih] 'lag', *[apih]. The trigger (or potential trigger) is in bold in (261).


There are a few forms for which do not have [] after a front vowel, e.g. [keks] 'cake' and [simsim] 'knife'. Because of these few forms, Bright 1957 assumes an underlying //. The former word, however, is clearly a borrowing and the latter looks like it might be a reduplicated form.

168 (261) a. /s/ immediately following [i]/[e] [tiih] 'dog' [piri] 'grass' [pi] 'soaked acorns' [ike] 'river' b. Intervening C is transparent [pikip] 'shadow' [ikup] 'to point' [ikus] 'to tear up a rat's nest' [arip-uru] 'to cut a strip off' [imut] 'spark' [it-uru] 'to leave, abandon' [pinura] 'beans'116 [usiwap] 'he plugged it up'117 c. Intervening V between [i] and [s] blocks palatalization [i tkus] 'walk arm in arm' [ihkus] 'deceased mother' d. /s/ immediately following [j] [tujip] 'mountain' e. Intervening V between [j] and [s] blocks [firipamjusah] 'place name'

These are the only two forms I found with an intervening coronal. Bright says that 'beans' is a borrowing and that [ura] is not identifiable. Forms with labials and velars were abundant. 117 This form is somewhat interesting and problematic. The suffix [-sap] does exist with the meaning of closing, e.g. [ikrup] 'to sew' and [ikrup-sap] 'to sew shut', though it is only found in a few forms (Bright 1957: 94). If this form contained this suffix and the [w] surfaces, then the implication might be that this is not an underlying glide (because the underlying glide would surface as a nasal). Then it must be the underlying vowel. Then the fact that it does not block the allophony is problematic. On the other hand, the form [siwap] 'to plug up' does not seem to have another related form like <siw>. Perhaps, this form [siwap] was lexicalized before a rule of w m. Since there exist sequences medially of w in coda that I must assume are /w/, as in (241b).


169 Unfortunately, the nature of Karuk syllable structure prevents an example in which [j] and [] are not adjacent. For example, a hypothetical form like [ajpa] has an illicit syllable structure; it violates *Complex. Since no example of this sort can exist, the only possible occurrence of [s/] and [j] are when the segments are strictly adjacent [...j...] or when a vowel intervenes [...jVs...]. Thus, it is impossible to test if the palatalization of [i]/[e] and [j] are the result of a single process or two different ones. It seems unlikely that two these are two different processes given the fact that it is clear that [i] and [j] are derived from the same source. 8.3.6 Summary From the above data we can conclude the following. First, Karuk has a surface labial sonorant that derives from two different sources. One source is the underlying glide /w/ that patterns with /r/ in final position by surfacing as a nasal before consonant-initial suffixes. It also patterns with other consonants when in final position before vowel initial suffixes, and harmonic suffixes. Finally, it patterns with consonants in allomorphy selection. The second source of surface [w] is the vowel /u/. It surfaces as a glide when this is the best syllabification. This [w] patterns with the surface palatal glide [j] and also with vowels in allomorphy selection, harmony, nasalization, and heteromorphemic vowel processes. 8.4 Previous analyses of glides in Karuk The crucial difference between the analysis presented here and previous analyses of Herman 1994 and Hume 1995 lies in the underlying source of output glides. Whereas this chapter assumes a single underlying labial glide, both Herman 1994 and Hume 1995 assume two distinct underlying labial glides which they distinguish as vocalic and consonantal. Their 'consonantal glide' is what I have argued to be an underlying glide. In their analysis, the 'vocalic glide' is also underlying but patterns with other vocalic segments. In the analysis proposed here, on the other hand, the second glide is merely derived from /u/ via regular syllabification processes. The fact that it derives from a vowel /u/ explains the fact that it patterns with other vowels. There is no evidence that this vocalic glide is underlying. On the contrary, all evidence points to its allophonic status. 8.5 Comparison of the representations There are three characteristics of Karuk phonology that must be explained by the four models. First, they must account for the fact that the two underlying sonorants /r/ and /w/ surface as nasals before heteromorphemic consonants, but that other sonorant codas derived from underlying vowels (/i/ [j] and /u/ [w]) do not. Recall that it is not simply that non-nasal sonorant codas are banned, since they do occur word finally. Second, they must explain the realization of vowel-initial harmonic suffixes. Finally, they must explain how surface [i], [j], and [e] all cause palatalization of a following /s/.

170 The other phenomena in Karuk are based on syllabification and cannot be used to distinguish the validity of the four models. 8.5.1 LEX In LEX, all features are identical between underlying glides and underlying vowels. The only difference is that the underlying glides are lexically marked to be non-nuclear. This means that the glides are forced into a non-syllabic position. If an underlying vowel surfaces as a glide, then it should pattern with the underlying glide in all other phonological phenomena. Clearly this is not the case in Karuk. For example, the process which turns sonorants into nasals would not be able to distinguish between the two forms in (262). Recall that Sonorant Nasalization targets the sonorants /r/ and /w/ and causes them to surface as the nasals [n] and [m] before another consonant. The problem with LEX is that once the two different underlying vocoids are syllabified their underlying difference is completely neutralized and they are expected to pattern identically. (262) a. Underlying representation N * / a s i u/ /ikriu/ b. Output [asiw] [ikriw]

c.In the environment of Sonorant Nasalization [asim-tak] *[ikrim-tak] (predicted), [ikriw-tak] (correct) As discussed in Chapter 3, the basic problem with LEX is that it predicts that all high vocoids that end up in non-nuclear positions will behave the same. Clearly, this is not corroborated by the evidence. The phonology of Karuk can distinguish glides that are derived from underlying vowels from those that are derived from underlying glides. 8.5.2 [CONS] Using the feature [consonantal] to distinguish between underlying glides and vowels can account for the data in Karuk. First, consider Sonorant Nasalization. Since the process targets /w/ and /r/ to the exclusion of /u/ and /i/, it is possible for the rule or constraint to utilize [consonantal] to differentiate these sets. A sample rule and constraint are given in (263). (263) a. +cons [+nasal] / ____ -[+consonantal] +son

171 b. *C-C: A non-nasal sonorant is banned before another heteromorphemic consonant.118 | -nasal +son Since the underlying glide is [+consonantal], it is possible to distinguish it from the underlying vowel, which is not. Since the underlying glide has both labial and dorsal features it has two potential nasal outputs [m] and []. However, only [m] is a segment in the language and therefore is the realization of the underlying labial glide /w/. [CONS] can also account for the suffix harmony. In consonant final stems, the suffix vowel takes on the backness of the preceding vowel (see (256)). An example of a set of rules that could account for this is given in (264). The rule in (264a) produces a high suffix vowel [i] or [u] after i/e and u/o respectively. The rule in (264b) produces a low vowel after a low vowel. Since the glides in [CONS] are specified as [+consonantal], they form part of the transparent set of consonants that intervenes between the stem vowel and the harmonic vowel. (264) a. -cons [ back] / -cons [+consonantal]0 - ____ +high -low back b. -cons +low / -cons [+consonantal]0 - _____ +high -high +low Harmony constraints in OT generally rest on the assumption of [CONS] by simply stating that vowels must agree in some feature (see Chapter 1, example (2)). By assuming that underlying glides are [+consonantal], OT can also account for their transparency. The drawback to [CONS], and in fact the challenge in VPT and RAT, is that it requires either multiple rules or multiple constraints. The representation in [CONS] also accounts for the s-palatalization. Because the underlying glides are represented as [+consonantal] they can be forced to pattern with other consonants. In s-palatalization, a consonant can intervene between the trigger (/i, e/) and the target (/s/) but a vowel cannot. The transparency of the glide is accounted for in the same way the transparency of the consonants is. It is simply stipulated as transparent in the rule. An agreement constraint could similarly be formulated.


In this constraint, C is simply shorthand for a [+consonantal] segment.

172 (265) -son +cont +cor [-ant] / -cons [+consonantal]0 ____ -back

8.5.3 Vowel-Place Theory Vowel-Place Theory can also account for the data in Karuk. Abbreviated representations of /w/, /u/, and /i/ are provided in (266) and (267). (266) /w/ Root | C-place

labial dorsal (267) a. /u/ Root | C-place | Vocalic | V-place labial dorsal b. /i/ Root | C-place | Vocalic | V-place | coronal

In the Sonorant Nasalization, a [+sonorant] segment with C-place features becomes a nasal before another consonant. This is illustrated in (268). (268) [+son] C-place [+nasal] [+cons] | C-place

(labial) (coronal) (dorsal) The two sonorants that have C-place features are /w/, which has [labial] and [dorsal] features, and /r/, which has [coronal] features. Vowels are not targets of this process because they do not have C-place features. Similarly, if the segments /h/ and // are considered sonorants, they will also not be targets of this process since they lack C-place features. Notice that the [labial]/[dorsal] segment /w/ surfaces as the [labial] nasal and deletes its [dorsal] feature. This output, though, is not surprising, given that the inventory of Karuk includes only the two nasals /n/ and /m/.

173 Accounting for the harmonic/copied vowel requires some explanation. The spreading rule in (269) below accounts for the surface forms [i] and [u] of the harmonic vowel. (269) Spreading Rule in VPT V C C-place C-place | Vocalic | V-place V C-place | Vocalic Aperture | [+high]

The two front vowels /i/ and /e/ have the V-place feature [coronal]. When the V-place is spread to the suffix vowel, it surfaces as [coronal] and [+high], namely [i]. Similarly, for the two back vowels /u/ and /o/, the V-place features [dorsal] and [labial] are spread to the suffix vowel which surfaces as [u]. This spreading rule also explains the transparency of consonants and of the underlying glide. Since these segments only have features under C-place, they are invisible to the spreading rule. The realization of the suffix vowel as [a] after /a/ requires some explaining. One possible way to account for this part of the harmony would be to stipulate that the first vowel in (269) be [-low]. Then a second spreading rule could account for the cases when the triggering vowel is [+low]. This is not as satisfying since it requires two separate rules. An alternative to resorting to two spreading rules is say that the [dorsal] feature of the low vowel spreads to the suffix vowel. However, the expected high back vowel [] is not present in the language and the repair is to change the height of the suffix vowel from [+high, -low] to [-high, +low]. Yet another alternative is to assume that the low vowel /a/ is placeless. Using examples of height harmony, Parkinson 1996 argues for the constraint in (270) that requires that non-low vowels have a place feature (i.e. something that makes them front or back). If the low vowel were placeless, then an empty V-place node is spread to the suffix vowel. If the suffix vowel were to acquire a place feature, perhaps as a default, then it would violate DEPPLACE, which would ban the insertion of a place feature.119

If the low vowel were placeless, then a reasonable question is why a preceding vowel (e.g. ...iCaC-V) could not spread through the placeless vowel. A possible reason for the non-transparency of /a/ is that the harmony might be prosodically driven. In all of the examples, a prosodically prominent vowel precedes the copy. Eleven cases have an acute accent on the vowel which immediately precedes the suffix. Two examples have monosyllabic stems with a long vowel that has a circumflex accent on the stem vowel. One final example has an acute accent two syllables before the suffix, but both of the final stem vowels are identical ([pikwirip-iwrin]). This leads me to believe that a situation like [iCaC-i] would not involve the


174 (270) HEIGHTPLACE (Parkinson 1996: 33) If a vowel is specified for an occurrence of [closed] (i.e. is not low--SVL), it must also be a front or back vowel. Of these three explanations for stem-final /a/ copy, the first is the least appealing. It is not ideal to have two different rules for what appears to be a single process. I will not choose between the remaining two (repair of [] vs. placeless /a/). Suffice it to say that both can explain the vowel copy/harmony in Karuk. The final process that must be dealt with is s-palatalization. The palatalization is somewhat problematic and requires some additional machinery. A first attempt at the palatalization is presented in (271). Spreading a V-place feature to a C-place is allowed in the standard VPT model (Clements & Hume 1995: 295). Spreading a V-place dependent explains the transparency of intervening consonants. Since they lack vocalic features, they do not block the spreading of V-place features.120 The problem with such a representation of palatalization is that it also incorrectly predicts the transparency of intervening vowels. Back vowels lack both [coronal] and [anterior] and therefore would not block the spread of [coronal] or [anterior] since each place feature is on a different plane. (271) S-palatalization in VPT: First attempt V C s C-place C-place C-place | | | Vocalic [feature] [coronal] | = V-place [+ant] | [coronal] | [-anterior]

spread of features from the preceding /i/ across the placeless /a/ since spreading appears to be prosodically bounded (i.e. only prosodically prominent vowels spread). 120 It is actually not clear to me if intervening consonants are actually transparent. If another coronal were between the trigger and target, as in [...ins...] then it might be a line crossing violation for the vocalic features to spread across the [anterior] feature of the nasal. A possible fix would be to say that the dental/alveolar coronals are not specified for [anterior]. However, an intervening /t/ would still be a problem since it would have a [-anterior] feature.

175 An obvious alternative is to spread a higher node. In this case, the next helpful higher node is either the V-place node or perhaps the Lingual node.121 But if the Lingual node is spread from the V-place to the C-place of a following /s/, then an intervening consonant which also has a lingual node should incorrectly block this spreading. Even if the target is specified as [coronal, +cont, -distrib], an intervening coronal is expected to block (see (272a)). A similar result holds if the V-place node as a whole spreads. (272) S-palatalization with a lingual node ? V n s C-place C-place C-place | | Vocalic Lingual Lingual | | | V-place [coronal] [coronal] | | Lingual [+anterior] | [coronal] | [-anterior] An alternative is for the V-place or the Lingual node under V-place to spread indirectly to the following /s/. It spreads to the following fricative and then there is interpolation of the remaining nodes (see (273)). (273) S-palatalization with interpolation (V-place or Lingual Nodes) V n s C-place C-place C-place | | Vocalic [feature] (Vocalic) [coronal] | [+anterior] V-place | [coronal] | [+anterior]

Chapter 4 on Turkish discusses the possible need for a Lingual node between C/V-place and the place features. Clements & Hume 1995 also discuss evidence for the Lingual node in both C- and V-places.


176 The resulting segment would have a major articulation and a minor articulation, resulting in [sj]. The parenthetical nodes are due to interpolation. In this case, there would need to be a repair to create the output []. Spreading either the Lingual or V-place nodes via interpolation correctly predicts that other vowels, which also have these nodes, will block the spreading. It also predicts that consonants and the underlying glide are transparent. There are no line crossing violations. 8.5.4 Revised Articulator Theory Like VPT, Revised Articulatory Theory (RAT) can also explain the data from Karuk. Abbreviated trees representing /w/, /i/, and /u/ are provided in (274) and (137). (274) /w/ Root | Place

Body | [lab] [+rd] [+bk] (275) a. /i/ Root | Place Body Lips b. /u/ Root | Place Body [dor] [+bk] [-lo] [+hi]


Lips [-rd]

[dor] [-bk] [-lo] [+hi] [+rd]

In the case of Sonorant Nasalization, the rule is similar to that in VPT. In (276), a sonorant that is followed by a consonant at a morpheme boundary becomes [+nasal]. In (276), the three place nodes are in parentheses to show that only one must be present. (276) Place [+son] Soft Palate [+cons] ...

(Lips) (Blade) (Body) [+nasal]

177 The designated articulator is what determines the realization of the nasal. Underlying /w/ has a designated articulator [labial], thus its nasal realization is [m]. Similarly, /r/ surfaces as [n]. Other sonorants in this position, for example the derived glides [w] and [j], cannot surface as a nasal because their designated articulator is dorsal and dorsal nasals are banned in Karuk. If the glottals are also considered sonorants, their lack of nasalization is also explained by the fact that they lack the correct designated articulators [coronal] or [labial]. We now turn to the harmonic suffixes. According to the marking statements in Calabrese 1995, a standard five vowel inventory has the contrastive features (boxed) given in (277). Given the spreading rule in (278), all of the surface patterns of the harmonic/copy vowel are explained. (277) Vowels in RAT i high low back round + e a + + o + + u + + +

(278) a. View contrastive features. b. [-cons] ... [-cons] | | Place Place | | Body Body [dor] [hi] [bk] [lo] [dor] [+hi]

The presence of [+high] on the harmonic suffix serves to block the spread of [-high] from the preceding mid vowels. Though RAT requires full specification, I have argued that it should be possible to allow harmonic vowels to be partially unspecified (see Chapter 4 for more discussion). Since the underlying glide is not contrastive for the features under the Tongue Body node, it does not block spreading. Similarly, the consonants are not contrastive for any of the Tongue Body features and are therefore also transparent. A schematic representation of how the harmony progresses is provided in (279). Here, the contrastive features are given in bold. In (279a&b), the feature [-back] (and [-low] for /e/) spreads to the suffix vowel. Since the only high front vowel is [i], the suffix surfaces with this vowel. In (279c&d), the feature [+back] (and [-low] for /o/)

178 spreads to the suffix vowel. It surfaces as the only high back vowel in the language, namely [u]. The presence of [+high] on the harmonic vowel prevents the spread of [-high] from the mid vowels. (279) a. +high -low -back -round b. e C -high -low -back -round c. u C +high -low +back +round d. o C -high -low +back +round e. a C -high +low +back -round i C V X +high iCi

V X +high


V X +high


V X +high


V +high


As mentioned in the previous two sections, to ensure a low harmonic vowel after [a] requires extra statements. In the case of RAT, the explanation is simple. The universal marking statement [+low, +high] is active. In this language, the repair strategy is to delink the [+high] feature. That is, in (279e) the copy vowel would surface with the features [+high, +low], but this is disallowed. Instead, the [+high] is delinked and the other compatible vowel features are filled in to produce [a].

179 A brief excursus on the locality of this harmonic/copy vowel is necessary. If only contrastive features are visible, then a pattern like that in (280) is expected. (280) a. +high -low -back -round b. a C -high +low +back -round i C -high +low +back -round i C +high -low -back -round a C V +high


V +high

Since the intervening features are not contrastive, we expect that the preceding features should be able to spread across the stem-final vowel. The resulting vowel would be [+high, +low, -back] in both (280a&b). The problem is that in (280a), some repair should result in [a], while in (280b) some other repair should produce [i]. The problem is that in one case the resulting [+high, +low, -back] vowel must change to [-high] and [+back] while in (280b) it must change these to [-low]. The trick to solving this problem is to realize that the copy interacts with the prosody. As discussed in footnote 119, the vowel which serves as the trigger is accented. Therefore, restricting the copy environment prevents the situation in (280) since the final vowel of the root is prosodically dominant. The final process of relevance is s-palatalization. The basic intuition is that the front vowels spread [-back] to the following /s/, generating []. If only contrastive features were visible, then the low vowel /a/ is expected to be transparent since it is not contrastive for [-back]. For example, a form like /iras/ could erroneously surface as *[ira]. However, if all features are visible to the spreading rule, then the underlying glide /w/ is incorrectly expected to block palatalization, since it is underlying [labial, +round, +back]. There are four ways to fix this problem. The first is to restrict the domain of palatalization to the preceding syllable. This will avoid the problem of /iras/ [ira] since the trigger is not in the immediately preceding syllable, but would not avoid incorrectly predicting /ias/ *[ja]. A second alternative is to restrict the domain of palatalization to the preceding [-consonantal] segment with a designated articulator [dorsal]. The rule can be sensitive to contrastive features and only spread [-back] to the following /s/. The underlying glide, which does not have contrastive [back], will be transparent. A third analysis is to argue that the feature [back] is actually contrastive on the low vowel /a/, contrary to Calabrese's conventions. In fact, allowing /a/ to be contrastive for [back] clears up the harmony process and the situation in (280) would no longer be a problem. A fourth alternative is to appeal to the fact that the realization of the

180 one underlying glide is actually [] and not produced with a dorsal articulation. In this scenario, the s-palatalization would be a later process, perhaps post-lexical, occurring at a point at which the underlying glide /w/ no longer has a dorsal articulation. Unfortunately, I was unable to find any examples where the palatalizing environment occurs across a word boundary. 8.6 Discussion The four representations handle the facts of Karuk to varying degrees. As usual, LEX falls short of accounting for the data since it is unable to maintain a distinction between underlying and derived glides once they have been syllabified. [CONS] is capable of handling the data. One reason that [CONS] actually works for this language is that the processes simply make a division between vowels and consonants. [CONS] had difficulty with Turkish since the harmony cut across clear consonant versus vowel distinctions. It had trouble with Tahltan and Pulaar because it was necessary to go to the featural level to target glides. In Karuk, the only relevant aspect is the glide's affiliation with consonantal segments. Sonorant Nasalization targets [+consonantal] segments. Harmony and s-palatalization do not affect [+consonantal] segments. Allomorphy selection looks for [+consonantal] segments. Both VPT and RAT can account for Sonorant Nasalization and the harmonic suffixes. In dealing with s-palatalization, both feature geometries must stipulate a repair from [sj] to []. In order to simultaneously deal with the transparency of the glide and the non-transparency of the low vowel /a/, RAT must either limit the domain of spalatalization or appeal to the surface phonetic quality of the underlying glide.


9 Sundanese

9.1 Introduction Sundanese is an Austronesian language of Indonesia spoken in West Java. Sundanese exhibits a well-known process of nasal harmony where nasality spreads from a nasal consonant to subsequent vowels (Robins 1957, Cohn 1990/1993). This spreading, however, is blocked by any supralaryngeal consonants (including glides) or a word boundary (Robins 1957: 90). Sundanese may actually be better suited in Chapter 10 among the languages which cannot help to distinguish between the four representations of underlying glides. We will see that none of the four models does a superb job explaining the behavior of transparent versus opaque segments in nasal harmony. The approach of Nasal Harmony detailed in Walker 1998 and 2003 described in §9.3.5, however, uses a universally ordered nasalizability scale similar to the sonority hierarchy. Since Sundanese presents a different type of evidence for underlying glides, I have included it in a separate chapter. Sundanese represents a Type VI language where underlying vowels surface as vowels and glides surface as glides. (281) Type VI i j i j

The segment inventory of Sundanese is given in (282) and (283). On the surface, [] also appears, but it is not phonemic, occurring only between sequences of like vowels and at some morpheme boundaries. (282) Phoneme inventory of Sundanese (Cohn 1990) Labial Alveolar Palatal Velar Stops pb td kg c 122 Fricatives s Nasals m n Liquids lr Glides w j Laryngeal h


Cohn 1992 writes that /s/ patterns as a palatal.

182 (283) Vowel inventory (Cohn 1992) i u e o a Roots in Sundanese are generally disyllabic with a shape like that in (284). Here, O represents any consonant. R1 is a nasal that is homorganic with the following consonant or /r/. R2 can be most consonants. C2 is /l/ or /r/ after a stop and is quite rare. (284) Root pattern (Cohn 1992: 205) (O) V (R1) (O) (C2) V (R2)

9.2 Evidence for underlying glides: Nasal harmony Evidence for underlying glides in Sundanese comes from their behavior in Nasal Harmony. In Sundanese, nasalization spreads rightward from a nasal stop, through vowels and the glottal consonants [h] and []. Nasalization is blocked by the presence of any supralaryngeal consonant, including glides. In order to confirm previous descriptions, Cohn 1990 measured the amount of nasal airflow produced during each segment of a word. The amount of nasal airflow is related to the amount of constriction in the oral cavity. For low vowels, even with a lowered velum, there is little airflow through the nasal cavity because the constriction in the oral cavity is not great and therefore does not force air to be rerouted through the nasal cavity. Therefore, Cohn avoided low vowels in her study. Her results confirm that glides, along with other consonants, block the spread of nasalization. The data in (285ac) show that nasalization spreads through vowels, derived glides, [], and [h]. It is true that [] is not phonemic and that one could account for its transparency via rule ordering. However, [h] is phonemic and patterns with [] as transparent. (285d-g) show that obstruents, liquids, and underlying glides block the spread of nasalization. (285) a. /iar/ b. /mi-asih/ /niis/ /nuus/ /aatkn/ [ijar] [miasih] [niis] [nuus] [aatkn] 'seek (active)' 'love (active)' 'relax in a cool place (active)' 'dry (active)' 'dry (active)'

183 c. /mahal/ /mihak/ /aho/ /nuhurkn/ d. /atur/ e. /isr/ f. /uliat/ /marios/ g. /awih/ /mawur/ /ajak/ [mahal] [mihak] [aho] [nuhurkn] [atur] [isr] [uliat] [marios] [awih] [mawur] [ajak] 'expensive' 'take sides (active)' 'know (active)' 'dry (active)' 'arrange (active)' 'displace (active)' 'stretch (active)' 'examine (active)' 'sing (active)' 'spread (active)' 'sift (active)'

In her work, Cohn uses the brackets |...| for the phonological output. The form in (285a) is represented as |iar|. The actual surface form has a transitional glide, which Cohn represents as [ijar]. These glides will be compared with the true underlying glides later. Not only do glides pattern with other supralaryngeal consonants in blocking the spread of nasal harmony, but several near minimal pairs exist between glides and vowels. In (286a), the underlying glide surfaces as a glide, while the underlying vowel in (286b) remains a vowel. (286) a. /najak/ |najak| b. /aian/ |aian| 'sift (active)' 'wet (active)'

The existence of minimal or near minimal pairs of vowels and glides does not in and of itself show the existence of underlying glides. Given the two outputs |najak| and |aian|, there are two possible hypotheses for the underlying forms. The first hypothesis is that there is a distinction between underlying vowels and underlying glides, the hypothesis assumed by the representations in (281) and (286). The alternative hypothesis is that Sundanese is an example of a language that has a distinction between underlying vowels which can alternate with glides and those which cannot (see Chapter 11). Consider for a moment this second analysis. If this were true, then the forms from (286) would actually have the representation in (287).

184 (287) a. /n a i a k/ |najak| N | b. / a i a n/ |aian| Let us entertain for the moment that the representations in (287) are correct for Sundanese. Under such an analysis, the only difference between the surface vowel and glide is their position in the syllable. Featurally, these segments are identical because they both derive from an underlying vowel, in this case /i/. That (287b) is prohibited from gliding does not affect its featural content. Thus, given that they have the same features, what accounts for their different behavior with respect to the spread of nasalization? The only difference between (287a&b) is the syllable position of the vocoid. Thus, the blocking would have to be due to the presence of a segment in the onset of the syllable in (287a). This is clearly not a viable analysis. Because nasalization spreads through the onsets [h] and [] in (285b-c), it is not possible to seek a syllablebased explanation to the blocking effects of the surface glides. Because syllable position is the only difference between [j] and [i] in (287) and because the onsets [h] and [] do not block the spread of nasalization, it is not possible to analyze of the facts of nasalization in Sundanese with the representation in (287). The nasal harmony that is described here differs from what we think of as vowel harmony in two ways. First, while vowel harmony can sometimes be initiated by a consonant, the general pattern is for the trigger to be a vowel. In Sundanese, the nasal harmony is only triggered by a nasal consonant (there being no underlying nasal vowels). Second, vowel harmony is a relationship between adjacent vowels where consonants are transparent. This is what makes it a process of vowel harmony and not an interaction between consonants and vowels. The general pattern is for consonants to be transparent to the vowel harmony. In the nasal harmony described here, the general case is for consonants to block the spread of nasalization. I bring up this point here to illustrate that appealing to a syllable-to-syllable approach to harmony (see Chapter 4 on Turkish) is unable to account for the spreading of nasalization. The analysis with underlying glides presented in (286) does not face the same problems as that in (287). Because they are two distinct segments, the glide and the vowel do not necessarily have the same featural representation. The featural differences between these segments will be dealt with in §9.3 and depends on what theory of features and feature geometry is adopted. The forms in (286) show a distinction between underlying glides and vowels. Now we will consider the difference between underlying glides and transitional glides. Cohn argues that the forms in (288a) derive from underlying /VV/ sequences while those in (288b) derive from underlying /VGV/ sequences.

185 (288) a. [ijar] [awur] [ajijan] b. [iwat] [mawur] [ajak] 'seek (active)' 'say (active)' 'wet (active)' 'elope' 'spread' 'sift (active)'

This transitional glides in (288a) surface as nasalized and allow nasalization to spread rightward through them. On the other hand, in (288b) the glides block the spread of nasalization. Cohn 1990 argues that the glides in (288a) are simply phonetic transitions between adjacent non-identical vowels. She bases her arguments on two acoustic measures. First, in the spectrograms, there is only slight weakening of the formant structure during the transitional glides, while there is "definite weakening of the formant structure and a prolonged transition" during the underlying glides (Cohn 1990: 65). Second, the duration of the VGV sequences is shorter than that of the VGV sequences (244 ms vs. 278 ms).123 Under Cohn's analysis, the transitional glides are not phonological and therefore do not have any status with respect to the rule of nasal spread. Still, we should recognize that just because these glides are transitional does not necessarily mean that they do not function in the phonology. Their participation in nasal harmony can be easily explained in any case. The transitional glides are derived from vowels and therefore have the same features as the vowels. Two possible representations of a derived glide are shown in (289). (289) a. N i N a r i a r b.


Since the transitional glides and surface vowels have the same features, they are expected to act the same with respect to nasal harmony since it is not sensitive to syllable affiliation. Indeed, Hume 1995 assumes that these transitional glides do have some status in the phonology of Sundanese. 9.3 Comparing the representations In Sundanese, nasalization spreads from nasals onto following vowels. The nasalization is blocked by supralaryngeal consonants and underlying glides. Thus, the set of segments


No statistics were run on these data because they were only taken from 5 tokens each.

186 in Sundanese fall into the two classes in (290). Ideally, the analysis of Sundanese nasal harmony will go beyond simply representing these two classes, but will also explain why these two classes exist. (290) Block nasalization consonants underlying glides Allow nasalization vowels derived glides laryngeals (h and )

We will see that none of the representations provides a complete and satisfying account of nasal harmony in Sundanese. In §9.3.5, I will summarize an alternative approach by Walker 1998 and 2003 that rests on a universal nasalizability scale. 9.3.1 LEX In LEX, underlying glides are represented by their inability to surface as nuclear elements. Though there is a structural difference between underlying vowels and glides, there is no featural difference. With no featural difference between underlying glides versus derived glides and underlying vowels, their differing behavior is hard to understand. While surface segments derived from an underlying vowel, as in (291a&c), allow nasalization to spread through them, a surface segment derived from an underlying glide does not (see (291b)). If the features of these segments were the same, it would be difficult to explain why they behave differently. (291) a. Transitional/derived glide (underlying vowel) nijar 'seek (active)' b. Underlying glide iwat 'elope' c. Underlying vowel mihak 'take sides (active)' In order to account for the blocking of the underlying glide in LEX, it would be necessary to appeal to the syllable. For example, an analysis could assume that syllable onsets block the spread of nasalization. As stated earlier, such an analysis runs into problems because [h] and [], which are presumably in the onset, do not block the nasalization. 9.3.2 [CONS] Whereas LEX could not explain the nasal harmony of Sundanese, [CONS] is able to characterize the nasalization rather easily. In [CONS], underlying glides are characterized as [+consonantal], along with other consonants. Vowels and laryngeal consonants ([h] and []), on the other hand, are characterized as [-consonantal]. Starting

187 with Jakobson, Fant, & Halle 1952 and Chomsky & Halle 1968, there has been a tradition of treating laryngeals as [-consonantal]. Given these representations, a rule like (292) accurately represents the harmony. Any segment that is [-consonantal] becomes nasalized after a nasal segment. (292) [-cons] [+nasal] / [+nasal] ____ The rule in (292) must apply iteratively. If it only applied once, then the incorrect output [mihak] would result. This representation also relies on laryngeals being [-consonantal] and therefore targets of the rule. In this rule, the target and the trigger must be strictly adjacent. In a form like [mihak], the final low vowel is nasalized because of the preceding nasal h not because of the initial nasal. 9.3.3 Vowel-Place Theory In Vowel-Place Theory (VPT), the difference between an underlying and derived glide is the location of the place features. An underlying glide has place features under the Cplace node, while a derived glide (and underlying vowel) has place features under the Vplace node. Thus, in VPT, the difference between the two classes in (290) is easily represented, since those segments which can block nasal harmony are those segments with a consonant place feature, while those that allow nasalization to spread across or through them do not have this consonant place. This is the analysis taken in Hume 1995. A representation of the blocking segments is given in (293). Vowels are not blockers because they lack place features under C-place. Similarly, the laryngeals lack place features under C-place, having the representation in (294). (293) Blocking segments Root | Oral Cavity | C-place ([lab]) ([cor]) ([dorl]) (294) Representation of laryngeals h ... ... Laryngeal Laryngeal | | [spread] [constricted]

188 While VPT accurately groups the two classes by looking at the feature geometry, it lacks a motivation for the harmony itself. Why is that consonants and underlying glides should block the spread of nasalization? Cohn 1990, who has a non-feature geometric answer, analyzes the blocking segments as being [-nasal] while vowels and laryngeals are not specified for nasal.124 This explains the blocking effects since [-nasal] segments are not allowed to surface as [+nasal]. In VPT, however, no such motivation can be given. Generally, in cases of spreading, it is segments that are specified for that feature or node which act as blockers. In the analysis presented in this section, however, no such segments exist. It would be necessary to represent all of the blocking segments as [-nasal] in order to account for the blocking, otherwise a situation like that in (295) is predicted. (295) * m Root | [nasal] a Root | ... | V-place r Root | C-place i o s 'examine (active)' Root | ... | V-place

9.3.4 Revised Articulator Theory In Revised Articulator Theory (RAT) it is even more difficult to unify either the class of blockers or the class of non-blockers. One possible way to differentiate the two classes is to target the presence or absence of the feature [nasal]. Since nasal or nasalized counterparts of stops, liquids, glides, and vowels exist in the world's languages, these segments would be specified as [-nasal] in the underlying representation since RAT requires full specification. The only way for a segment to be unspecified for a feature is for it to be permanently unspecified, meaning that no segment exists with the opposite value. A possible spreading rule is provided in (296). (296) a. View all features. b. [+cons] | Soft Palate | [+nasal] [-cons] Soft Palate = [-nasal] Place | Body | [dorsal]

Actually, Cohn says that [-continuant] segments are specified for [-nasal], but that [+continuant] consonants are not specified for [nasal]. There is also a filter which bans [+nasal, +continuant, +consonantal] segments, which explains why the lateral and the underlying glides act as blockers. In her analysis, glides are [+consonantal].


189 Any [+consonantal] segment which is [-nasal] will block this rule. Similarly, any [-consonantal] glide will block since it is not a target because it does not have a designated articulator [dorsal]. The problem with this representation rests with the laryngeals. One possible representation of these segments relies on their being "permanently underspecified" for the feature [nasal].125 From a phonetic point of view, it is plausible that these segments are not specified for [nasal]. In a glottal stop, the closure is behind the velum, thus no air can pass through the nasal cavity. Whether the velum is raised or lowered does not make a difference in the acoustic output of this segment. According to Ladefoged & Maddieson 1996, there is little distinction between plain and nasalized h from an auditory standpoint. This is essentially the stance taken by Cohn 1990, who assumes the absence of nasal features on laryngeals. If the laryngeals were universally underspecified for [nasal], then they can be unspecified in RAT without violating the notion of full specification. Such a view is espoused by Durand (1987: 102) who writes, "I therefore want to suggest that, from an articulatory point of view, [] and [h] differ from the other obstruents in that the behaviour of the velum is not crucial to the production of these sounds." The second analysis is to assume that laryngeals are not underspecified for [nasal]. This analysis is bolstered by the existence of languages which have phonologically nasalized laryngeals that contrast with plain laryngeals (Walker & Pullum 1999). These authors argue that any phonetically possible segment can also be phonologically possible. It is phonetically possible to lower the velum during the articulation of the laryngeals. Therefore, this can be made available to the phonology. They argue that there is an underlying contrast between /h/ and /h/ in several languages (e.g. Kwangali, Arabela, Aguaruna, and Seimat). Their example of glottal stop does not show a contrast between // and / /, but shows that the glottal stop does not block the spread of nasalization. That laryngeals can be specified as [+nasal] in the phonology of some languages means that their nasality must be specified in all languages under a full specification analysis. Since RAT relies on full specification, this poses a serious problem to the analysis of spreading in (296). If the laryngeals are [-nasal] underlyingly in Sundanese, then they are expected to block the spread of [+nasal]. 9.3.5 An alternative approach: Nasalizability Scale None of the four models offered a satisfying analysis of nasal harmony in Sundanese. LEX was unable to differentiate derived from underlying glides, predicting that they should pattern together in the spread of nasalization. [CONS], on the other hand, was actually able to account for the nasal harmony, but required an iterative rule. VPT


The term permanent underspecification is used by Steriade 1995, citing Archangeli 1988. This refers to features which never form possible contrasts.

190 managed a unifying account of the class of blockers and non-blockers, but did not add any insight into why the blockers should actually block the spread of nasalization. It simply characterized them as a natural class. RAT would have been able to represent the harmony were it not for the existence of a few languages that contrast plain and nasalized laryngeals. Since RAT is based on a full representation, the laryngeals too must be [-nasal], which causes problems for a spreading rule. Walker 1998 and 2003 offer an alternative analysis of nasal harmony that is not based on the features of specific segments, but is based on a hierarchy of nasalizability.126 This hierarchy expresses the compatibility of segments with nasalization. The nasalizability scale is given in (297). (297) Vowels > Semivowels > Liquids > Fricatives > Obstruent Stops Vowels are the most compatible with nasalization and obstruent stops are the least compatible. Were nasal stops included in this scale, they would occur to the left of vowels. Walker points out that the scale in (297) resembles the sonority hierarchy, with the exception that nasal stops would not be found between liquids and fricatives, but to the left of vowels, as they are the most compatible with nasalization. Walker's analysis of the hierarchy explains why there is always one clear division in the nasal harmony languages: segments which are nasalizable and those which are resistant to becoming nasalized. If a language allows liquids to become nasalized, then it will also nasalize all segments to their left on the hierarchy and none to the right. Walker's analysis is couched within Optimality Theory where she argues the harmony constraint SPREAD occurs between two points in the hierarchy. Using the constraints *NASALSEG universally ranked according to the nasalizability scale, and the harmony constraint SPREAD[+nas], Walker derives a factorial typology. Some examples are given in (298). The example in (298a) exhibits the pattern found in Sundanese, while that in (298b) shows the pattern in Ijo (Walker 1998). (298) a. *NASSTOP » *NASFRIC » *NASLIQ » *NASGLIDE » SPREAD[+nas] » *NASVOWEL b. *NASSTOP » *NASFRIC » SPREAD[+nas] » *NASLIQ » *NASGLIDE » *NASVOWEL The position of the laryngeals in the scale requires some discussion. In most languages in Walker's 1998 study, laryngeals act as highly nasalizable segments. In fact, they generally pattern as the most nasalizable after vowels. She notes that there are a few cases where glottals pattern with obstruents. This claim is made in order to account for a few cases in which vowels and glides are nasalizable, but liquids, obstruents, and glottals

Walker's scale is based on part on Schourup 1972 and Cohn 1993, though Walker's survey includes many more languages.


191 are not.127 I believe that an alternative to Walker's claim rests not on the variable ranking of glottals, but instead on the variable ranking or nature of glides. Walker's basic scale is provided again in (299) with numbers indicating the possible locations of the division between nasalizable and non-nasalizable segments. (299) Vowels Glottals Glides Liquids Fricatives Stops

Let us examine the set of languages in Walker's study more closely. In her study, there is one language that allows spreading only across vowels (location ) and 8 which allow spreading across vowels and glottals, but not glides (one of which is Sundanese).128 This marks a division at location . The next set of languages allows nasalization to spread through glides but is blocked by liquids and obstruents. Of the 28 languages and dialects that Walker lists as having a division at point , 15 have blocking by both glottals and glides and 11 either do not have glottals or the data is unclear. These 26 cases cannot actually determine the relative order of glottals and glides in the hierarchy since either order would produce the same results. The remaining two languages treat glottals and glides differently than is expected by the hierarchy in (299). In Rejang, glottals block nasalization, while glides do not. In Madurese, the same holds, but there is evidence that the glides are derived from vowels. In fact, of these 28 languages, 26 do not provide any evidence of the relative order of glides and glottals in the hierarchy. Rejang and Madurese appear to show the reverse order. Though I do not have a full analysis worked out at this point, I would like to suggest the revised hierarchy in (300). (300) Revised Hierarchy Vowels, Derived Glides > Laryngeals > Underlying Glides > Liquids > Frics > Stops To reinforce this new hierarchy, I will provide a brief excursus on Madurese nasalization. Madurese inserts glides between two vowels (Stevens 1968: 29). These derived glides are subject to nasalization, as shown in (301). (301) /ma/ /neat/ /N+sm/ [m w a] [n j a t] [ j m] 'fact' (Cohn 1993: 357) 'intention' 'kiss' (active)

Cohn 1993 also noticed the special status of glottals and argued that the nasalizability scale has two tracks, one with supralaryngeal consonants (in which she includes glides) and a second with the laryngeals. 128 Four of these eight are different dialects of one language.


192 There are some glides which are not derived in this way, but no data exists with them in the nasalization context (e.g. [apj] 'fire') (Cohn 1993: 357). Cohn argues that the glides in Madurese are more vocalic than those in Sundanese because they do not show weakening of the formant structure. It is not clear to me if there must be a phonetic difference between the realization of underlying versus derived glides. It may be more likely that the difference in the realization of glides in Sundanese and Madurese is simply a language specific difference. I believe that Cohn 1993 is on the right track in pointing out that the glides in Madurese are derived whereas those in Sundanese are underlying. I believe that this is the basis for their difference in behavior in the nasalization hierarchy. Further evidence that the variable behavior of glottals and glides is not the result of an alternative ranking of glottals with obstruents comes from their order with respect to the other segments on the scale. Of the 16 languages with a division at point in (299) where liquids are also targets of harmony, zero show the glottals as patterning with the obstruents and being less compatible with nasalization. Likewise, of the 4 languages which allow nasalization of fricatives (location ), none have blocking by glottals. If glottals truly had variable ranking, then we might expect some cases in which liquids or fricatives were nasalizable, but glottals were not. In fact, this does not happen. The only place where glottals appear to have a variable ranking is which respect to glides. I believe this is further evidence that there are two types of glides and that they are located in two different places on the nasalizability scale. Thus, the different behavior of the two types of glides in Sundanese can be explained using the nasalizability scale. The language simply makes its division between the laryngeals and the underlying glides. All segments to the right of and including the underlying glides block harmony, while all those to the left can be nasalized. 9.4 Discussion Sundanese nasal harmony was not elegantly explained by any of the four models that targeted the feature differences between segments. Instead, a nasalization hierarchy that mimics the sonority hierarchy clearly identifies a division between those segments that can become nasalized and those that cannot. Perhaps Sundanese should have been relegated to Chapter 10 which discusses the languages which have underlying glides but which cannot distinguish among the four representational theories. 9.5 Appendix: Excursus on [continuant] and nasalization One final aspect of nasal harmony that relates to the features of Sundanese is the transition from nasal to non-nasal segments. In terms of the actual phonetic implementation, Cohn 1990 shows that there is a gradual transition during [w, j, l] from the preceding nasalized vowel and the following non-nasalized vowel. This is illustrated with the mark to illustrate a gradual decline in nasalization.

193 (302) a. /iwat/ [iwat] 'elope' b. /ajak/ [ajak] 'sift (active)' c. /uliat/ [uliat] 'stretch (active)' Contrastingly, [r] and the obstruents do not show this gradual decline of nasalization. Cohn 1990 attributes this to the fact that these segments are [-continuant].129 In her 1990 analysis, the [-continuant] consonants are underlyingly [-nasal] and therefore block the spread of nasalization. The [+continuant] consonants, on the other hand, block the spread of nasalization because of a constraint that blocks [+nasal] on [+continuant] consonants. The problem with this account is that it implies that [s] should behave like [j, w, l] and show a gradual decline of nasalization, but instead it patterns with the obstruents and shows a quick decline in nasal airflow (Cohn 1993). In order to account for the behavior of [s], Cohn 1993 assumes that the voiceless fricative must be oral due to a third property. She argues that a lowered velum during the production of a fricative would prohibit enough oral airflow to produce frication. She thus concludes that [s] is oral (i.e. [-nasal]) due to an articulatory constraint. Ladefoged & Maddieson 1996 point out that there is little auditory difference between oral and nasal voiceless fricatives. They imply, however, that the assimilation of voiceless fricatives to nasalized segments is more common than generally reported. And Walker 2003 shows that fricatives can be nasalized (e.g. in Applecross Gaelic and Tuyuca). In these two languages, nasal harmony spreads through the fricatives. The existence of nasal fricatives implies that Cohn's extra constraint to ban nasal fricatives is not entirely motivated. The different behavior of [w, j, l] on the one hand, and [r] and the obstruents on the other must be due to something else.

The implication is then that [w, j, l] are [+continuant], an assumption that is challenged by the lateral in Pulaar and Tahltan and also Kaisse 2002, though see Mielke 2004 for further discussion of the variable nature of laterals.



10 Other languages with underlying glides

10.1 Overview This chapter will look at three additional languages with underlying glides: Cree, Pashto, and Imdlawn Tashlhiyt Berber. These languages contrast with those in the preceding chapters in not having processes that featurally distinguish vowels from glides. Instead, the evidence here comes only from syllabification and related syllable-based processes. Therefore, these languages will not help in distinguishing among the four representations. Any of the four models can account for syllable-based differences between vowels and glides. Still, these languages are important in providing more sound cases of languages with underlying glides. 10.2 Moose Cree 10.2.1 Introduction Cree is an Algonquian language of Canada. The particular dialect discussed here is Moose Cree, as described by Ellis 1983. In the remainder of this section, I will simply use the name Cree to refer to this dialect. The evidence for underlying glides in Cree comes from the presence of reverse sonority clusters. Because the only evidence for underlying glides in Cree is distributional, it will not help distinguish among the four representations. Nonetheless, it is another example of a language with underlying glides. Cree represents a Type II language where there is no corresponding underlying /u/, though there is /w/.130 The status of the palatal glide is unclear. It does not occur in any reverse sonority clusters. (303) /w/ | [w] The consonant inventory of Cree is given in (304). Pre-aspiration of stops and affricates, though "light", is contrastive post-vocalically (see (305)). Voicing is not distinctive. Stops are voiced medially when preceded by a long stressed vowel or by a nasal ([madi] 'begin to' vs. [mati] 'bad'). [w] is less rounded than in English, with the lips held straighter (Ellis 1983: 19).


If one were to analyze the back round vowel as /u/, then Cree would represent a Type V language with /u/ and /w/ but no alternations.

195 (304) Consonant inventory of Cree Labial Alveolar Stop p t Pre-aspirated p t Fricative s Nasal m n Liquid l Glide w (305) petaw akotin 'he brings it' 'he hangs' Post-alveolar t t j petaw akotin 'he waits for it' 'he floats' Velar k k h

The vowel phonemes are provided in (306). The back round vowel varies in height from mid to high. The low vowel varies from central to front [æ]. (306) Vowel phonemes of Cree i i e o o a a 10.2.2 Basic phonology and morphology of Cree Cree has two numbers (singular and plural) and two genders (animate and inanimate). The inanimate class contains only inanimate nouns. The animate class contains all animate nouns, as well as some inanimate nouns. For example, [astis] 'mitten' is part of the animate noun class. Cree has the three persons 1st, 2nd, and 3rd, but also the obviative or farther 3rd person which Ellis 1983 marks as 3'. The nominal morphology that will be relevant in the rest of this section is given in (307). The verbal morphology is given in (308). (307) Nominal morphology Obviative Plural Stative (it is...) Locative Animate -a -ak -iwiw -ik Inanimate -iliw -a -iwan -ik

196 (308) Verbal morphology a. Animate Intransitive (AI) 1p: ni--n pl: ni--n-an 2p: ki--n 3p: -w pl: -w-ak b. Inanimate Intransitive (II) 3p: -w pl: -w-a 3'p: -w pl: -w-a c. Direct marker (Transitive Animate) -ad. Inverse marker (Transitive Animate) -ikwHiatus is disallowed in Cree. When the person prefix is attached to a vowel initial verb, an epenthetic [t] surfaces to break the hiatus. For example, /ni-otamahw-ikw/ surfaces as [nitotamahok] 'he hits me' (E231). Similarly, /o-askikw-ik/ surfaces as [otaskikok] 'in her kettle' (E590). This form also illustrates another process in Cree. When the /w-i/ come together at a morpheme boundary, the result is [o]. In word initial position, all consonants can surface with the exception of preaspirated stops and [h]. Word finally, all consonants appear except [t], [l], and [h]. The affricate [t] often surfaces as [t] when the following sound is coronal and when spoken with a faster speaking rate (Ellis 1983: 155). Any of the vowels in (306) can surface word initially. A survey of Ellis 1983 shows that words in Cree may end in [i, e, o, a] (in addition to the consonants). I assume the missing vowels in final position are simply gaps in the source. It is necessary to know which segments are allowed in word-final position to rule out an analysis of Cree that does not have an underlying glide. This will be discussed in 10.2.3. (309) a. Word-initial singleton consonants p 'first' pitama t 'ball' tohwan 'canoe' t timan k 'boil, sore' kasip s sipij 'river' 'duck' iip m maskwa 'bear' n niska 'goose' l 'it is windy' lotin w '3 p pronoun' wila j 'Look out!' jakwa

197 b. Word-final singleton consonants p 'boil, sore' kasip 'hardly' t akawat k 'duck pl.' iipak 'dress' p akop 'some, several' t atit 'now' t anot 'across' k akamik s 'moose (nom.)' mos 'pup' atimoi m 'snowshoe' asam n 'tent' apakwason w 'he barbecues' apwew j apwoj 'paddle' 'ghost' tipaj Word initially, the only clusters that surface involve [w]. In addition to the clusters in (310a) below, [wap] 'shop' surfaces, but only in this loan. Intervocalic clusters are more numerous and have the form Cw, sC, C, and nt. The medial sequence [js] occurs only in the loan [lajsins] 'license'. Word medial sequences of three consonants are all of the form STw where S is a sibilant and T is a stop. These are given in (310c). The only word-final cluster that surfaces is [sk]. The sequence [ns] surfaces in the two loans [lajsins] 'license' and [sens] 'cent'. We will see in the next section that some words can end in the cluster /Cw/, which is reduced to [C] on the surface. (310) a. Word initial clusters: C + w p 'late' pwastaw t 'he goes through the ice' twain k kwajask 'properly' m 'exactly' mweti

198 b. Word medial clusters (2Cs) pw 'thigh' mipwam tw 'say!' itwe kw 'what?(inan.)' kekwan 'tent' kw apakwason sw 'seven' niswas 'wigwam pole' w apawoj mw 'before' (used in Lk Winnepeg area) emwaje nw 'it is long' kinwaw lw 'it is calm' alwatin hw 'ball' tohwan sp aspin 'since' st astis 'mitten' 'he is worn out' st mestipaliw sk 'kettle' askik 'above' p ipimik 'it is calm' t alwatin 'mud' k aikij nt anta 'there' c. Word medial clusters (3Cs) skw apiskwesimon 'pillow' maskwa 'bear' 'door' kw ikwatem spw ospwakan 'tobacco pipe' d. Word final clusters (2Cs) sk amisk 'beaver' The two surface glides pattern differently in terms of clusters and general distribution. While [w] appears in numerous clusters, [j] does not. The status of the palatal glide remains unclear. The remainder of this section will deal with the labial glide, since there is evidence for its underlying status.

199 10.2.3 Evidence for underlying glides: reverse sonority clusters The evidence for underlying glides in Cree comes from reverse sonority clusters involving glides. In particular, all of the evidence comes from final sequences of /...Cw/ where the reverse sonority cluster is banned from surfacing.131 There are several repairs that are available cross-linguistically to remedy this banned cluster; for example, a language could choose vowel epenthesis, glide vocalization, or deletion of one of the segments. In Cree, the last option is selected and these underlying sequences surface without the final glide. Evidence will come from nominal and verbal roots that end in a sequence of /...Cw/ and the third person (3p) and obviative person (3'p) verbal suffix /-w/. The first set of data comes from nominal stems that end in an underlying sequence /...Cw/. When these forms are uninflected, the final glide cannot surface in this cluster. The result is a stem that ends in [...C]. When vowel-initial suffixes are added to these stems, the underlying glide is allowed to surface. In (311a), the stems end in an underlying labial glide /w/. The glide only surfaces when the animate plural or obviative markers are added. These forms contrast with (311b) where no labial glide surfaces in the plural or obviative. The fact that these two sets of words contrast rules out the possibility that the plural and obviative are actually /-wak/ and /-wa/. (311) Uninflected a. askik

Pl. (anim.) /-ak/ askikw-ak atimw-ak atikamekw-ak Obviative /-a/ UR /askihkw/ /atimw/ /atihkamekw/ /mosw/ /napakahtikw/ /napakitapanaskw/ /ajkikw/ /apowoj/ /awai/ /astis/ /aokan/ /iip/ Gloss 'kettle' 'dog' 'white fish' 'moose' 'plank' 'toboggan' 'seal' 'paddle' 'child' 'mitten' 'warf' 'duck'

atim atimw-a atikamek mos mosw-a napakatik napakatikw-ak napakitapanask napakitapanaskw-a Akik akikw-ak


apowoj awai astis Aokan iip

awai-ak astis-ak aokan-ak iip-ak

apwoj-a awai-a aokan-a

When suffixes beginning with /i/ are attached to /...Cw/ stems, the existence of the underlying glide is revealed by its effect on the following vowel. Recall, across a morpheme boundary, /w-i/ surfaces as [o]. As in (311), the uninflected stems in (312a) surface without the final /w/. When /i/-initial suffixes are added, the result is [o],

This is the opposite of what occurs in Pashto where such reverse sonority clusters are allowed to surface. See § 10.3.3.


200 showing the existence of /w/ at the end of the stems. These forms contrast with those in (312b) where a /w/ (or its resulting [o]) never surface. (312) a.

Uninflected askik kiik piwapisk ministik mistik mos Locative /-ik/ o-t-askikok kiikok ministikok mistikok Possessed /-im/ UR /askihkw/ /kiikw/ /piwapiskw/ /ministikw/ /mistikw/ /mosw/ /nohtim/ /timan/ Gloss 'kettle' 'sky' 'motor' 'island' 'tree/stick' 'moose' 'bush' 'canoe'

ni-piwapiskom ni-mistikom ni-mosom notimik timanik


notim timan

The reason the glide does not surface in the unsuffixed forms is simple; Cree chooses deletion over the other possible repair strategies. In other words, Cree does not allow glide vocalization or surface reverse sonority clusters. Though this type of cluster can exist in some languages, they are rare. In Cree, the reverse sonority cluster *[atimw] is banned and the repair strategy is simply to delete the glide. Similar to the nouns that end in /...Cw/, there are verbs that underlyingly end in a consonant-glide sequence. When stems that end in a Cw cluster are in the imperative (null suffix), the cluster is simplified. For verbs that end in /...sw/ or /...w/, the glide is deleted. For verbs that end in /...hw/, the /h/ is deleted (Ellis 1983: 379). This difference may be due to the absence of [h]-final words in Cree (see (309b)). When the indicative /-ew/ is attached, the glide surfaces. When the inverse suffix /-ikw/ is attached, the result is [o].132 (313) Evidence for final Cw Inverse /-ikw/ 'he Xs him' /-ew/ a. [otamohw-ew] /ni-otamahw-ikw/ 1p. ­ hits - inverse [nitotamahok] 'he hits him' 'he hits me'

Imperative /-Ø/

The inverse suffix causes a reversal of the transitive verbs. Examples without this suffix would mean 'we saw him' or 'you hit her'. The inverse suffix causes a reversal of the theta roles, yielding 'he saw us' or 'she hit you'.


201 b. [paskisw-ew] /ki-paskisw-ikw-ina-w-ak/ 2p-shoot-inverse-incl. 1st p.-3p-pl. [kipaskisokonawak] 'they are shooting at us (inclusive)' [paskis_]

'he shoots him' c. [tistahw-ew] 'he punctures him'

'shoot him' [tista_w] 'puncture him'

The last type of evidence for underlying glides and the ban on reverse sonority clusters comes from the verbal morphology. Intransitive animate verbs either end in a vowel or in /n/.133 The third person singular suffix is /-w/. When this suffix attaches to an n-final stem, it does not surface. When it attaches to a vowel-final stem it does surface. It also surfaces on n-final stems when the plural suffix /-ak/ is also attached. (314) a. api-w nipa-w itote-w tatamo-w b. takoin-w [takoin] pimiin-w [pimiin] pakiin-w [pakiin] c. takoin-w-ak 'he sits' 'he sleeps' 'he goes (to a place)' 'he sneezes' 'he sits' 'he lies down' 'he falls' 'they sit'

Similarly, inanimate intransitive verbs take this /-w/ (third person singular) suffix. Again, when the stem ends in /n/, the glide suffix does not surface (Ellis 1983: 75). (315) a. tipiska-w milopali-w b. itakwan kimiwan otakoin 'it is night' 'it is going well' 'it is (at a place)' 'it is raining' 'it is evening'

Ellis refers to verbs in Cree as either Animate Intransitive, Animate Transitive, Inanimate Intransitive, or Inanimate Transitive.


202 Let us consider the alternative hypothesis that there is no underlying labial glide in Cree, and that the verbal suffix and the final segment on the nominal and verbal stems is an underlying vowel. It is certainly possible that this segment, call it /V*/, surfaces as a glide when adjacent to another vowel. However, the problem for this hypothesis arises when V* is not adjacent to another vowel and where the optimal parse would be for the vowel to surface as a vowel. What would prevent this vowel from surfacing in final position, as for example [atimV] or [otakoinV]? A survey of the glossary in Ellis 1983 shows that words can end in a variety of vowels.134 There is therefore no reason why the V*-final words could not end in a vowel. The evidence presented above shows that Cree has an underlying labial glide. All of the evidence comes from underlying sequences that would produce reverse sonority clusters. Cree does not permit such clusters to surface, opting instead to delete the glide. Because evidence from reverse sonority clusters is based on syllables and syllabification, and is not dependent on featural specifications of these segments, it is not possible to use such evidence to distinguish among the four representations. 10.3 Pashto 10.3.1 Introduction Pashto is an Indo-Iranian language spoken in Afghanistan and Pakistan. Most of the data in this section will be drawn from the Kandahar dialect (Western) (Penzl 1955), though the Kabul dialect (Eastern) will be used as well (Bell & Saka 1983). Pashto is relevant for our purposes because of its abundant word initial clusters. Not only does Pashto contain a wide variety of normal clusters, but also many reverse sonority clusters. We will see that the labial glide and vowel in Pashto exhibit a Type VI pattern, where underlying vowels surface as vowels and glides, and where the underlying glide surfaces only as a glide. The palatal glide, on the other hand, exhibits a Type III pattern where an underlying vowel has two surface realizations. (316) Type VI /u/ /w/ [u] [w] [i] Type III /i/ [j]

The consonant inventory is listed in (317). In addition to these consonants, Pashto also contains /q/, //, //, and /f/ in elegant, formal speech. There is also a palatal glide [j] on the surface, but the evidence does not indicate that it is underlying (see §10.3.3).


The survey shows that words end in [i, e, o, a].

203 (317) Consonant phoneme inventory (Kandahar) Labial Dental/Alveolar Retroflex Velar Pharyngeal Stops p b t d k g Affricates ts dz t d Fricatives h s z x Nasals m n Liquids r l Glides w The specific place of the segments is described differently in different sources. The inventory in (317) resembles most closely the description found in Penzl 1955. He describes /s z/ as dental and / / as post-dental or alveolar, though Bell & Saka consider all of them to be alveolar. Penzl 1955 describes / / is a retroflexed lateral flap. Penzl does not regard the affricates as single segments, though Elfenbein 1997 argues that they are. The affricates behave as single segments by forming clusters with two other consonants. If the affricates were actually two segments, they would form the only clusters of four consonant in the language. Furthermore, the affricates behave as single units in metathesis (Elfenbein 1997). The vowel inventory is given in (318). According to Penzl 1955, the two low vowels differ only in quantity. This differs from the dialect described in Bell & Saka 1983, where they claim that the two low vowels primarily differ in quality. Both works assert that the mid front and mid back vowels are phonetically long. In elegant, formal speech, [i] and [u] also occur. (318) Vowel phonemes Penzl 1955 (Kandahar) i u e o a a Bell & Saka 1983 (Kabul) i u e o a

10.3.2 Two aspects of Pashto phonology One salient aspect of Pashto phonology is the presence of many consonant clusters. Not only does Pashto contain clusters such as [tsmx] 'lining' (Standard Pashto) (Elfenbein 1997: 741), but also several reverse sonority clusters. The possible word-initial two consonant clusters are provided in (319). The initial consonant of the cluster is along the left side and the second consonant along the top row. The shaded area of the chart highlights the reverse sonority clusters.

204 (319) Word-initial two-consonant clusters (Penzl 1955: 16) w r l m n s x z p w r l m n s x z p t k b d g t k b d


Crucially, Pashto contains several reverse sonority clusters in which the initial segment is the glide [w]. These will be discussed in greater detail in the next section. In addition to the clusters in (319) are those that include the affricates: [mdz, nd, tsw, tsr, tsk, tsx, t, dzw, dz]. A second important point is that the bilabial glide is in fact a glide and not a fricative. Shafeev 1964 describes this segment as a bilabial fricative []. However, according to Penzl 1955 and Bell & Saka 1983, it is produced with no frication. In terms of the phonology, the glide patterns with other sonorants, not obstruents. In Pashto, only sonorants can combine with either voiced or voiceless obstruents. Obstruent clusters, on the other hand, must agree in voicing. The glide occurs with both voiced and voiceless obstruents (Penzl 1955: 16), showing that it is indeed a sonorant. 10.3.3 Evidence for underlying glides: reverse sonority clusters As with Cree (§10.2), the evidence for an underlying glide in Pashto comes from reverse sonority clusters. Whereas Cree bans such clusters and deletes the glide when it cannot be syllabified, Pashto allows these clusters to surface.

205 The fact that the labial glide occurs in reverse sonority clusters at all provides some evidence for the notion that it is underlying. If this segment were underlyingly a vowel, basic syllabification would maintain it as such, and no reverse sonority clusters would exist. Furthermore, in the Kandahar dialect, minimal pairs such as [wradz] 'day' and [uriadz] 'cloud' exist (Penzl 1955: 24).135,136 In an acoustic study, Bell & Saka 1983 examine two different reverse sonority clusters ([wr-] and [wl-]) in an eastern dialect from Kabul. Their phonetic study shows that the output segment in these clusters is phonetically a glide. Their study also implies that this glide is not simply a reflex of a vowel but is a glide underlyingly. They compare initial glides in reverse sonority clusters (e.g. [wr-]) with those in prevocalic position (e.g. [wV-]), and also with the short unstressed vowel [u] in [bura] 'sugar'. They find that glides in reverse sonority clusters have a similar spectral shape to those in singleton onsets. They also find that the short unstressed [u] in [bura] 'sugar' is significantly longer than the glide in [wr] 'wedding procession' (p<.001), illustrating that the glide is not simply a short syllabic vowel. An additional piece of evidence comes from stretching, a method used to increase the overall duration of an utterance. Here, the speaker is asked to make an utterance last two seconds or five seconds. When [wr] was stretched, the duration of the glide increased, but it did not turn into its cognate vowel [u]. They conclude that the initial segment in these reverse sonority clusters is a glide. Let us consider an alternative underlying representation where the initial segment in these reverse sonority clusters is a vowel. The existence of the near minimal pair [wradz] 'day' and [uriadz] 'cloud' in and of itself does not constitute evidence that one is an underlying glide and the other a vowel. An alternative analysis could argue that the latter be represented with a prelinked nuclear vowel, with the former having a normal vowel. This would then be similar to Spanish or Korean (see Chapter 11). Such nuclear prelinking, however, should only be used when a vowel surfaces in a place where we would expect the glide to surface. This is not the case in Pashto. The vowel is expected to surface when a more sonorous segment is not available. Instead, in Pashto, we find the glide in a place where we expect a vowel, not a vowel where we expect a glide. Pashto does have an underlying vowel that can surface as a glide in certain contexts. In (320), the morpheme final vowel /u/ surfaces as a vowel when no suffix is added. However, when the vowel-initial plural suffix is added, it surfaces as a glide, avoiding hiatus. We see that Pashto is indeed a Type VI language, where the underlying

In the Eastern dialects, like the one described in Bell & Saka 1983, 'cloud' surfaces as [wriadz] (Penzl 1955, Bell & Saka 1983). 136 The data discussed in this section all deal with onset clusters. It is unclear whether reverse sonority coda clusters exist. In Penzl 1955, the word dalw '11th Afghan month' is presented in the section on the labial glide, but in the glossary, it is listed as dalw(a). It is unclear to me whether it can be pronounced without the final [a].


206 vowel surfaces as either a vowel or a glide, and where the underlying glide surfaces only as a glide. (320) hindu kandu 'hindu' 'wheat bin' hindwan kandwan 'hindus' 'wheat bins'

Whereas there is evidence for the phonemic status of /w/, no such evidence exists for the palatal glide.137 First, the distribution of the palatal glide is always predictable. It either surfaces word initially before a vowel, intervocalically, or word finally after another vowel. The palatal glide also appears medially before another consonant after [a] and [a] in [pajse] 'money', [ajnaki] 'spectacles', and [pajda] 'advantage'. Page numbers are from Penzl 1955. (321) a. jawa je jad ji b. wajlj aja majn c. kj zuj swaj 'one fem.' '(you) are' 'memory' '(he) be' (P18) (P21) (P24) (P24)

'said' (perfect participle) (P18) 'interrogative particle' (P21) 'lover' (P24) 'point' (P18) 'son' (P20) 'could' (optative) (P24)

Furthermore, the palatal glide never forms part of an onset cluster, either regular or reversed sonority.138 In Pashto, [i] can surface in hiatus when the glide parse would produce a cluster. In fast speech, [Cj] clusters are formed (Penzl 1955: 24).

This view differs from that of Penzl 1955 who believes that both of the glides are phonemic. A search of the 1386-word glossary from Penzl 1955 shows that the only word medial glides that are not intervocalic are in these six words [harjaw] 'everybody', [kumjaw] 'which', [beljaw] 'somebody else', [terjak] 'opium', [tarjak] 'weight measure', and [turjalaj] 'brave'. In all of these words, it is possible to analyze these words with a syllable boundary between the consonant and the glide. In the first three, there may also be a morpheme boundary before [jaw]. No word initial clusters with this glide exist.



207 (322) bia diarls piaz 'again' (P24) 'thirteen' (P24) 'onion' (P24)

Where the glide does appear to contrast with the vowel [i], the morphology dictates this distinction. (323) spaj dj 'dog' '(he) is' spi li 'female dog' 'girl'

The sequence [i] is made with slightly more stress on the second vowel and contrasts with [j] (Penzl 1955: 23). It is sometimes called the 'strong feminine y' (Penzl 1955: 23). The fact that this high vocoid does not glide in these forms can be explained. The morphological structure of the word might prevent gliding in this morpheme. Furthermore, the two diphthongs do not minimally contrast for i/j because they also differ in stress. That is, [i] bears stress on the [i], while [j] has stress on []. The distribution of the palatal glide is regular and can be based on principles of syllabification. The distribution of the labial glide, on the other hand, cannot be explained purely by principles of syllabification since the glide appears in syllable positions where it is not expected to occur. However, the only distinction between the labial glide /w/ and the labial vowel /u/ has to do with syllabification and not with featural differences. Therefore, as with Cree, it is not possible to use Pashto to evaluate the validity of the four representations. 10.4 Imdlawn Tashlhiyt Berber 10.4.1 Introduction Imdlawn Tashlhiyt Berber is the dialect of Berber spoken in the Western Higher Atlas of Morocco (henceforth ITB) (Dell & Elmedlaoui 1985). This particular dialect is the subject of several articles by Mohamed Elmedlaoui, a native speaker of this dialect, and François Dell (Dell & Elmedlaoui 1985, 1988, 1989, 1991, 1996a, and 1996b, Elmedlaoui 1985 and 1988). In the remainder of this chapter, these references will be presented as ME85 and D&E88, for example. One of the hallmarks of Berber is that any segment can form a syllable nucleus. That is, all segments show allophonic variation between nuclear and non-nuclear realizations. I will show that ITB has underlying glides that differ from basic underlying high vowels. It is in exactly such a language that one would expect the underlying glides to show glide/vowel alternations, given that all segments are allowed to occupy the syllable nucleus. ITB represents the only example of a language I found in which underlying glides show alternations. In fact, ITB exhibits alternations of both underlying vowels and glides, a Type VIII language, as shown in (324).

208 (324) Type VIII /i/ /j/ [i] [j] The phoneme inventories are given in (325) and (326). (325) Consonant Inventory (Elmedlaoui 1985) Labial Alveolar Alveo- Palatal palatal Stops b t t d d Fricatives f s s z z Nasals m n n Liquids r r l l Glides w j (326) Vowel phonemes of ITB i u a Before we continue, several of the symbols in (325) require some comment. Here <h> represents [], a "murmured glottal fricative ('voiced h')" (Dell & Elmedlaoui 1996b: 219), <r> is a voiced alveolar flap or trill (D&E 1996b), C represents 'emphasis' or dorsopharyngealization, and <x> represents a voiceless uvular fricative []. In the data that follows, I have elected to remain true to the source when it comes to the representation of emphasis. The various articles differ in whether the underlying form is listed as /...C.../ or as /!...C.../, where <!> indicates that at least one segment in the word has secondary dorsopharyngealization. In my transcriptions, both of these underlying representations should be understood has having a consonant with secondary dorsopharyngealization. Dell & Elmedlaoui state explicitly that in the surface form, emphasis spreads through the entire word. Furthermore, in accordance with D&E's transcriptions, underlying geminates will be written with two consecutive segments. Thus, /nn/ or [nn] should be interpreted as a geminate n. 10.4.2 Basics of syllabification One of the major advances in the work of Dell & Elmedlaoui is their analysis of syllabification in ITB. Unlike many languages where only vocalic or only sonorant segments can surface as syllabic nuclei, all segments in ITB have a syllabic allophone Velar k kw g gw Uvular q qw x xw w Pharyngeal/ Laryngeal


209 (ME85:26). Thus, the challenge is to predict in a given string of segments which ones actually surface as syllabic nuclei. In this, and all that follows, syllabic segments will be represented in capitals. For IPA characters for which no capital character exists, syllabic segments will be represented with the standard C. The exception to this convention is high vocoids which are represented by [i] and [u] when they are syllabic and by <j> and <w> when they are not syllabic. In the underlying representations, /i, u/ represent the basic high vocoids. I will often use the term "vowel" to refer to these segments. The segments /j, w/ will represent underlying glides. The data in (327) show the allophonic variation of several consonants of ITB. (327) 2 p. sg. [arM-t] [kS-t] [azN-t] 'taste it' 'make it graze' 'send it' 2 p. pl. (ME85:27) [arm-at] 'taste!' [ks-at] 'graze!' [azn-at] 'send!'

Not only do consonants surface with both syllabic and non-syllabic allophones, high vowels too show this alternation. The data in (328), (329), and (330) show three types of glide/vowel alternations. The first shows root allomorphy with the addition of a vowel initial suffix. The second shows allomorphy in affixes where the surface realization of the underlying vowel is a vowel when adjacent to a non-syllabic segment and is a glide when adjacent to a syllabic segment. The third type in (330) shows that glide/vowel alternations are allowed across a word boundary. The alternating segments are written in bold. (328) /zri/ [zri] 'pass! (' [zRj-at] 'pass! (' (ME85:41)

(329) a. Pronominal clitic suffix allomorphy (D&E1989:169) Pronoun Dative At X's place 2 m pl /-un/ a-wn dar-un 2 f pl /-unt/ a-wnt dar-unt b. 3rd person masculine singular allomorphy /i-/ i-fk 'he gives (aorist)' (D&E1989:172) i-kti 'he remembered' (ME85:27) 139 is-i-kti 'did he remember' (ME85:27)

Locative gi-wn gi-wnt


/is-/ is the yes/no question prefix.

210 j-ak j-aggug j-ura ra-j-kti 'he goes (aorist)' 'he is far' 'he wrote' 'he will remember'140 (D&E1989:172) (ME85:27) (ME85:27) (ME85:27)

(330) a. In isolation [zri] 'Pass!' (ME85:41) [usixk] 'I carried you' (ME85:42) b. In a phrase [zriwsixk] 'Pass! I took charge of you.' (ME85:42) The syllabification of ITB is achieved via Core Syllabification (CS). This basic algorithm creates onset-nucleus units (ON). The process of CS creates ON sequences starting with the most sonorous segments and progresses until it reaches the least sonorant segments. The sonority hierarchy in (331) does not include underlying glides. We will see later that they behave as if they have a very low sonority. (331) Sonority Hierarchy of ITB (D&E85): a > high vowel > liquid > nasal > [+vc] fric. > [-vc] fric. > [+vc] stop > [-vc] stop Core syllables are built according to the convention in (332). Type T refers to one of the classes of sounds in the sonority hierarchy of ITB. First, Core syllables are formed where Z is the low vowel. Next, syllables where Z is a high vowel are formed. The process continues until the string is completely syllabified.141 (332) Core Syllabification: Associate a core syllable with any sequence (Y)Z, where Z is a segment of type T (D&E88:4). The algorithm in (332) accounts for most of the syllabifications. However, there are a few additional requirements that are necessary to accurately account for all of the data. First, given a plateau in sonority, the algorithm applies left-to-right. In other words, the nucleus is formed from the leftmost member of the plateau with the segment to its left forming the onset.

/ra-/ is the future prefix. Clements 1997 offers a non-derivational analysis of ITB syllabification that includes many of the aspects discussed in this chapter, including prepausal annexation. Prince & Smolensky 1993 also consider ITB, but do not address all of the data relating to syllabification (e.g. prepausal annexation).



211 (333) /gli-izm/ /sbudu=urti/ /rks-x/ /bain-n/ (334) L-R /gli-izm/ (li) ------(zm) (glij) [glij.zM] [glij.zM] *gL.ji.zM 'drive back the lion' (ME85:35) [sbu.du.wR.ti] *sbud.wur.ti 'dig up the garden' (ME85:35) [R.kSx] *Rk.sX 'I hid' (D&E85:114) [ba.jNn] *baj.nN 'they (m) appear' (D&E85:114) Incorrect R-L /gli-izm/ (ji) (gl) (zm) *[gL.ji.zM] UR CS(HV) CS(Liq) CS(Nas) Coda, Onset augmentation Output

The second stipulation to (332) is that all syllables must have an onset, with the exception of word initial syllables (D&E85:113). The L-R parse in (334) creates the Core Syllable (li). Despite the fact that the second /i/ is also a high vowel, it is not allowed to surface as a nucleus because there is no onset available *[gli.i.zM]. The same is true in cases without a plateau. Forms such as /i-aul-tn/ 'he made them (m.) plentiful', surface as [i.a.wL.tN], and not as *[i.a.ul.tN] (D&E85:113). Though the high vocoid /u/ is more sonorous than the liquid /l/ and therefore a better nucleus, it is banned from a nuclear position in this token because the result would be an onsetless syllable. The ban on onsetless syllables is only active for non-initial syllables. Thus, in word initial position, onsetless syllables are allowed, as in (335). (335) [ikru] [RgL] 'kid' (D&E85:113) 'lock' (D&E85:113)

There is one caveat to the general acceptance of word-initial onsetless syllables and to the left-to-right syllabification. If the two initial segments form a sonority plateau, then the word surfaces with an onset in the first syllable (i.e. satisfies the criterion of having an onset whenever possible). Some examples are given in (336). The expected output using L-R syllabification is provided in (337). With this caveat, the correct outputs can be realized. (336) /sxxn/ /i-ut=i-il/ /i-ullt/ [sX.xN] [jutiil] [jullt] *Sx.xN *iwtiil *iwLlt 'dip (in sauce)' (D&E85:121) 'he hit his arms' (ME85:17) 'he raised it' (ME85: 45)

212 (337) Expected Syllabification /sxxn/ /i-ut=i-il/ -----(i) (ti)(i) (xn) ----------(S) ----------(Sx) (iw) (il) *[Sx.xn] *[]

UR CS(HV) CS(Nas) CS(Fric) Coda Formation Output

Codas are formed from leftover segments. For example, Core Syllabification of /rksx/ (see (333)) forms the two syllables (R) and (kS). The final /x/ cannot form a nucleus since there is no onset available. Instead, it forms a coda with the preceding syllable, yielding (kSx). There is also onset augmentation. Elmedlaoui 1985 discusses a complex system of rhyme 'stability'. In /gli-izm/, core syllabification is expect to produce [G.lij.zM] with an onsetless [G]. Because of the sonority profile (stop followed by liquid), onset augmentation occurs, creating [glij]. A full discussion of rhyme stability will not be given here; the reader is referred to E85. Finally, geminates require another slight modification to Core Syllabification. Tautomorphemic geminates never surface as a CV unit. The two parts of a geminate may surface across a syllable boundary as coda + onset ([j-at.tuj] 'it is high' (D&E85:105)), with both halves in the coda ([t-xZ.n=akkw] 'she even stockpiled' (D&E85:113)), or across a syllable boundary as nucleus + onset ([] 'sew (imperfect)' (D&E85:121)). However, two identical segments that are separated by a morpheme boundary can be syllabified as onset + nucleus units as in (338). Here, the adjacent nasals are simply treated as two separate segments. (338) /mun-n-t/ [mu.nNt] /t-!azum-m-t/ [!ta.zu.mMt] 'they resemble each other' (ME85:37) 'you (pl. f) fasted' (ME85:37)

10.4.3 Evidence for underlying glides in ITB: Syllabification In this section, we will consider several types of evidence for the existence of underlying glides that relate to syllabification. First, there are sequences of high vocoids that contradict basic L-R syllabification. Second, underlying glides pattern differently than underlying vowels with respect to syllabification of other segments. Third, prepausal annexation applies differently to the two types of underlying high vocoids. Fourth, geminate glides and high vowels pattern differently. Sequences of high vocoids Sequences of segments which form a sonority plateau are subject to left-to-right syllabification. Examples of this were given in (333). More examples with the expected pattern of L-R syllabification of two high vocoids are given in (339). In non-initial

213 position, the first high vocoid surfaces as syllabic in accordance with the directionality of syllabification. (339) /i-sui/ /i-attui/ /t-ikiu-t/ /i-uti=i-il/ [i-suj] [j-attuj] [t-ikiw-t] [jutijil] 'he passed' (ME85: 39) 'it is high' (ME85: 39) 'kind of plant' (ME85:39) 'it's more than an arm (length)' (ME85:17)

There are some forms where the surface forms appear to be contradictions to L-R syllabification. In these forms, it is the second of two vocoids that surfaces as syllabic. (340) [i-zwi] [i-rwi] [i-swi] [a-jul] [!a-kjud] [tt-izwi] *izuj *iruj *isuj *aiwl *!akiwd *ttizuj 'he beat/ shook down' (ME85: 39) 'he mixed' (ME85: 39) 'excrement' (ME85: 42) 'donkey' (ME85: 42) 'braid (hair)' (ME85:42) 'become red' (ME85:88) /i-zwi/ /i-rwi/ /i-swi/ /a-jul/ /!a-kjud/ /tt-izwi/

Following Dell and Elmedlaoui, I conclude that the vocoids that surface as glides in these forms are underlying glides. Thus, the difference between the forms in (339) and those in (340) is that the latter possess sequences of /CGV/ while the former possess /CVV/. Let us assume for the moment that the forms in (340) were actually regular sequences of /CVV/. The derivation below shows the inadequacy of such an account. In particular, it is obvious that identical sequences of /sui/ 'pass' and /sui/ 'excrement' cannot surface with two different syllabifications [suj] 'pass' and [swi] 'excrement' without some lexical difference between the two forms. (341) UR CS(HV) Coda Output /sui/ (su) (suj) *[suj] 'excrement'

It is clear from the different outputs [suj] and [swi] that they must be represented differently. The analysis that I followed above was that the forms in (339) were the regular forms, while those in (340) showed the unexpected pattern and therefore had some special underlying status (i.e. an underlying glide). Let us briefly consider an alternative representation with no underlying glides. The alternative is to treat the forms in (340) as regular and those in (339) as anomalous. The vocoids in (339) that surface as vowels (supposedly unexpectedly) could be represented as lexically pre-linked to a nuclear position. (See Chapter 11 form more information on lexically nuclear vowels.) Such an analysis has three basic flaws. First, it would require that the forms in (339) be the exceptions. It would be these forms, which follow the basic L-R pattern of

214 syllabification that would have to be lexically marked in some way as exceptional. Second, there would still be no explanation for the apparent R-L syllabification of the forms in (340). Thus, there would need to be some stipulation that plateaus of high vocoids were syllabified R-L while other plateaus would be syllabified L-R. Finally, marking the forms in (339) as pre-linked to a nuclear position does not explain forms such as [t-wNza] 'fringed hair (bound state)' (*[tunza]) where assuming the underlying form /t-unza/ would never produce the actual output.142 Given the flaws in assuming that the difference between [suj] and [swi] is based on pre-linking the first /u/ to the nucleus, I dismiss this as a possible solution. Therefore, the conclusion is that it is the forms in (340) that are exceptional in some way. It is not possible to assume that the exceptional behavior of the forms in (340) is due to an underlying lexically nuclear vowel /zui=N/. This analysis still offers no explanation of the form [] *[tunza]. The analysis that assumes an underlying glide simultaneously accounts for the data in (340) and forms like []. These forms have underlying glides, while the expected and regular forms are simply sequences of underlying vowels. There is also evidence for underlying glides from the syllabification of high vocoids in initial position. Recall from (336) that in word initial sonority plateaus, it is not the leftmost segment that surfaces as syllabic, but instead is the second. The following forms do not fit with this pattern. (342) /i-wri/ /i-wrr/ [!iw.ri] [!iwRr] *!juri *!jurr 'yellow colored vomit' (ME85: 44) 'it/he is yellow' (ME85: 44)

Were the forms in (342) simply sequences of underlying vowels /iu/, then they would surface as [ju] as do the forms in (336). Thus there must be some difference between the forms in (336) and (342). We cannot analyze the forms in (342) as having a prefix /i=N/ that is lexically nuclear because this prefix can and does surface as a glide in the forms in (336). Therefore, the forms in (342) contain an underlying glide in the second position and thus not a plateau. The glide behaves here like other consonants in forcing the initial /i/ to surface as syllabic without an onset. Interaction between underlying glides and other consonants The evidence for underlying glides in this section comes from otherwise unexpected surface syllabifications of high vocoids and other consonants. If the vocoids in (343) below were merely underlying vowels, there would be nothing to prevent them from surfacing in a nuclear position. However, if they are glides, then it is not surprising that they behave differently from underlying vowels. The forms in (344) provide some minimal pairs for the two types of high vocoids.

We will see in the next section that underlying glides actually pattern with obstruents in terms of sonority.


215 (343) [t-wNza] [t-w .k-in] [!talwRtt] [t-wRta-t] [twR.jut] [!twRr] [lwR] *tunza *tukin *!talurtt *turtat *turjut *!turr *lur 'fringed hair (bound state)' (ME85: 43) 'metal pendants (bound state)' (ME85: 43) 'rose tree' (D&E88:fn4) 'kind of feline (bound state)' (D&E85:115) 'rein without bit (bound state)' (ME85: 44) 'it/she is yellow' (ME85: 44) 'run away' (D&E85:115) Underlying glide [lwR] 'run away!' [t-wRta-t] 'kind of feline (f)' (D&E85:115) (D&E85:115)

(344) Underlying vowel [lur] 'give back!' [t-urti-t] 'garden (f)'

The interesting thing about these forms is that the glides appear to be less desirable as syllabic nuclei than the other adjacent segments. For example, the form 'rose tree' surfaces with a syllabic [R] even though the glide presumably is more sonorous and in a position where it could have been put in nucleus. There is nothing that should bar the glide from syllabifying as nuclear; it would have had an onset (the /l/). It is not only sonorants that behave as more sonorant than the underlying glides. The form for 'metal pendants' shows a preference to syllabify [] in the syllable nucleus, instead of the glide. The sonority hierarchy that can be gleaned from these and other forms is given in (345). (345) Sonority Hierarchy of ITB (Elmedlaoui 1985): a > High Vowel > Liq > Nasal > [+vc] Fric > [-vc] Fric > [+vc] Stop > [-vc] Stop > Glide A reasonable question to ask is why Berber would treat the glide as less sonorous than these other segments. Perhaps the answer is if underlying glides were simply ranked between underlying high vowels and liquids on the sonority scale, there would be few cases where the difference between the two types of high vocoids would be noticeable. After all, if a language has underlying glides, one way to show the difference from vowels is to prevent them from surfacing as syllabic. One might expect that underlying glides would be protected from surfacing as syllabic vowels, lest they be indistinguishable from them. The data above show that in syllabification, underlying glides are treated as less sonorous than other consonants. Berber is a language in which any segment can surface as syllabic or non-syllabic. Even obstruents are allowed in the syllable nucleus. In fact, underlying glides in ITB are allowed to surface as syllabic in restricted environments as well. There appear to be some asymmetries between GV units and VG units where G is a glide. Elmedlaoui 1985 provides charts on onset and rhyme stability. Whereas the glide is an allowable and stable onset in [tw.kin], it is a dispreferred rhyme in *[ti.fLwt] and instead this surfaces as [tiflut] 'door' (ME85:49). If we were to simply follow the rules of core syllabification, we would expect the outputs in (346).

216 (346) UR /t-i-krw-t/ 'hair' CS(HV) (ti) CS(Liq) (kR) Coda (kRwt) Output *ti.kRwt Actual Output: [tikrut] /t-i-flw-t/ 'door' (ti) (fL) (fLwt) *ti.fLwt [tiflut]

It is not possible to reanalyze the forms in (346) as having underlying vowels /t-i-kru-t/ and /t-i-flu-t/ because when the plural suffix /-in/ is added, the forms surface as [] and []. If underlying forms of 'hair' and 'door' had only underlying vowels, then the plurals would surface as [tik.rujn] and [tif.lujn] with adjustment from rhyme stability or as [] and [] from Core Syllabification because of the L-R directionality of syllabification. (347) CS(HV) CS(Liq) CS(nas)

Coda and Rhymal Stability Adjustment

Assuming no underlying glide /t-i-kru-in/ /t-i-flu-in/ (ti) k (ru) i n (ti) f (lu) i n (ti) k (ru) (jN) (tik)(ru)(jN) * (ti) f (lu) (jN) (tif)(lu)(jN) *

Assuming underlying glides /t-i-krw-in/ /t-i-flw-in/ (ti) k r (wi) n (ti) f l (wi) n (ti)(kR)(wi) n (ti)(fL)(wi) n (ti)(kR)(win) (ti)(fL)(win)

Again, it is not possible to analyze the plural suffix as lexically pre-linked to a nucleus because according to Rosenthall 1994 (via Dell personal communication), there are no high vocoids in ITB that must surface as vowels. Therefore, the plural suffix cannot be lexically nuclear. Prepausal Annexation The rule of prepausal annexation provides yet another type of evidence for underlying glides. In ITB, clause-final light (open) syllables can adjoin to the preceding syllable as a complex coda. (348) Prepausal Annexation | / \ / | \ X A B XAB If B is [+sonorant] then the rule is optional. If B is [-sonorant], the rule is obligatory. However, prepausal annexation never applies when B is /a, i, u/ (D&E 1985: 120). The data in (349a) show possible alternations in the final syllable when the final segment is an

217 obstruent. In (349b), the final segment is a sonorant, but not a vowel. (349c) shows cases where the final segment is a vocoid and where prepausal annexation is allowed to apply. (349) a. Phr. Medial innanS !isbudD irkS ist Phr. Final innans isbukd irks ist (B=obstruent, obligatory) 'he said: spend the night' (Clements 1997) 'he poked (hi father's) eye out' (Clements 1997) '(his father) hid' (Clements 1997) 'he cracked (the door)' (Clements 1997) (B=sonorant, optional) 'she fled' (ME85:98) 'eagle' (D&E85:119) 'lock!' (D&E85:119) 'they are going' (D&E85:121) (B=glide, optional) 'until it is ripe' (ME85:98) 'drain' (ME85:98) 'shadow' (ME85:98) 'room' (ME85:98)

b. Phr. Medial Phr. Final t-RwL t-Rwl igidR igidr RgL Rgl ska.rN skarn c. Phr. Medial ard=i-nu a-saru a-malu a-anu Phr. Final ard=i-nw a-sarw a-malw a-anw

Recall that D&E85 note that prepausal annexation does not apply to forms that end in an underlying vowel /a,i,u/. The forms in (349c) surface with a final vowel in phrase medial position. We might be puzzled then by their behavior in undergoing prepausal annexation. However, these forms do not end in underlying vowels, but in underlying glides and are therefore able to participate in this process. The forms in (349c) illustrate another property of underlying glides. They show that underlying glides pattern with all other segments of Berber in being able to be syllabified in the nucleus. The only way to syllabify these forms by Core Syllabification is by putting the final glide in the nucleus. It is not possible for the glide to be in the coda phrase medially because the resulting form would contain hiatus, e.g. *[sa.Rw]. Geminate vocoids Further evidence for the difference between underlying glides and underlying vowels comes from the behavior of geminates. According to D&E 1985, in tautosyllabic "diphthongal" vowels, there is no discernable diphthongization. That is, the vowels are simply pronounced as long. In (350), an underlyingly geminate vowel can surface across a syllable boundary, or it can surface in the same syllable as a long vowel. (350) UR /!uur/ By CS [!u.wR] Prepausal Annexation [!uwr] = [!ur] 'to wait (perfect)'

218 Some more data from underlying geminate vowels is given in (351). Recall that geminates behave differently than simple sequences of equally sonorous segments. Most importantly, geminates never surface as onset + nucleus chunks. This prevents the form /iidar/ from surfacing as [jidar], as would be expected given the directionality of syllabification. The second thing to note is that in sequences of three identical segments, it is the first two which are geminate and the third simply behaves as a separate segment. Thus, /aiiis/ is actually /a(ii)is/ and not as /ai(ii)s/. The banned output *[ajijs] would be the expected output of /ai(ii)s/. (351) /iidar/ /aiiis/ /afauuu/ [ijdar] [ajjis] [afawwu] *[jidar] *[ajijs] *[afawuw] 'man's name' (D&E85:122) 'horse' (D&E85:122) 'blanket' (D&E85:122)

The data thus far show that underlying geminate vowels are allowed to surface as tautosyllabic [ij] as in [ij.dar]/[i.dar] or across a syllable boundary as in [aj.jis]. Elmedlaoui 1985 also provides data on geminate vowels that comes from assimilation. When the nasal /n/ in (352) assimilates, the result is a geminate glide. (ME85:147) (352) a. /u-i ==n/ 'the one ==of'143 /u-rgaz/ 'the man' w [win urgaz] ~ [wiw wRgaz] *[wigg Rgaz] b. /ian/ 'one' /u-ass/ 'day' [jan wass] ~ [jaw wass] (ME85:147) *[jaggwass]

The data in (352) show that an underlying vowel can surface as a geminate glide. We will see next that underlying glides which are geminated pattern differently. In the imperfect, the onset that is derived by Core Syllabification surfaces as a geminate. Certain segments are not allowed to surface as geminates, including the underlying glides. For example, /(w)/ surfaces as [qq(w)] and /w/ as [ggw] as the result of gemination. Unfortunately it is not possible to have an exact comparison between /w/ and /u/ in gemination contexts. Two reasons exist for this lack of comparison. First, for a stem to be geminable in the imperfect, it cannot have a syllabic vocoid in a non-final position (see § Second, gemination targets an onset segment. An underlying high vowel, as in [lur] is not in the onset and therefore not geminable.144

Here, == represents a prepositional boundary ("frontière du syntagme prépositionnel"). Elmedlaoui 1985 argues that underlying geminate glides can exist. He argues that the way they surface is due to the constraints on syllabification. He assumes the underlying form /!wwr/ [!uwR] 'act with patience' (ME85:148). It is not clear to me what the evidence is for these being underlying glides. D&E 1991, which discusses these forms, does not deal with the underlying representation of these. Stems of the form CHC (H=high vocoid) do not always behave as expected.

144 143

219 (353) Perfect zwi rwl [lwR] lwr [rwL] Imperfect zggwi rggwl lggwr

'shake down (fruit)' 'flee' 'flee'

(ME85: 51) (ME85: 51) (D&E88:14)

Geminate vowels are allowed to surface as long vowels or across a syllable boundary, but retain their identity as vocalic. Underlying glides which geminate due to morphological pressures occlusify and surface as stops. Summary This section has looked at different types of evidence for the existence of two distinct vocalic segments (underlying vowels and underlying glides) by looking at syllabification. First, sequences of glide-V behave differently than V-V sequences. Second, underlying glides behave as less sonorous than all other segments, whereas underlying vowels behave as more sonorous in syllabification. Third, underlying glides allow prepausal annexation which is banned when the final segment is an underlying vowel. Finally, underlying glides and vowels surface differently when they are geminates. According to Elmedlaoui 1985, at least some of these underlying glides are historically derived from stops. A few examples are provided in (354). Unfortunately these forms do not overlap with the data in this section. (354) Glide cognates in ITB (all are verbs) (ME85:150) ITB Hebrew Wza 'skin' gaza 'cut/take out' Wll 'raise' 'go up' ala Wnf 'divert/deflect/veer' kanaf 'wing' 'father/breed/produce' para 'father/breed/produce' Wrw Wrw 'be pregnant' hara 'be pregnant' Arabic qaa 'cut' 'go up' ala hanaf 'divert/deflect/veer' 'create' bara

10.4.4 Evidence for underlying glides in ITB: Morphology This section will look at the different behavior of underlying glides and vowels as they interact with the verbal morphology of ITB. The section on the aorist formation will deal only with underlying glides. The sections on imperfect formation and negative formation will show the contrasting behavior of underlying glides and vowels. Aorist Verbs of the form CXa fall into two classes. Here, C refers to a non-vocalic segment and X, any other segment. The choice of which class the verbs fall in is arbitrary, witness the minimal pair /mla/ with two different meanings that form the aorist in two different ways. Class I, where the final vowel is changed to [u], will be referred to as mutable. Class II has a detachable final vowel.

220 (355) a. Class I Perfect gra mla b. Class II Perfect !ra mla

Aorist gru mlu Aorist !R mL

'to gather' 'be wet'

(ME85:46) (ME85:46)

'read' 'show'

(ME85:46) (ME85:46)

Not surprisingly, perfect stems with an underlying glide in the second position also form the aorist in the two different ways illustrated in (355). Class I verbs simply replace the final [a] with [u], while Class II verbs delete the [a]. As a result of [a] deletion, the final segment surfaces as syllabic, as was the case in (355b). (356) a. Class I Perfect wa xwa !rwa b. Class II Perfect nwa swa zwa Aorist wu xwu !rwu Aorist nu su zu

'be good' 'evacuate' 'be full'

(ME85:47) (ME85:47) (ME85:47)

'cook' 'drink' 'dry'

(ME85:47) (ME85:47) (ME85:47)

Elmedlaoui 1985 argues against an analysis that the high vocoids in (356b) are underlying vowels. When the negative /-i/ is added to these forms, they surface as [nwi] and [swi] and not as *[nuj] or *[suj] as would be expected by the directional syllabification. The reason that the aorist of these forms is not [Nw] or [Sw] is due to the instability of these rhymes and the late readjustment of syllables that we saw in § A legitimate question to ask is whether the argument against the "vowel" analysis based on the negative formation is reasonable. That is, the paradigm nwa-nwi might be the result of some sort of analogy. The argument would be as follows. The underlying form of the perfect 'cook' is /nua/ and surfaces as the result of basic syllabification as [nwa]. Because the vocalic element is non-syllabic in the perfect (the stem form), then by analogy, when the negative marker is added, the stem vowel will again surface as nonsyllabic in [nwi]. There are two reasons to disregard this possibility. First, the notion that the morpheme boundary between the root and the negative /nu-i/ (supposed UR) somehow

221 blocks the basic syllabification that would produce [nuj] is not compatible with the facts of ITB. In ITB, the basic syllabification algorithm even applies across word boundaries. In (357a), the initial segment of 'boy' surfaces as syllabic because it is between two less sonorant segments. In (357b), on the other hand, the initial segment of 'boy' surfaces as non-syllabic because (i) there is a more sonorous segment next to it, namely the final [a] of the previous word and (ii) if it were syllabic then there would be an onsetless syllable. The fact that syllabication is not affected even by word boundaries implies that the morpheme boundary before the negative marker is not likely to be of importance. (357) a. Surfaces as syllabic (D&E89:168) i-ga tn gi-s 3ms-put:perf obj. 3 m. p. in-3s 'the boy put them (m) into it (f)' [u]frux boy

b. Surfaces as non-syllabic (D&E89:168) i-ga [w]frux ifullusn tgmmi 3ms-put:perf boy chickens in house 'the boy put the chickens into the house' The second piece of evidence against the notion that the morpheme boundary somehow prevents continuous syllabification comes from the general facts of syllabic alternations. One of the most noted facts of Berber is the fact that any segment can act as a syllable nucleus. In fact, the alternations between syllabic and non-syllabic allophones are one of the things that makes Berber Berber. Consider the forms in (358). (358) Syllabic alternations (D&E85:106) 3 p. f. /t-/, dat. 3 p. m. sg. obj /-a-s/ Root 2 p sg. /t-...-t/ rgl tRgLt tRglas xzn tXzNt txZnas lb tLbt tLbas nf tNFt tNfas Gloss 'lock' 'store' 'idem' 'graze (skin)'

From this data, it is clear that the syllabic association of the root final consonants in one form have no impact on the syllabic association in the other form. In these examples, the root final consonants are syllabic in the second column, but not in the third. This alternation is based purely on the surrounding segments and the general rules of syllabification in ITB. Furthermore, if this "vowel" analysis were to rely on analogy or syllable association to force the supposed /nu-i/ to surface as [nwi], then we might also expect some type of analogy to force the vocoid to surface as a glide in the aorist [nw] or [Nw] which it does not. The fact that the supposed underlying /u/ could surface as [u] in [nu] means that it should also be allowed to surface in nucleus in [nuj]. The fact that is does

222 not surface in this way, shows that the second segment in the forms in (356b) cannot be underlying vowels, but must be underlying glides. Additional evidence for these forms containing underlying glides comes from imperfect formation where perfect [nwa] surfaces as imperfect [nggwa]. The next section will deal with the imperfect in more detail. Imperfect There are three ways to form the imperfect from the perfect base: (1) /tt-/ affixation, (2) insertion of a 'chameleon' vowel, or (3) gemination (D&E 1991: 85). Some forms use a combination of these three processes. There are some conditions that must be satisfied for a perfect base to allow chameleon vowel insertion or gemination, but satisfying these conditions does not insure that the bases will use this method to form the imperfect. The evidence for underlying glides will come from geminable roots. To illustrate the complete paradigm of imperfect formation, some examples that take /tt-/ are given in (359). (359) /tt-/ prefixation mun tt-mun iwn tt-awn duj tt-duj sul tt-sul 'gather' 'climb' 'be awake' 'still be alive' (D&E91:85) (D&E91:86) (D&E91:86) (D&E91:86)

In order for perfect bases to use the chameleon vowel to form the imperfect, several conditions must be met (D&E 1991:88). (1) The bases must not be those of the class with a variable "a". (2) The penultimate segment must not be a syllabic vocoid.145 (3) There can be no geminate in the penultimate position. (4) The bases cannot begin with a syllabic vocoid. The chameleon vowel surfaces either as a copy of the last syllabic vowel or as [a] if there is no preceding syllabic vowel.146 The chameleon vowels are written in bold in (360).


This particular requirement is puzzling given the two perfect forms [lu] 'throw' and [!sug] 'drive' which are listed with the imperfect forms [tt-lwa] and [tt-!swag] with the chameleon vowel. The first form also has a variant [tt-lu] which is the expected form. I have no explanation why these forms are allowed to have the chameleon vowel given this particular requirement about the second segment not being a syllabic vocoid. 146 The chameleon vowel surfaces as [u] in some bases that begin with a coronal fricative and in bases that end in a labial. These will not be discussed here.

223 (360) Perfect a. sli b. skkiws c. mmujlt d. ggiwr e. ss-iwd f. !iwr Imperfect slaj tt-skkiwis tt-mujlut tt-giwir ss-iwid tt-!iwir 'graze' 'sit' 'feel sick' 'sit' 'frighten' 'take advice' (D&E91:88) (D&E91:88) (D&E91:88) (D&E91:96) (D&E91:96) (D&E91:96)

Notice that the chameleon vowel must surface as a copy of the preceding syllabic vocoid. The stipulation that it be syllabic is necessary to prevent [tt-mujlit] as the output of (360c). As far as I know, there is no evidence that the surface glide [j] is derived from an underlying glide. Thus, I simply assume that the underlying representation of (360c) is /mmuilt/. The third way in which imperfect verbs can be formed is by geminating one of the segments of the perfect base. There are several conditions which must be satisfied in order for a base to be geminable, though satisfaction does not guarantee that this is the way in which the imperfect will be formed. (1) The roots must be native (ME85: 51). (2) They must be triliteral (ME85: 51, D&E91:85).147 (3) The roots cannot contain a geminate already (ME85: 51, D&E91:85). (4) The root cannot have a syllabic vocoid in non-final position (ME85: 51, D&E91:85). (5) The base must be primary (i.e cannot be a causative or passive etc.) (D&E91:85). The segment that is geminated is the one that forms an onset after core syllabification. The forms in (361a) geminate the initial segment, while those in (361b) geminate the second. Because the geminated segment is always an onset, the final segment will never surface as a geminate. After core syllabification, the perfect forms in (361) would be (dL), (mRz), (xNg), (kRf), (M)(sX), (K)(sa), (G)(ra), though some readjustment might apply to create complex onsets. Thus, we see that in all cases, the segment which forms an onset is geminated. (361) Perfect Imperfect (D&E91:85) a. dl ddal 'cover' mrz mmrz 'be wounded in the head' xng xxng 'strangle' krf kkrf 'tie up' b. msx ks(a) gra mssx kssa grru 'transform 'pasture' 'pick up'

This requirement is a bit confusing given dl ddal does not appear to be trilateral. If /a/ were in the perfect, then it would violate the fourth condition.


224 As with the forms in (361), roots where the second segment is an underlying glide follow the same pattern of geminating the onset segment. Because the underlying glide behaves as though it is less preferred as a syllable nucleus, core syllabification will produce these syllabifications of the perfect forms: (N)(wa), (R)(wi), (Z)(wi), (R)(wL), (L)(wR). (362) Perfect nwa rwi zwi rwl lwr Imperfect nggwa rggwi zggwi rggwl lggwr 'cook' (ME85: 51) 'mix' (ME85: 51) 'shake down (fruit)' (ME85: 51) 'flee' (ME85: 51) 'flee/run away' (D&E88:14)

Compare the forms in (362) with those in (363). Recall that one of the conditions necessary for a verb to be geminable is for there to be no syllabic vocoids in the first or second positions. All the forms in (363) have a syllabic vocoid in the second position. A good pair to consider is /lwr/ 'flee' and /lur/ 'put back'. In /lwr/ gemination is possible because the second segment does not surface as a syllabic vocoid. In /lur/ on the other hand, the second segment does surface as syllabic and therefore does not fit the conditions of being a geminable verb. Notice that none of the verbs below in (363) surface as [suwi] or [swwi] as might otherwise be expected if it were possible to geminate the second segment. (363) Perfect suj duj lur lul mun sul Imperfect tt-saj tt-duj tt-lar tt-lul tt-mun tt-sul 'pass' 'be awake' 'put back' 'be born' 'gather' 'still be alive' (ME85: 53) (ME85: 53) (ME85: 53) (ME85: 53) (D&E91:85) (D&E91:86) Negative There are three ways in which the negative is formed from perfect bases. The first way is to optionally insert an /i/ between the final two segments of the base, as seen in (364). In order for this option to be available, the root must end in a consonant which is preceded by another consonant or a non-syllabic vocoid. In other words, bases which end in [...a/i/uC] will not allow i-insertion. The ban on syllabic vocoids in either of the last two positions confirms that the form in (364b) cannot be /rul/ for if it were the perfect form would be [rul] and it would not allow i-insertion to form the negative. Another interesting fact about this type of negative formation is that the first form in (364c) that contains [uw] or [u] acts as though the second part of the long vowel is not syllabic. The optional i-insertion does not apply to an underlying form, but must be sensitive to the

225 surface forms since it is allowed to apply in (364c). My proposed URs are also listed in (364). (364) a. Roots that end in CC dl ~ dil 'cover' (D&E91:83) skr ~ skir 'do' (D&E91:83) b. Roots with an underlying glide rwl ~ rwil 'flee' (D&E91:83) /dl/ /skr/ /rwl/

c. Roots with an underlying vowel that is not the head of the syllable !duwr ~ !duwir 'put around' (D&E91:83) /!duur/ !awr ~ !awir 'take advice' (D&E91:83) /!aur/ ajjn ~ ajjin 'record' (D&E91:83) /aaiin/ The second way to form the negative only applies to stems that end in /a/. In these stems, the final /a/ is simply replaced with /i/. Some other examples of this type were discussed in § Notice that the last two forms in (365) surface as GV and not as VG as would be expected by basic syllabification if they were simply sequences of underlying high vowels. (365) Stem ga uka nwa swa Negative gi uki nwi swi 'be, put' 'come' 'cook' 'drink' (D&E91:84) (D&E91:84) (ME85:47) (ME85:47)

Verbs that do not fit into either of the above classes form the negative through no change at all. Here, the negative stem is the same as the perfect. The verbs that fall into this last category are those that end in a variable "a", those that end in [i] or [u], or those that contain a syllabic vocoid in the penultimate position. (366) mun 'gather' uru 'give birth' !ksud 'be afraid' (D&E91:84) (D&E91:84) (D&E91:84)

The contrast between the forms [rwil] 'flee' in (364) and [mun] 'gather' in (366) is in the second root segment. In 'flee', the second segment is an underlying glide and therefore does not surface as syllabic adjacent to these sonorants. As a result, it is able to form the negative via i-insertion. The vocoid in 'gather' on the other hand is an underlying vowel and therefore surfaces as syllabic in this form. As a result, i-insertion is not a possible method of negative formation.

226 Summary This section discussed several types of morphological evidence for the existence of underlying glides. The aorist forms provided evidence that underlying glides can surface as syllabic under the right conditions. The imperfect data showed that there is a difference between bases that have an underlying glide in the second position and those that have an underlying vowel in this position. The glide bases fit into the geminable class, while those with an underlying vowel in the second position were not allowed to form the imperfect via gemination. Finally, evidence from negative formation showed that [rwl] is indeed syllabified with a glide and not a vowel and is therefore allowed to form the negative with optional i-insertion. 10.5 Representation The three languages in this chapter cannot be used to choose the best representation of underlying glides since all of the processes which diagnosed glides in this section were based on syllabification. In Cree, a ban on reverse sonority clusters forced the deletion of final /w/ when not adjacent to a vowel. In Pashto, similar clusters are allowed to surface. If the segments in Pashto and Cree were underlying vowels, then they would be expected to surface as vowels between a consonant and a word boundary. Since they are not vowels, they either surface as reverse sonority clusters (Pashto) or are deleted (Cree) based on the types of syllables that are allowed in the language. Similarly, the evidence for underlying glides in IT Berber came from syllabification. Since any segment can surface in nuclear position (i.e. can be syllabic), the surface difference between underlying glides and vowels is potentially at risk. The underlying vowels in Berber surface as either vowels or glides depending on the segmental environment. Underlying glides also surface as either syllabic (vowels) or non-syllabic (glides) depending on the environment. In a standard sonority hierarchy we expect underlying glides to be less sonorous than vowels but more sonorous than liquids, nasals, and obstruents. If this were the case in a language like Berber, then the difference between underlying vowels and underlying glides would only be visible when the two segments were adjacent. By targeting underlying glides as less sonorous than all other consonants, their special status as underlying glides is maintained. The one place where features of underlying glides and vowels come into consideration is in the morphological geminates in IT Berber. The underlying glides show the alternation [w] ~ [ggw] in morphologically derived geminates. Both LEX and [CONS] treat underlying glides as having the same features as vowels. In particular, underlying glides are considered to be [dorsal]. In VPT, the underlying glide is both [labial] and [dorsal]. In RAT, the underlying glide has the designated articulator [labial] and is [+round]. In addition to features under the Lips node, it also has features under the Tongue Body node. The expected output of gemination might therefore be a labial stop [b] which does not actually exist in the language. The alternative without the secondary velarization [b] is also not the actual realization. I suggest that it is more important in

227 Berber to retain all of the features from the segment than to maintain the designated articulator. Since [b] does not exist, the alternative is for /w/ to surface as [gw] under gemination contexts.


11 Languages without underlying glides

11.1 Types of languages without underlying glides Languages without underlying glides come in one of two types. I will refer to the first type as "normal" and the second as "vocalic". In the "normal" set, there is one underlying vocoid (which I will call a vowel) that surfaces either as a vowel or a glide depending on the parameters of the language (e.g. whether it allows hiatus, how word minimality is satisfied, cluster restrictions, etc.). I conclude that this set of languages represents the normal situation since the parameter settings generate completely predictable outputs. It is evidence from these "normal" languages that I believe led to the notion in some circles (e.g. Kaye & Lowenstamm 1984) that underlying glides do not exist. Languages in this set might show glide/vowel alternations or no alternations at all. What unifies the languages of this group is that there are no vocoids that surface in unexpected syllable positions. That is, there are no glides in places where a vowel would be a better syllabic parse. Similarly, there are no vowels in places where a glide would be expected. In the "normal" languages, glides surface when they are adjacent to better vowels/nuclei, while vowels surface when such a segment is absent. Of course languages can stray from this general pattern. In Tukang Besi, there appear to be no surface glides (Donohue 1999, Levi 2004). This language would simply require that any [-consonantal] segment always surface as nuclear. Rosenthall 1994 provides an overview of the possible manifestations of vowel sequences, which encompasses the various ways in which high vowels can surface. The languages that he treats are of the "normal" type. Crucially, in these languages, there never appears to be a contrast between a vowel and a glide, as there might be in languages with underlying glides or in the "vocalic" languages. These "normal" languages correspond to Type III (or I) from the typology outlined in Chapter 1. The second type of language without underlying glides will be referred to as "vocalic". In these languages, there appears to be some contrast in the vocoids. Instead of exhibiting an over-abundance of glides, implying that the language has underlying glides like the nine languages dealt with in this dissertation, the "vocalic" type of languages exhibit an over-abundance of vowels. Vowels surface in places where a glide, under normal circumstances, is expected. For example, in French there a few near minimal pairs of the form [Vi] vs. [Vj] (e.g. [abi] 'abby' and [abj] 'bee'). These languages are what have caused linguists to seek ways to distinguish alternating from non-alternating vowels. I will argue that these languages contain some vowels that are lexically marked as nuclear. Since they are already marked with some syllabic structure, they must remain syllabic even if they would be expected to surface as a glide under

229 normal syllabification. The notion of lexically specifying some structure on a limited set of underlying vowels originates in Levin 1985.148 The languages that have this "vocalic" pattern look similar to Type VII on the surface. Type VII was one of two types that appear to be non-existent. In a Type VII language there is an underlying glide which shows a predictable realization as either a vowel or a glide on the surface, and there is an underlying vowel which only surfaces as a vowel (see (367b) below). The underlying form is not given to a linguist. Instead, we begin with (367a) and must determine the true nature of the language. Since the pattern exemplified in (367a) has an abundance of vowels on the surface, the alternative in (367c) is possible. In a language with the pattern in (367c), high vowels show a "normal" pattern, where they alternate depending on the parameters in the language. In addition to these normal vowels, there are also some segments that surface as vowels when, all things being equal, they should not. (367) a. /?/ [i] /?/ [j] b. /i/ [i] /j/ [j] c. /i=N/ /i/ [i] [j]

To truly distinguish between the analyses in (367b) and (367c), it would be necessary to have some process which could distinguish the features of these segments. If the segments that refused to surface as glides interacted with other vocalic processes while the alternating segments did not, then the representation in (367b) would be accurate. If, on the other hand, the alternating segments and the non-alternating ones patterned together with respect to the featural process, then the representation in (367c) would be correct since these underlying segments have the same features. Another possible way to distinguish between the two analyses would be to look at the proportion of these segments. If the basic pattern is to alternate, but there is a small and limited set that does not alternate, then it implies the analysis in (367c). This is the case in Korean (see §11.3.3). If the non-alternating set of vowels is also limited to borrowed vocabulary, this can also help establish the analysis in (367c). That is, in (367c), the segment which always surfaces as a vowel is considered anomalous. An overall diagram of languages with respect to glide typology is given in (368). First, languages divide into two sets, those with underlying glides and those without. Of those languages that do not have underlying glides, there is again a split. On the left are those languages with only predictable realizations. These languages have only /i/. On the right are those languages that have an underlyingly lexically nuclear vowel in addition to having a predictably alternating /i/.

This is similar to Deligiorgis 1988 and 1993, though instead of using structure to mark some vowels as exceptional, she uses [-consonantal]. She assumes that most (high) vowels are [0consonantal] and therefore can surface as nuclear or non-nuclear. The lexically marked [-consonantal] vowels, however, are forced to always surface with the syllabic parse.


230 (368)

Set of all lgs.

Lgs. with underlying glides (see rest of dissertation)

Languages without underlying glides

/i/ ·V/G alternations depend on parameters in the language. ·E.g. -word minimality -ban on hiatus -ban on clusters -ban on secondary articulation ·E.g. Rosenthall 1994

/i/ and /i./ ·In addition to "normal" Vs, has lexically nuclear Vs. ·Characteristics of /i./: -surfaces as V when not expected to. -generally more marginal in the lexicon. -patterns with /i/ in feature-based processes ·E.g. Levin 1985

The remainder of this chapter will provide a brief overview of some languages that do not have underlying glides. Kimatuumbi, Karuk, and Sanskrit will serve as representative examples from the "normal" set, while Ait Seghrouchen Berber, Spanish, and Korean will do so for the "vocalic" set. 11.2 "Normal" languages A sketch of three "normal" languages with predictable glide/vowel realizations follows. In Kimatuumbi, underlying high vowels can surface as vowels, glides, or secondary articulations on consonants depending on the surrounding environment. The palatal glide in Karuk patterns with vowels in phonological processes. In Sanskrit, the glides alternate with vowels and are transparent to long-distance consonant harmony. 11.2.1 Kimatuumbi Kimatuumbi is a Bantu language spoken in Tanzania. In Kimatuumbi, the distribution of surface high vowels and glides, and in fact of all vowel sequences, is predictable. The data and basic analysis in this section is taken from Rosenthall 1994 (citing Odden 1992), though some of the constraints have been modernized to reflect the changes in the OT literature.

231 In Kimatuumbi, sequences of non-high vowels surface in hiatus. If the second vowel in a sequence is high, it also surfaces in hiatus (c.f. (369a)) since falling diphthongs are disallowed. (369) a. Non-high Vowels in first position /ma-oto/ [] 'large fires' /ka-ula/ [] 'small frog' b. C + high vowel /mi-oto/ [] /ki-ula/ [] c. High vowel-initial words /i-ula/ [] c.f. /i-tabua/ [i.ta.bwa] /u-a-teleke/ [] c.f. /u-telike/ []

'fires' 'frog'

'frogs' 'books' 'you should cook' 'you cooked'

When the first vowel is high, there are two different ways it can surface. If the high vowel is in pre-vocalic position and preceded by a consonant, then it surfaces as a secondary articulation on the preceding consonant. In word-initial position, with no preceding consonant, the high vowel surfaces as a glide. In both of these latter cases, the vowel following the non-moraic high vowel is long. A sample of the data is provided in (369). Rosenthall 1994 uses an OT approach to account for the realization of high vowels in a variety of languages. The constraints that will be used for the Kimatuumbi data are given in (370). The names of some of these constraints have changed and their more modern forms are provided. The names used by Rosenthall are given in parentheses. (370) Constraints for Kimatuumbi a. DEPC: Do not insert a consonant. (FILL) b. MAXµ: Do not delete a mora from the input. (PARSE-µ) c. MAXV: Do not delete a vowel. d. {A}=V: Non-high vowels must be syllabic. e. ONSET: Syllables must have onsets. f. NLV (NOLONGVOWELS): Long vowels are banned. g. SECART: Consonants with secondary articulations are banned.

232 The three tableaux that follow show how the interaction of the constraints determines which form surfaces.149 In all of these examples, a light diphthong is notated by putting both vowels in parentheses. Rosenthall argues that all ongliding diphthongs are light.150 As a result, they all violate MAXµ. If they are heavy, the constraint NODIPH, which prohibits diphthongs, can simply be added. The tableau in (371) illustrates the ranking DEPC, MAXµ, MAXV,{A}=V >> ONSET. The tableaux in (372) and (373) show ONSET >> NLV, SECART. (371) Tableau for /ma-oto/ 'large fires' DEPC MAXµ MAXV a. b. *! a c. m d. *! e. m(ao).to *! f. *! * *! g. (372) Tableau for /mi-oto/ 'fires' DEPC MAXµ a. b. *! j c. m d. *! e. m(io).to *! f. *! g. {A}=V * *! ONSET * NLV SECART * *









* *!


Rosenthall 1994 shows a full factorial typology. The illustration here of Kimatuumbi is simply intended to provide an example of a language with predictable high vowel alternations. 150 It is actually not the case that all ongliding diphthongs are light. In Spanish they pattern as heavy for stress assignment (Harris & Kaisse 1999).


233 (373) Tableau for /i-ula/ 'frogs' DEPC MAXµ a. .(iu).la *! b. c. *! d. e. *! * j *! f. C g. *!* {A}=V ONSET * *!* NLV SECART

* * * *

This section has shown that the distribution of high vowels in Kimatuumbi is predictable. If no other vowel is near, high vowels surface as syllabic vowels (see (369c) 'books'). If an onset consonant is available, the high vowel surfaces as a secondary articulation. If no onset consonant is available, the high vowel surfaces as a glide in the onset position. Losing the mora from the underlying vowel results in compensatory lengthening of the following vowel. 11.2.2 Karuk A second example of a language with an underlying high vowel that lacks a corresponding underlying glide is Karuk. In Chapter 8, we saw that Karuk has an underlying labial vowel /u/ that surfaces as either a vowel or a glide, but it also has an underlying labial glide /w/. The high front vowel /i/, on the other hand, does not contrast with its corresponding underlying glide /j/. This underlying vowel can surface as a glide, but whether its surface realization is a vowel or a glide depends on the syllable phonotactics and other characteristics of the language. Thus, Karuk is a language with a high vowel that lacks an underlying glide counterpart. Evidence for the lack of an underlying glide /j/ in Karuk comes from three properties of the language. First, the realization of /i/ as either a vowel or a glide is predictable based on the surrounding segments. As we saw in Chapter 8, high vowels surface as vowels when surrounded by consonants ([...CiC...]) and as glides when adjacent to non-high vowels ([...aj...] or [...ja...]). When two high vowels are adjacent in word-medial position, the leftmost high vowel surfaces as a glide. When two high vowels are adjacent in word-final position, it is the rightmost one that surfaces as a glide. Two tableaux from Chapter 8 are repeated below.

234 (374) /imiuha/ [imjuha] 'soap plant' *[imiwha] *COMPLEX ONSET DEP MAX *STRUC a. im.ju.ha b. i.miw.ha c. i.mi_.ha d. i.mi.u.ha *** *** *** ****

*! *! *[tatwi] *V# *STRUC ** ** *** **

ALIGNR 1 2 3 **** ** ***** **! **** ** ***** *** **

(375) /tatui/ [tatuj] 'to sweep' ONSET DEP MAX a. ta.tuj b. tat.wi c. ta.tu.i d. ta.tu_

*! *!

*! * *

ALIGNR 1 2 *** ** *** * **


The second type of evidence for the lack of /j/ comes from the interaction of vowels at morpheme boundaries. Though tautomorphemic sequences of vowels exist underlyingly as shown in (243) and (247), heteromorphemic sequences of vowels result in deletion and coalescence. The data in (254a) show that morpheme-final vowels are deleted before vowel-initial suffixes. In (254b), the same is true for /u/ even if it surfaces as [w] in related forms. The final vowel /u/ is deleted and the remaining two vowels coalesce. In (254c), the same is shown for /i/. Here, final /i/ is deleted and the remaining vowels coalesce. All words that end in a surface glide /j/ behave as though this segment were a vowel when another vowel-initial suffix is added. (376) a. Final /V/ [V] when unsuffixed /tuphi-a/ [tuph-a] 'speech' /ina-is(rih)/ [in-i(rih)] 'to kneel down' b. Final /u/ [w] when unsuffixed /ikriu/ [ikriw] /ikriu-is(rih)/ [ikrî(rih)] /ikriu-at/ [ikrêt] /ihiiu/ [ihjiw] /ihiiu-unis/ [ihjûni]

'to sit' 'to sit down' 'lived' 'to shout' 'to shout at'

235 c. Final /i/ [j] when unsuffixed /iiuai/ [ijwaj] 'to pour' /iiuai-is(rih)/ [ijwê(rih)] 'to pour down' The third piece of evidence that surface [i] and [j] are derived from the same source comes from s-palatalization. In Karuk, both [i] and [j] cause a following /s/ to surface as []. Palatalization is blocked by intervening vowels, but not by intervening consonants. Due to the phonotactics of Karuk, however, it is only possible to find [j] strictly adjacent to the target /s/. Some data is repeated below. The trigger is in bold. (377) a. Palatalization immediately following [i] [tiih] 'dog' [piri] 'grass' [pi] 'soaked acorns' b. Intervening C is transparent [pikip] 'shadow' [ikup] 'to point' [ikus] 'to tear up a rat's nest' [usiwap] 'he plugged it up' c. Intervening V blocks (between [i] and [s]) [itkus] 'walk arm in arm' [ihkus] 'deceased mother' d. Palatalization immediately following [j] [tujip] 'mountain' e. Intervening V blocks (between [j] and [s]) [firipamjusah] 'place name' As stated in Chapter 8, Karuk shows that the two high vowels in a single language can differ in whether or not they contrast with a corresponding glide. The labial vowel /u/ and glide /w/ contrast in Karuk, showing that these segments have a Type VI relationship. The front vowel /i/, on the other hand does not contrast with a high front glide /j/, though it does show both vowel and glide surface realizations. Thus, the front vowel exhibits the "normal" behavior of a single vocalic segment with two surface realizations.

236 11.2.3 Sanskrit Sanskrit is another example of a language which does not have underlying glides and in which the realization of underlying high vowels as either syllabic or non-syllabic is predictable. Sanskrit is also interesting for our purposes because it has a type of consonant harmony known as ati, or n-retroflexion.151 The consonant inventory is provided in (378). (378) Consonant Inventory of Sanskrit (Cardona 1987) Labial Dental Retroflex "Palatal" Stops p p b b t t d d t t d d Fricatives s Nasals m n Liquids l r Velar k k g g Pharyngeal h

A few notes about the transcription are necessary. The symbol <h> is described as a voiced pharyngeal. The symbols <t, t, d, d> replace the traditional transcriptions <c, ch, j, jh>. Cardona 1987 also includes a few segments which either have a limited distribution or an allophonic status (,,m, h). The two nasals and generally occur before velars and palatals. The nasal m is a nasal off-glide. Finally, h is a voiceless pharyngeal found in word-final position. In addition to these segments, the two glides <j> (traditional y) and <v> are generally included in the consonant inventory.152 They have been omitted in (378) since they do not have phonemic status. The non-nasal sonorants have two realizations, one syllabic and one nonsyllabic.153 Because we are interested in the two glides j and v, this discussion will be primarily limited to them. Whitney remarks that the alternations i ~ j and u ~ v are frequent (Whitney 1879: §55-58).154 Uhlenbeck 1898 calls j the "consonant i" and describes the relationship between u and v to be the same. I interpret this to mean that the glides are the non-syllabic counterparts of the high vowels i and u. Before a different vowel (i.e. syllabic segment), the segments i/u/r become their non-syllabic counterparts j/v/r (Whitney 1879: §129, Cardona 2003). Some examples are provided in (379) and

Hansson 2001 does not consider the Sanskrit n-retroflexion to be an example of true consonant harmony. I simply use the term here to refer to a long-distance assimilation between consonants. 152 [v] acts as the non-syllabic version of /u/. It is described as being produced "with just the slightest contact between the upper teeth and lower lip; slightly greater than that used for English w (as in "wile") but less than that used for English v (as in "vile")" (Goldman & Goldman 2002: 7). 153 The syllabic lateral has a limited distribution, being found only in klp 'be fit, arrange, imagine', with the past participle klpta- (Cardona 1987). 154 When discussing i/j he writes, "the two exchange with one another in cases innumberable."


237 (380). The data from Whitney 1879 are not glossed. The transcriptions have been adapted to reflect the inventory in (378).155 (379) Syllabic and non-syllabic alternations in Sanskrit a. i ~ j itj aha (from iti + aha) W129 strj asja (from stri + asja) W129 vjui (from vi + ui) W130 dadj atra '(there is) yogurt here' Cardona 2003 strj atra 'the woman (is) here' Cardona 2003 b. u ~ v madhv iva vadhvai svia vadvagarac. r ~ r duhitrarthe (from madhu + iva) W129 h (from vad u + ai) W129 (from su + ii) W130 'bride's house' (vadu + agara) Cardona 2003

(from duhitr + arthe)


(380) Alternations in a paradigm (Cardona 1987: 458) Root Acc. Dat. Loc. agniagni agni-bjam agnj-os vajuvaju vaju-bjam vajv-os madumaduni madu-bjam madv-os devidevjau devi-bjam devj-os pitrpitarau pitr-bjam pitr-os

'fire' 'wind' 'honey' 'goddess' 'father'

Given the ease with which the syllabic and non-syllabic segments in (379) and (380) alternate, I conclude that they represent a single segment underlyingly. I assume for the glide/vowel pairs that the vowel is underlying and that the glide surfaces when it results in the best syllabic parse. There are other alternations where j/v is expected but the meter requires an additional syllable. In these cases, they surface as ij and uv. This does not nullify the analysis because in this case, there are conflicting metrical and syllabic demands. The glide is ideal for the general syllabification, but an extra syllable

Here, W# refers to Whitney 1879 §#, SS# refers to Schein & Steriade 1986 page #, and B# to Burrow 1955 page #.


238 is needed for the meter. Clearly, a further look at the glide/vowel alternations and distinction is necessary, but will not be attempted here. What makes Sanskrit especially relevant for this dissertation is the (non)interaction of glides in n-retroflexion. The rule of Sanskrit n-retroflexion (ati) changes a dental nasal n to a retroflex nasal when it is preceded by a retroflex continuant (i.e. , r, r, or r). N-retroflexion is restricted to occur before a non-liquid sonorant (i.e. a vowel, j, v, n, or m).156 It is blocked if a coronal (i.e. "palatal", retroflex obstruent, or dental) intervenes between the trigger and the target (Whitney 1879). N-retroflexion, however, is not blocked by the palatal glide j (Whitney 1879, Emeneau & Nooten 1968). The data in (381) show examples where n-retroflexion occurs. In (381a) the trigger and target are adjacent. (381b&c) show the target and trigger separated by a vowel and a consonant respectively. Finally, in (381d) the target and trigger are separated by an even greater distance. (381) Examples of n-retroflexion a. Adjacent vr-o-ti 'covers' pitr-am 'of the fathers' pu-a-ti 'makes to thrive' i-a 'seek-pres.' pr-a'fill-pres.'

Cardona 2003 Cardona 2003 Cardona 2003 SS717 SS717

Schein & Steriade 1986 note that this restriction on the environment can be eliminated once other factors of Sanskrit phonology are realized. They show that (1) before non-syllabic continuants (s, , , r, and h) n surfaces as a nasal vowel (anusvra) and therefore bleeds n-retroflexion, (2) nl sequences do not appear to occur word-medially and across a word boundary, the sequence is realized as a nasal lateral, (3) word-final nasals are neutralized to n (except for inflexional suffixes), and (4) n+stop clusters show nasal place assimilation which overrides the n-retroflexion.


239 b. Separated by V vari-i 'water-acc.' vari-os 'water-loc.' vari-a 'with water' purua-am 'of the men' rudrea rudraam dveai iai takaa c. Separated by C ( coronal) vrk-a'cut up-passive' grb-a-ti 'seizes-3 sg. active'

Cardona 1987 Cardona 1987 Cardona 2003 Cardona 2003 W190 W190 W190 W190 W190

SS717 Hansson 2001

d. Separated by several segments brh-aa'reinforcing, supporting' Cardona 2003 kub-aa 'quake-middle' SS717 ruga W190 dravia W190 rekas W190 krpamaa W190 The data in (382) show that the retroflex and plain n alternate. In these examples, the suffixes surface with a retroflex nasal when a retroflex continuant precedes them, but as a dental nasal when no such segment exists. In the last two forms, it is the root-initial nasal that alternates.

240 (382) Alternations [] ~ [n] Retroflexed vr-o-ti 'covers' pitr-am 'of the fathers' pu-a-ti 'makes to thrive' vari-a 'with water' brh-aa'reinforcing' kub-aa 'quake-middle' pra amati 'bows' pra ajati 'leads forth' (Cardona 2003)157 'presses' 'of the gods' 'purifies' 'with fire' 'saying' 'hum-middle' 'bends' 'guides'

Unchanged su-no-ti deva-nam pu-na-ti agni-na vat-anaked-ana namati najati

When the target occurs in final position, it does not surface as retroflexed. All final nasals are neutralized to [n] (see footnote 156 and Schein & Steriade 1986 for more clarification). However, when a suffix is attached, the stem-final nasal surfaces as retroflexed. (383) Root-final alternations W190 /ran-/ brahman (voc.) raanti brahmaa rajati brahma-i (loc.) raraa brahma-ja (relating to X) brahmavant N-retroflexion is blocked if a coronal intervenes between the trigger and the target, as shown in (384). Other works have not provided evidence showing that retroflexes block, though Whitney 1879 states so explicitly. The examples in (384c) come from Whitney 1885 and represent new evidence that retroflexes indeed block ati. Examples of this type are rare since the retroflexes represent an innovation from ProtoIndo-European (Burrow 1955: 95).


The forms for 'quake' and 'hum' come from Schein & Steriade 1986, page 717.

241 (384) Blocking by coronals a. Dentals mrd-na *mrd-a 'be gracious-pres.' mrd-nati *mrd-ati kved-ana *kved-aa 'hum-middle' krt-a-mana- *krt-a-maa- 'cut-middle' b. "Palatals" mard-anaard-una kro-ana-

SS717 SS718158 SS717 SS717

*mard-aa- 'wipe-middle' *ard-ua 'white' *kro-aa'shouting'

SS717 B84/151 B150

c. Retroflex obstruents roamana *roamaa kru jamana *kru jamaa mrana *mraa veana *veaa

'torment-pres.' 'thicken' 'be gracious' 'wrap'

Whitney 1885 Whitney 1885 Whitney 1885 Whitney 1885

While coronal consonants block the spreading of retroflexion, the palatal glide does not (Whitney 1879). In (385), the trigger and target are separated by several segments, including [j], but retroflexion is still allowed to occur. (385) turv-ajaa har-ajaa dak-ajaa 'victory-adj.' B155 'proper name-adj' B155 no gloss/patronymic B151/198

The non-interaction of the palatal glide can be explained using both feature geometries since the glide is not phonemic, but derived from an underlying vowel. In Vowel-Place Theory (VPT), the glide has features under V-place since it is derived from a vowel. Its transparency is thus explained, since it lacks any C-place features that could interact with the n-retroflexion. Similarly, in Revised Articulator Theory (RAT), the glide lacks the spreading feature. N-retroflexion is a relationship between a coronal trigger and a coronal target which is blocked by other coronals. Since the glides in Sanskrit are not underlying, but are derived, they have a designated articulator dorsal.

I believe these two forms were mis-glossed in Schein & Steriade 1986. The root mrd is glossed as 'rub, crush', while the root mr with a retroflex stop is glossed 'be gracious' in Whitney 1885. Therefore, the two forms in (384a) should be glossed as 'rub, crush', contrasting with the forms in (384c).


242 Sanskrit was the third example of a "normal" language. Underlying high vowels surface as either vowels or glides depending on the environment. Because surface glides are derived from underlying vowels, they do not act as blockers (unlike coronal consonants) in n-retroflexion. 11.3 "Vocalic" languages This section will provide a sketch of three "vocalic" languages which have predictable glide/vowel alternations as well as some extra occurrences of vowels. As discussed above, the instances where we find an over-presence of vowels (i.e. vowels where glides are expected) should be less numerous than those which exhibit the standard glide/vowel predictability. In this section, both i=N and i. (i followed by a period) will be used to symbolize the underlying vowels that are lexically nuclear. This section will discuss Ait Seghrouchen Berber, Spanish, and Korean. 11.3.1 Ait Seghrouchen Berber The first example of a language with lexically nuclear vowels is Ait Seghrouchen Berber (AS Berber). In this dialect of Berber there is a contrast between vocoids which show surface alternations and those that always surface as nuclear/syllabic (Guerssel 1986). In Guerssel's analysis the distinction is between segments with a lexically prelinked rhyme slot and those without such a specification. He uses the term vowel to refer to the lexically nuclear vowels and glide to refer to those unmarked segments which are allowed to alternate. I will use the term vowel for all of these segments. Those vowels that are lexically marked as nuclear will be referred to as lexically nuclear. The basic alternation between surface glides and vowels is given in (386). In (386a) the third person subject prefix /i-/ and the construct state prefix /u-/ surface as a vowel when adjacent to a consonant in word-initial position and as a glide in prevocalic position. (386b) shows that these prefixes even surface as a glide across a word boundary in order to satisfy syllabification. These examples also show that the glide can occur in the coda position. In (386c), a stem-final high vowel surfaces as a vowel before the consonant-initial third person plural object marker /-ten/, but as a glide before the vowelinitial first person plural object marker /-ax/. (386) Basic alternation pattern (data from Guerssel 1986) a. i-ru 'he cried' j-ari u-mazan 'messenger' w-ansa b. is i-ru x u-mazan c. tusi-ten tessu-ten tuli-ten 'did he cry' 'on the messenger' 'he writes' 'place'

aha j-ru 'then he cried' arra w-mazan 'messenger's kids' 'she carried us' 'she made us drink' 'she ascended for us'

'she carried them' tusj-ax 'she made them drink' tessw-ax 'she ascended them' tulj-ax

243 In addition to these alternating segments, AS Berber contains some vowels which are prohibited from surfacing as glides. These segments surface as vowels even in places where a glide is expected. The data in (387) contrast with those in (386c). In (387), the stem-final vowels do not surface as glides before a vowel-initial suffix. Instead, they surface as nuclear and force epenthesis to prevent hiatus. Because these data show an over-abundance of vowels, the analysis with lexically marked nuclear vowels is optimal. The forms in (386) simply have underlying high vowels /i/ and /u/, while those in (387) have lexically nuclear vowels /i=N/ and /u=N/. (387) Lexically nuclear vowels (data from Guerssel 1986) tessu-ten 'she made them into a bed' tessu-j-ax *tessw-ax tettu 'she forgot' tettu-j-ax *tettw-ax turi-ten 'she wrote them' turi-j-ax *turj-ax tini-ten 'she says them' tini-j-ax *tinj-ax 'she made us a bed' 'she forgot us' 'she wrote us' 'she tells us'

Guerssel's analysis differs slightly from the one outlined in this section. He assumes that all underlying vowels which only surface as vowels are marked as nuclear. This includes vowels that are morpheme internal and inter-consonantal. Under his analysis, [tusi] 'she carried' has the underlying representation in (388a). Since the surface realization of this segment is predictable, I assume the representation in (388b) where the segment is unmarked.159 (388) a. Guerssel's UR N | /t u s i/ b. Alternative UR /t u s i/

The data in this section showed that AS Berber has two types of underlying vowels. The first is a set of unmarked high vowels that surface optimally as syllabic or non-syllabic depending on the environment. The second is a set of lexically nuclear high vowels that always surface as nuclear even when the environment would call for a nonsyllabic parse.


From the data in Guerssel 1986, I assume that this form must actually be /t-usi/ with a prefix /t-/. There is no data in that paper to determine if this root is prohibited from surfacing with a glide. Vowel-initial roots show differences in behavior. For example, the forms 'I did not fall' and 'I did not fold' both surface with a vowel after a consonant-final word [ur ud!i-x] ([d!] represents pharyngealization). When a vowel prefix is attached, however, the difference in these two roots is realized: [ur j-ud!i]/[ur i-wd!i] 'he did not fall' but [ur j-ud!i] 'he did not fold'. Having this vowel prefix shows that 'fold' must have a lexically nuclear vowel, while 'fall' does not. This type of evidence was absent for the form in (388).

244 11.3.2 Spanish The second example of a "vocalic" language is Spanish. Harris & Kaisse 1999, citing Navarro Tomás 1918[1965], argue that there is a distinction in Spanish among high vocoids. Both Roca 1997 and Harris & Kaisse 1999 argue that the distinction in Spanish is not between underlying glides versus underlying vowels, but instead between underlying vowels which are either marked or unmarked with some syllabic structure.160 The data in (389) provide some surface contrasts between glides and vowels. In Spanish, orthographic h has no realization and is put in brackets in the examples below. (389) Peak161 a. [i] = /i=N/ vac[i.a]ba d[i.a]blo m[o.i]na [o.i]remos b. [u] = /u=N/ z[u.a]vo s[u.e]co <h>[u.i]da <h>[u.i]dizo Non-peak [j] = /i/ vic[ja]ba bon[ja]to b[oj]na [oj]gamos [w] = /u/ s[wa]ve z[we]co c[wi]da c[wi]dada (Harris & Kaisse 1999) 's/he vitiated' 'sweet potato' 'beret' 'we hear (subj.)'

's/he emptied' 'devil' 'annoyed' 'we will hear'

'Zouave' 'Swedish' 'fight' 'fleeting'

'smooth' 'wooden shoe' 's/he cares' 'care'

Harris & Kaisse point out that surface glides and vowels occur in the same environments. In particular, the realization of a segment as a vowel or a glide is not based on stress; for example both segments can occur immediately before a stressed vowel, vac[i.a]ba vs. vic[ja]ba. The question is how to represent this lexical distinction. Harris & Kaisse argue that the segments in the second column of (389) are simply underlying vowels, while those in the first column are underlyingly marked as nuclear, having the underlying structure in (390). I have used the notation /i=N/ to represent these lexically nuclear vowels.

Harris & Kaisse avoid the term vowel in favor of the more neutral vocoid in order to highlight the fact that these segments can surface as either a vowel or a glide depending on the environment. In order to be consistent with the terminology in this dissertation, I will use the term vowel for such an alternating segment in order to distinguish it from an underlying glide, which is also a vocoid. 161 Levin 1985 argues that lexically marking a vowel as nuclear only occurs at morpheme edges. The data from Spanish, however, show that lexical marking can occur morpheme internally as well.


245 (390) N | /i/ N | /u/

In order to prove that this lexical marking is the correct representation, it necessary to consider the distribution of these segments in more detail. First, the realization of an underlying vowel is predictable based on its environment (Harris & Kaisse 1999: 124). When the underlying high vowel is the most sonorous segment, it surfaces as a vowel. For example, the underlying vowel in 'floor, story' surfaces as a vowel since it is surrounded by consonants, []. When a better nucleus is available, however, the underlying vowel surfaces as a glide as in [] 'sweet potato'. The lexically nuclear vowels, on the other hand, do not show this predictable alternation.162 Instead, they surface as nuclear/syllabic when the glide realization would otherwise be expected, as in [] 'hiatus'. The second fact that lends credence to this analysis is the distribution of these vowels. While those underlying vowels that surface as glides are abundant, those which surface as vowels in gliding contexts are not (Harris & Kaisse 1999: 185). To illustrate this point, Harris & Kaisse provide a 'representative' list of present-tense verb forms. In (391a), the underlying high vowel surfaces as a glide as expected. In (391b) on the other hand, the high vowel surfaces as a vowel, forcing hiatus. Harris & Kaisse point out that the proportions of examples in (391) represent the general pattern. That is, the cases in which the high vowel surfaces as a glide are much more common. The lexically marked high vowels, which require extra structure and do not show the expected alternations, are not numerous, just as expected.

There is an optional process of denuclearization where these lexically nuclear vowels can surface as glides, but this is the result of other factors such as speech rate (Harris & Kaisse 1999: 125). By contrast, a form like [] can never appear as *[] at any speech rate.


246 (391) a.[j]a 's/he caresses' a.fe[j].ta 's/he shaves' vic[j]a 's/he vitiates' o.d[j]a 's/he hates' re[j].na 's/he rules' pe[j].na 's/he combs' ba[j].la 's/he dances' ca.lum.n[j]a 's/he slanders' di.vorc[j]a 's/he divorces'[j]a 's/he envies' e.lo.g[j]a 's/he praises' ra.b[j]a 's/he rants' b. re.<h> 's/he misguides' 'it buzzes' 's/he empties' 's/he sprinkles'

Harris & Kaisse 1999 argue that their analysis, which uses lexically nuclear high vowels, is superior to others which utilize a glide/vowel distinction because it also provides a unified account of stress in Spanish. They show that in verbs, stress is consistently realized on the penultimate syllable in the present indicative and subjunctive (Harris & Kaisse 1999: 186). Three relevant examples are deposíta 's/he deposits', víc[j]a 's/he vitiates', and rocí.a 's/he sprinkles'. It is not possible to reduce this analysis to the penultimate vocoid since stress in 's/he vitiates' is on the antepenultimate vocoid. Lexical marking of the vocoid in 's/he sprinkled' explains why it receives stress; it surfaces as a syllable nucleus. 11.3.3 Korean Korean is another language in which some vowels are lexically marked as nuclear, thereby prohibiting their realization as non-syllabic glides. The Korean examples show that the lexically nuclear vowels are rare, occurring considerably less frequently then the unmarked vowels. These marked vowels surface as vowels where glides would normally be expected, illustrating that it is not a distinction between underlying glides vs. vowels, but a distinction in the vowels themselves. The general behavior of underlying vowels in Korean is to surface as nuclear in between two consonants (e.g. [ti-ta] 'lose') and as a glide intervocalically (e.g. [moj-t-ta] 'gathered') (Yun 2003). There is also a ban on coda glides/falling diphthongs,

247 forcing the output [moi-ta] *[moj-ta] 'gather' (Yun 2003). In the environment C__V there is variation between a glide and high vowel. When the underlying segment in C__V is a high front vowel, there are two patterns that emerge. In the unmarked pattern, a high front vowel /i/ surfaces as a glide as shown in (392a). In these forms, it is not possible for the vowel to surface as nuclear. In (392), /s*/163 is the past tense suffix and /ta/ is the declarative suffix. In accordance with coda neutralization in Korean, the /s*/ surfaces as [t]. (392) a. Unmarked /i/ (Yun 2003) h h [p i.ta] [phjt.ta] /p i-/ /ti-/ [ti.ta] [tjt.ta] /t*i-/ [t*i.ta] [t*jt.ta] b. Lexically nuclear /i=N/ /pi.-/ [pi.ta] /p*i.-/ [p*i.ta] h /p i.-/ [phi.ta] *[phi.t.ta] *[ti.t.ta] *[t*i.t.ta]

'straighten' 'lose' 'boil'

[pi.t.ta] [p*i.t.ta] [phi.t.ta]

*[pjt.ta] *[p*jt.ta] *[phjt.ta]

'be empty' 'sprain' 'bloom'

In the forms in (392b), however, the high vowel must surface as nuclear.164 Yun 2003 points out that the forms in (392b) are part of a rare pattern. Only six out of 428 forms with a stem-final high front vowel exhibit the pattern in (392b). An alternative analysis of (392) exists, where (392a) has an underlying glide /j/ and (392b) a normal underlying vowel /i/. The underlying representation of 'lose' would therefore be /tj/. This alternative is less ideal in that it predicts that the vowel /i/ only occurs in a few verb roots. It also predicts multiple verb roots that consist of a cluster and that lack a vowel. Furthermore, it predicts that other roots lacking a vowel should exist, but such forms are non-existent (Emily Curtis, p.c.). I conclude that the analysis of Yun 2003 illustrated in (392) is correct. The back round vowel /u/ differs from /i/ in that it is never lexically marked as nuclear. Again, this confirms the notion that the lexically nuclear vowels are (a) rare and (b) outside the basic pattern of glide/vowel realization. The round vowel /u/ further differs from /i/ in that both of the forms in (393) are possible. According to Yun 2003, however, the forms with the glide are preferred, while those with vowels sound distinctly old-fashioned.

The vowel in this suffix alternates as [] ~ [a] depending on the identity of the preceding vowel. [] surfaces after /i, , u, e, / while [a] surfaces after /o, a/. The /s*/ of this suffix is a fortis fricative (Yun 2003). 164 According to Yun 2003, the forms with an underlying nuclear /i=N/ undergo an optional rule of gliding that is dependent on speech rate.


248 (393) Unmarked /u/: Preferred /pusu-/ [pu.swt.ta] /pak*u/ [pa.k*wt.ta] /nanu-/ [na.nwt.ta] /taru-/ [ta.rwt.ta] Not preferred [] [pa.k*u.t.ta] [] []

'smashed' 'changed' 'divided' 'handled'

Finally, the mid vowel /o/ shows glide/vowel alternations as well as one case of a lexically nuclear segment. The vowel /o/ can alternate with a glide, as evidenced by the alternation [o-ta] 'come'/[w-at-ta] 'came' *[o-at-ta]. In the environment C___V, it has two realizations as shown in (394a). Whereas there was a preference for the glide form in (393), no such preference exists for the forms in (394a) (Yun 2003).165 There is one form that does not allow this alternation and it is given in (394b). In this example, it is not possible for /o/ to surface as a glide and it is analyzed as containing a lexically nuclear mid vowel (Yun 2003). Interestingly, the underlying nature of the surface glide is recoverable since the past tense suffix surfaces with [a] instead of []. (394) a. Unmarked /o/ /po-/ [pwat.ta] /s*o-/ [s*wat.ta] b. Lexically marked /o=N/ /t*o.-/ *[t*wat.ta] [] [s*] 'see' 'shoot'



Korean exhibits two sets of underlying vowels. The unmarked set exhibits a normal pattern of glide/vowel alternations depending on the environment. The second set exhibits an over-abundance of vowels. That is, members of this second set surface as vowels when a glide would otherwise be expected. Furthermore, it is clear that the second set represents the marked behavior since the forms that make it up are rare. 11.4 Summary Unlike the rest of this dissertation, this chapter has dealt with languages that do not have underlying glides. This set of languages can be divided into two groups, "normal" languages and "vocalic" languages. The "normal" languages have underlying glides that surface predictably as vowels or glides. Languages in this set can of course vary. Some might allow clusters with a surface glide, while others may prefer to realize the underlying vowel as a secondary articulation. Still others may choose to never allow


I hypothesize that the lack of preference with the mid vowel may be due to the fact that is mid and not high. I have not worked out the details of such an analysis.

249 gliding. In any of these cases, though, the realization of an underlying vowel is completely predictable. Some languages, in addition to the "normal" set of vowels, contain underlying vowels that must surface as syllabic. Their realization is not predictable from the environment, but from the segment itself. These segments are marked in the lexicon with additional structure, namely prelinking to the nuclear position. Since these segments are marked, they are expected to be rare or at least less numerous than the basic pattern exhibited by the normal high vowels. Indeed, this marginal status is confirmed by Spanish and Korean.


12 Conclusion

This dissertation provides a set of languages that have underlying glides. From the data presented here, we can put to rest the question of whether or not underlying glides exist as distinct segments. In a set of genetically and geographically diverse languages (Turkish, Pulaar, Yawelmani, Tahltan, Karuk, Sundanese, Cree, Pashto, and IT Berber), data from a variety of phenomena showed that underlying glides pattern as structurally and featurally different from underlying vowels. Furthermore, the range of cases of underlying glides provides no ambiguity; no other satisfactory solution exists for the data. In addition to providing definitive cases of underlying glides, this dissertation also provides a set of criteria that can be used to test whether other languages have underlying glides. While representations are likely to change in the face of new evidence, the types of processes that provide evidence for underlying glides will not change. We may find new criteria that can be added, but the evidence itself must still be explained. A further benefit of this dissertation is that it provides a large amount of evidence for underlying glides. A frequent shortcoming of typological studies is a lack of data, which makes challenging the conclusions difficult. This dissertation provides a sufficient set of examples for anyone interested in exploring this topic. 12.1 Typology In conducting the research for this dissertation, it quickly became clear that the complexity of glide/vowel alternations and underlying vocoids was significantly greater than what had previously been assumed. Not only can vocoid systems vary in which segments are underlying, they too can vary in how those segments are realized on the surface. There are not two types of languages, but eight. Of the eight types discussed in Chapter 1, Types V and VI were the most common of the languages that included underlying glides. Type V maintains a clear separation between underlying and surface vowels and glides. Type VI combines the common property of allowing underlying vowels to surface as either vowels or glides with an underlying glide that shows no such alternations. From the survey in this dissertation, it seems as though there is a dispreference for allowing underlying glides to surface with both a non-syllabic and syllabic parse. The only example of such a language was IT Berber. This language is unique in having a balanced pattern (Type VIII), where both underlying vowels and glides can surface as syllabic and non-syllabic. The typology that I outlined in Chapter 1 shows the types of vocoids that can exist underlyingly and how they can be realized in surface forms. I believe that the typology clarifies the complexity of the issue of underlying and derived glides by providing a

251 range of possible vocoid systems. Furthermore, the typology elucidates the complexity of this issue and explains why there has been confusion about possible vocoid systems. 12.2 Representations In this dissertation, I compared four possible representations of underlying glides. In order to evaluate the quality of a particular model, it is imperative to consider crosslinguistic evidence, as well as a variety of language-internal evidence. The four representations LEX, [CONS], VPT, and RAT were tested with the following feature-related data: backness harmony in Turkish; ATR-harmony and consonant gradation in Pulaar; vowel harmony in Yawelmani; consonant harmony in Tahltan; vowel harmonic suffixes, sonorant nasalization, and s-palatalization in Karuk; and nasal harmony in Sundanese. In the face of these processes, the two feature geometries, Vowel-Place Theory and Revised Articulator Theory, are significantly better equipped to deal with the range of data. Nonetheless, both models have a few difficulties. Vowel-Place Theory has difficulty dealing with two processes: consonant gradation in Pulaar and consonant harmony in Tahltan. The consonant gradation in Pulaar is difficult for VPT since it considers front vowels and front glides both to be coronal. Since the derived glide alternates with a velar, and not a coronal stop, this is a problem. Some might wonder if altering the representation of front vowels to be dorsal would fix this problem. This is not a satisfying solution because the reasons for using the same features to describe vowels and consonants are intertwined with the separation of place features on two levels in the feature tree. Getting rid of Unified Feature Theory by forcing all vowels to be dorsal in VPT eliminates the motivation for having two levels of place features. Therefore altering VPT so that vowels are dorsal vowels is not a realistic option. RAT does not have a problem with consonant gradation in Pulaar because it represents all vowels as dorsal and all place features on a single level. Since vowels are dorsal, so too are derived glides, and their alternation with velars (i.e. dorsal consonants) is explained. The second area where VPT runs into problems is with consonant harmony in Tahltan. In order to explain why all dentals do not pattern identically in the harmony, VPT must assume that some dentals are underspecified for place features (i.e. [d, t, t', n, n', dl, t, t', l, ]) while others are not (i.e. [dz, ts, ts', z, s]). RAT does not run into this problem because it rests on the markedness theory of Calabrese 1995. In this model, the dentals that participate in the harmony have contrastive coronal features and therefore interact with consonant harmony. Those consonants that do not participate are not contrastive for their coronal features. Though amending VPT to rely on Calabrese's markedness theory could work to explain harmony in Tahltan, it still leaves the problem of the representation of consonant gradation in Pulaar. Revised Articulator Theory has some difficulty representing s-palatalization in Karuk. The problem with RAT is that it cannot simultaneously account for the nontransparency of the low vowel /a/ and the transparency of the glide /w/. If it can be

252 shown that s-palatalization is a later process, even post-lexical, then the fact that /w/ surfaces as [] would explain the transparency of the glide. Since it does not actually have a dorsal articulation, it is of course transparent to s-palatalization just as other consonants are. VPT does not have this problem because the set of blocking segments is simply the segments with V-place features. I conclude that RAT is superior to VPT, though only marginally so. Both feature geometries account for the majority of the data. 12.3 Other languages In addition to the nine languages with underlying glides, I also briefly considered languages that do not have underlying glides. These languages behave differently from those with underlying glides in terms of their features, as well as their structure. In Karuk, for example, the high vowel /i/ has a predictable distribution on the surface, as either [i] or [j]. The surface glide [j] also patterns with vowels in hiatus resolution. In Sanskrit, the surface glide is also derived from an underlying vowel and is therefore transparent to n-retroflexion. Other languages do not exhibit this predictable distribution, but still show no evidence for underlying glides. In these cases the languages show evidence for lexically marked nuclear vowels which are forced to surface as syllabic.



Alderete, John (in press) On tone and length in Tahltan (Northern Athabaskan). In Athabaskan Prosody (Sharon Hargus and Keren Rice, eds.) Amsterdam: John Benjamins. Alderete, John and Tanya Bob (in press) A corpus-based approach to Tahltan stress. In Athabaskan Prosody (Sharon Hargus and Keren Rice, eds.) Amsterdam: John Benjamins. Anderson, Stephen R (1976a) On the description of consonant gradation in Fula. Journal of African Linguistics 7 (1). 93-136. Anderson, Stephen R (1976b) On the description of multiply-articulated consonants. Journal of Phonetics 4(1). 17-27. Archangeli, Diana B. (1984) Underspecification in Yawelmani phonology and morphology. PhD dissertation, MIT. Archangeli, Diana (1988) Aspects of underspecification theory. Phonology 5, 183-208. Archangeli, Diana (1991) Syllabification and prosodic templates in Yawelmani. Natural Language and Linguistic Theory 9. 231-283. Bakovic, Eric (2003) Vowel harmony and stem identity. ROA 540-0902. Barker, Chris (1989) Extrametricality, the Cycle, and Turkish word stress. Phonology at Santa Cruz. 1-34. Bright, William (1957) The Karok language. Berkeley and Los Angeles: University of California Press. Bell, Alan and Mohamad M. Saka (1983) Reversed sonority in Pashto initial clusters. Journal of Phonetics 11. 259-75. Bob, Tanya Marie (1999) Laryngeal phenomena in Tahltan. MA Thesis, University of British Columbia. Boyce, Suzanne E. (1990) Coarticulatory organization for lip rounding in Turkish and English. JASA 88 (6), 2584-2595. Browman, Cathrine Pl & Louis Goldstein (1989) Articulatory gestures as phonological units. Phonology 6. 201-252. Burrow, T. (1955) The Sanskrit language. (Reprinted 1959, 2nd edition, Great Britain: R. MacLehose and Company Limited). Cardona, George (1987) Sanskrit. In The world's major languages (Bernard Comrie, ed.). New York: Oxford University Press. 448-469. Cardona, George (2003) Sanskrit. In The Indo-Aryan languages (George Cardona and Dhanesh Jain, eds.) London & New York: Routledge. 104-160. Catford, J. C. (1977) Fundamental problems in phonetics. Bloomington: Indiana University Press. Calabrese, Andrea (1995) A constraint-based theory of phonological markedness and simplification procedures. Linguistic Inquiry 26 (3). 373-463. Chitoran, Ioana (1997) The phonology and morphology of Romanian glides and diphthongs: a constraint-based approach. PhD dissertation, Cornell University.

254 Chomsky, Noam and Morris Halle (1968) The sound pattern of English. Cambridge, MA: MIT Press. Clements, G. N. (1989) Place of Articulation in Consonants and Vowels: a Unified Theory. Unpublished Ms, Cornell. Clements, G. N. (1997) Berber syllabification: Derivations or constraints? In Derivations and constraints in phonology (Iggy Rocca, ed.) Oxford: Clarendon Press. 289330. Clements, G. N. and Elizabeth V. Hume (1995) The Internal Organization of Speech Sounds. In The Handbook of Phonological Theory (John A. Goldsmith, ed.) 245306. Clements, George N. and Samuel Jay Keyser (1983) CV Phonology: A generative theory of the syllable. Cambridge, MA: MIT Press. Clements, G. N. and Engin Sezer (1982) Vowel and consonant disharmony in Turkish. In The Structure of Phonological Representations, pt. 2 (H. van der Hulst and N. Smith, eds.) 213-55. Cohn, Abigail C. (1993) The status of nasalized continuants. In Phonetics and phonology Volume 5: Nasals, nasalization, and the velum (Marie K. Huffman and Rena A. Krakow, eds.) 329-367. Cohn, Abigail C. (1992) The consequences of dissimilation in Sundanese. Phonology 9. 199-220. Cohn, Abigail C. (1990) Phonetic and phonological rules of nasalization. PhD dissertation, UCLA. Cole, Jennifer and Charles Kisseberth (1997) Restricting Multi-level constraint evaluation: Opaque rule interaction in Yawelmani Vowel Harmony. Proceedings of the Arizona Phonology Conf 18-38 (also ROA 98-0000) Comrie, Bernard (1997) Turkish Phonology. In Phonologies of Asia and Africa II (Alan S. Kaye, ed.) 883-98. Deligiorgis, Ioanna (1988) Glides and syllables. PhD dissertation, University of Iowa. Deligiorgis, Ioanna (1993) The status of [consonantal]. Cahiers de linguistique théorique et appliquée 30. 3-15. Dell, François and Mohamed Elmedlaoui (1985) Syllabic consonants and syllabification in Imdlawn Tashlhiyt Berber. Journal of African Languages and Linguistics 7. 105-130. Dell, François and Mohamed Elmedlaoui (1988) Syllabic consonants in Berber: Some new evidence. Journal of African Languages and Linguistics 10. 1-17. Dell, François and Mohamed Elmedlaoui (1989) Clitic ordering, Morphology and phonology in the verbal complex of Imdlawn Tashlhiyt Berber: Part I. Langues orientales anciennes philology et linguistique 2. 165-194. Dell, François and Mohamed Elmedlaoui (1991) Clitic ordering, Morphology and phonology in the verbal complex of Imdlawn Tashlhiyt Berber: Part II. Langues orientales anciennes philology et linguistique 3. 77-104

255 Dell, François and Mohamed Elmedlaoui (1996a) On consonant releases in Imdlawn Tashlhiyt Berber. Linguistics 34. 357-395. Dell, François and Mohamed Elmedlaoui (1996b) Nonsyllabic transitional vocoids in Imdlawn Tashlhiyt Berber. In Current trends in phonology: Models and methods (Jacques Durand and Bernard Laks, eds.) 217-244. Diop, Abdul Aziz (1993) Vowel deletion in Pulaar: Rime and nuclear mergers and the issue of the syntax-phonology interface. Studies in the Linguistic Sciences 23(2). 33-71. Diop, Abdul Aziz Maalik (1999) Aspects of Pulaar non-linear phonology. PhD dissertation, University of Illinois at Urbana-Champaign. Donohue, Mark (1999) A grammar of Tukang Besi. Berlin & New York: Mouton de Gruyter. Durand, Jacques (1987) On the phonological status of glides: the evidence from Malay. In Explorations in Dependency Phonology (John Anderson and Jacques Durand, eds.) Dordrecht, Holland: Foris Publications. 79-107. Elfenbein, Josef (1997) Pashto phonology. In Phonologies of Asia and Africa (Alan S. Kaye, ed.) Winona Lake, Indiana: Eisenbrauns. 733-60 Ellis, C. Douglas (1983) Spoken Cree. Edmonton, Alberta: The Pica Pica Press. Elmedlaoui, Mohamed (1985) Le parler berbère chleuh d'Imdlawn (Maroc): segments et syllabation. PhD dissertation, Université de Paris VIII. Elmedlaoui, Mohamed (1988) De la gémination. Langues orientales anciennes philology et linguistique 1. 117-156 Emeneau, M. B. & B. A. van Nooten (1968) Sanskrit sandhi and exercises. Berkeley and Los Angeles: University of California Press. Flemming, Edward (2001) Scalar and categorical phenomena in a unified model of phonetics and phonology. Phonology 18. 7-44. Flemming, Edward (2002) Auditory representations in phonology. New York and London: Garland Publishing, Inc. Goldman, Robert P. and Sally J. Sutherland Goldman (2002) An introduction to the Sanskrit language (3rd edition). Berkeley: Center for South Asia Studies. Goldsmith, John A. (1979) Autosegmental Phonology. New York: Garland Publishing. Guerssel, Mohamed (1986) Glides in Berber and syllabicity. Linguistic Inquiry 17 (1). 112. Grammont, M. (1933) Traite de phonetique. Paris: Delagrave. Hall, Tracy Alan (1992) Syllable structure and syllable-related processes in German. Tübingen: Max Niemeyer Verlag. Halle, Morris (1995) Feature Geometry and Feature Spreading. Linguistic Inquiry 26 (1). 1-46. Halle, Morris, Bert Vaux, and Andrew Wolfe (2000) On feature spreading and the representation of Place of Articulation. Linguistic Inquiry 31 (3). 387-444. Hansson, Gunnar Ólafur (1999) Opacity and the loss of phonological alternations: vowel harmony in present-day Yowlumne. Ms, University of California-Berkeley.

256 Hansson, Gunnar Ólafur (2001) Theoretical and typological issues in consonant harmony. PhD dissertation, University of California, Berkeley. Hardwick, Margaret F. (1984) Tahltan Morphology and Phonology. MA Thesis, University of Toronto. Harris, James W. and Ellen M. Kaisse (1999) Palatal vowels, glides and obstruents in Argentinian Spanish. Phonology 16. 117-190. Hayes, Bruce (1989) Compensatory lengthening in moraic phonology. Linguistic Inquiry 20 (2). 253-306. Hayes, Bruce (1995) Metrical stress theory: Principles and case studies. Chicago: The University of Chicago Press. Herman, Rebecca (1994) "La double vie de w" or The status of [w] in Karuk. Studies in the Linguistic Sciences 24 (1/2). 233-244. Hualde, José Ignacio (1988) Affricates are not contour segments. In Proceedings of the Sevenths West Coast Conference on Formal Linguistics (Hagit Borer ed.) Stanford: The Center for the Study of Language and Information. 143-157. Hualde, José Ignacio (1989) Autosegmental and metrical spreading in the vowelharmony systems of northwestern Spain. Linguistics 27, 773-805. Hulst, Harry van der and Jeroen van de Weijer (1995) Vowel Harmony. In The Handbook of Phonological Theory (John A. Goldsmith, ed.) 495-534. Hume, Elizabeth (1994) Front vowels, coronal consonants, and their interaction in nonlinear phonology. New York and London: Garland Publishing, Inc. Hume, Elizabeth (1995) Representing the duality of glides. Les actes du congrès: langues et grammaire 1. Hume, Elizabeth and David Odden (1996) Reconsidering [consonantal]. Phonology 13. 345-376. Hyman, Larry M. (1985) A theory of phonological weight. Dordrecht, Holland: Foris Publications. Hyman, Larry M. (1975) Phonology: theory and analysis. Fort Worth, TX: Holt, Rinehart, & Winston. Inkelas, Sharon (1999) Exceptional stress-attracting suffixes in Turkish: representation versus the grammar. In The prosody-morphology interface (René Kager, Harry van der Hulst, and Wim Zonneveld, eds.) Cambridge: Cambridge University Press. 134-187. Jakobson, Roman, C. Gunnar M. Fant, and Morris Halle (1952) Preliminaries to speech analysis: the distinctive features and their correlates. Cambridge: MIT Jones, Daniel (1918) An outline of English phonetics. (Reprinted 1956 (8th edition) New York: E.P. Dutton and Co., Inc) Kabak, Bari and Irene Vogel (2001) The phonological word and stress assignment in Turkish. Phonology 18, 315-360. Kaisse, Ellen M. (1985) Some theoretical consequences of stress in Turkish. CLS 21. 199-209 Kaisse, Ellen M. (1992) Can [consonantal] spread? Language 68(2), 313-32.

257 Kaisse, Ellen M. (2002) Laterals are [-continunat]. Ms, University of Washington. Kaisse, Ellen M. and Susannah V. Levi (2004) Vowel harmony: nucleus to nucleus or vocalic node to vocalic node? Talk presented at the annual meeting of the LSA. Boston. Kaun, Abigail (1995) The typology of rounding harmony: An Optimality Theoretic Approach. PhD dissertation, UCLA. Kaye, Jonathan D. and Jean Lowenstamm (1984) De la syllabicite. In Forme sonore du langage: structure des representations en phonologie (F. Dell, D. Hirst, & J.-R. Vergnaud, eds.) 123-159. Kisseberth, Charles W. (1969) Theoretical Implications of Yawelmani Phonology. PhD dissertation, University of Illinois. Klatt, Dennis H. (1987) Review of text-to-speech conversion for English. Journal of the Acoustical Society of America 82. 737-793. Kornfilt, Jaklin (1997) Turkish. New York: Routledge. Krauss, Michael and Jeff Leer (1981) Athabaskan, Eyak, and Tlingit Sonorants. Fairbanks: Alaska Native Language Center. Labov, William (1994) Principles of linguistic change: internal factors. Oxford, UK and Cambridge, USA: Blackwell. Ladefoged, Peter and Ian Maddieson (1996) The sounds of the worlds languages. Cambridge, MA: Blackwell Publishers Latta, Frank F. (1977) Handbook of Yokuts Indians, 2nd ed. Santa Cruz: Bear State Books. Leer, Jeff (1987) Navajo and Comparative Athapaskan. In The Navajo Language: A grammar and colloquial dictionary (Robert Young and William Morgan, eds.) Albuquerque: University of New Mexico Press. 264-301. Lees, R. B. (1961) The Phonology of Modern Standard Turkish. Bloomington: Indiana University Press. Levi, Susannah V. (2000) Glides, laterals, and Turkish Vowel Harmony. MA Thesis, University of Washington. Levi, Susannah V. (2001) Glides, laterals, and Turkish Vowel Harmony. CLS 37. 379394. Levi, Susannah V. (2004) Underlying vowels vs. glides: typology and case studies. Talk presented at the third North American Phonology Conference. Montreal, Canada. Levin, Juliette (1985) A metrical theory of syllabicity. PhD dissertation, MIT. Lewis, G. L. (1967) Turkish Grammar. Oxford: Oxford University Press. (Reprinted 1975). Liberman, Alvin, Pierre Delattre, Louis Gerstman, Franklin Cooper (1956) Tempo of frequency change as a cue for distinguishing classes of speech sounds. Journal of experimental psychology 52 (2). 127-137. Lombardi, Linda (1990) The nonlinear organization of the affricate. Natural Language and Linguistic Theory 8. 375-425. Maddieson, Ian (1984) Patterns of sounds. Cambridge: Cambridge University Press.

258 Maddieson, Ian and Karen Emmorey (1985) Relationship between semivowels and vowels: cross linguistic investigations of acoustic difference and coarticulation. Phonetica 42(4), 163-174. McCarthy, John J. (2002) A thematic guide to Optimality Theory. Cambridge: Cambridge University Press. Mester, Armin and Jaye Padgett (1994) Directional syllabification in Generalized Alignment. Phonology at Santa Cruz, Vol. 3. 79-85 (and on ROA). Mielke, Jeff (2004) What ambivalent segments can tell us about the universality of distinctive features. Talk presented at the 78th Annual Meeting of the Linguistic Society of America, Boston, MA. Nater, Hank F. (1989) Some comments on the phonology of Tahltan. International Journal of American Linguistics 55 (1). 25-42. Navarro Tomás, Tomás (1965) Manual de pronunciación española. Consejo Superior de Investigaciones Científicas. Newman, Stanley (1944) Yokuts Language of California. Publications in Anthropology, 2. New York: The Viking Fund, Inc. Niang, Mamadou Ousmane (1997) Constraints on Pulaar phonology. Lanham, MD: University Press of America, Inc. O'Connor, J.D., L.J. Gerstman, A. M. Liberman, P.C. Delattre and F.S. Cooper (1957) Acoustic cues for the perception of initial /w, j, r, l/ in English. Word 13 (1). 2443. Odden, David (1992) Kimatuumbi phonology and morphology. Ms, Ohio State University. Paradis, Carole (1987a) Strata and syllable dependencies in Fula: The Nominal Classes. Journal of African Languages and Linguistics 9. 123-39. Paradis, Carole (1987b) Glide alternations in Pulaar (Fula) and the Theory of Charm and Government. In Current Approaches to African Linguistics 4 (David Odden, ed.) Dordrecht, Holland: Foris Publications. 327-338. Paradis, Carole (1988) Towards a theory of constraint violations. McGill Working Papers in Linguistics 5.1. 1-43 Paradis, Carole (1992) Lexical Phonology and Morphology: The nominal classes in Fula. New York: Garland Publishing, Inc. Paradis, Carole and Jean-François Prunet (1989a) On coronal transparency. Phonology 6. 317-348. Paradis, Carole and Jean-François Prunet (1989b) Markedness and coronal structure. In Proceedings of NELS 19 (Juli Carter and Rose-Marie Dechaine, eds.) 330-344. Parkinson, Frederick B. (1996) The representation of vowel height in phonology. PhD dissertation, Ohio State University. Penny, Ralph J. (1969a) El habla pasiega: ensayo de dialectología montañesa. London: Tamesis Books Ltd. Penny, Ralph J. (1969b) Vowel harmony in the speech of Montes de Pas (Santander). Orbis 18, 148-166.

259 Penzl, Herbert (1965) Review of "A short grammatical outline of Pashto" by D. A. Shafeev, translated and edited by Herbert H. Paper. Language 41 (3). 529-534. Penzl, Herbert (1955) A grammar of Pashto. Washington, D.C.: American Council of Learned Societies. Pike, Kenneth L. (1943) Phonetics: A critical analysis of phonetic theory and a technic for the practical description of sounds. U. of Michigan Press: Ann Arbor. Prince, Alan and Paul Smolensky (1993) Optimality Theory: constraint interaction in generative grammar. Technical Report 2 of the Rutgers Center fro Cognitive Science. Rice, Keren (1997) A reexamination of Proto-Athabaskan *y. Anthropological Linguistics 39 (3). 423-436. Rijkhoff, Jan & Dik Bakker (1998) Language sampling. Linguistic Typology, 2. 263-314. Robins, R. H. (1957) Vowel Nasality in Sundanese, a phonological and grammatical study. In Studies in Linguistics (special volume of the Philological Society). Oxford: Basil Blackwell. 87-103. Roca, Iggy (1997) There are no "glides", at least in Spanish: An optimality account. Probus 9. 233-265. Rosenthall, Samuel (1994) Vowel/glide alternation in a theory of constraint interaction. PhD dissertation, University of Massachusetts-Amherst. Rosenthall, Samuel (1997) Metrical influences on syllable structure in Lenakel. Linguistic Inquiry 28 (1). 187-193. Rubach, Jerzy (2000) Backness switch in Russian. Phonology 17, 39-64. Sagey, Elizabeth C. (1986) The representation of features and relations in non-linear phonology. PhD dissertation, MIT. Schein, Barry and Donca Steriade (1986) On geminates. Linguistic Inquiry 17 (4). 691744. Schourup, Lawrence (1972) Characteristics of vowel nasalization. Papers in Linguistics 5, 530-548. Sezer, Engin (1983) On non-final stress in Turkish. Journal of Turkish Studies 5. 61-69. Shafeev, D. A. (1964) A short grammatical outline of Pashto (translated by Herbert H. Paper Bloomington, Indiana: Mouton & Co.) Shaw, Patricia A. (1991) Consonant harmony systems: the special status of coronal harmony. In Phonetics and Phonology, Volume 2, The special status of coronals (C. Paradis and J-F Prunet, eds.) 125-157. Steriade, Donca (1984) Glides and vowels in Romanian. In Proceedings of the 10th Annual Meeting of the Berkeley Linguistics Society (Claudia Brugman and Monica Macaulay, eds.) 47-64. Steriade, Donca (1986) Yokuts and the vowel plane. Linguistic Inquiry 17 (1). 129-146. Steriade, Donca (1993) Closure, release, and nasal contours. In Phonetics and phonology Volume 5: Nasals, nasalization, and the velum (Marie K. Huffman and Rena A. Krakow eds.) San Diego: Academic Press, Inc. 401-470.

260 Steriade, Donca (1994) Complex onsets as single segments: The Mazateco pattern. In Perspectives in phonology (Jennifer Cole and Charles Kisseberth, eds.) Stanford, CA: CSLI Publications. 203-291. Steriade, Donca (1995) Underspecification and Markedness. In The Handbook of Phonological Theory (John A. Goldsmith, ed.) 114-174. Stetson, R. H. (1928) Motor Phonetics: a study of speech movements in action (Reprinted 1951, Amsterdam: North-Holland Publishing Co.) Stevens, Alan M. (1968) Madurese Phonology and Morphology. New Haven, Connecticut: American Oriental Society. Straka, Georges (1964) A propos de la question des semi-voyelles. Zeitschrift fuer Phonetik Sprachwissenschaft und Kommunikationsforschung 17. 301-23. Sussman, H.M., H. A. McCaffrey, & S. A. Matthews (1991) An investigation of locus equations as a source of relational invariance for stop place categorization. Journal of the Acoustical Society of America 90. 1309-1325. Sweet, Henry (1907) The sounds of English: an introduction to phonetics (2nd Edition 1910, Oxford: The Clarendon press) Sweet, Henry (1890) A primer of phonetics. (3rd edition 1906, Oxford: The Clarendon Press) Sweet, Henry (1877) A handbook of phonetics (Reprinted 1970, College Park, MD: McGrath Publishing Co) Troubetzkoy, N. D. (1939) Principes de phonologie. (2nd Edition 1957, J Cantineau, tr., Paris: Librairie C. Klinkcksieck.) Uhlenbeck, C. C. (1898) A manual of Sanskrit phonetics. (Reprinted 1960, Delhi: Jayyad Press) Underhill, Robert (1976) Turkish Grammar. (Reprinted 1997, Cambridge: MIT Press) Underhill, Robert (1986) Turkish. In Studies in Turkish Linguistics (Dan Slobin and Karl Zimmer, eds.) 7-21. Waksler, Rachelle (1990) A formal account of glide/vowel alternation in prosodic theory. PhD dissertation, Harvard. Walker, Rachel Leah (1998) Nasalization, neutral segments, and opacity effects. PhD dissertation, University of California Santa Cruz. Walker, Rachel (2001) Round licensing, harmony, and bisyllabic triggers in Altaic. NLLT 19, 827-878. Walker, Rachel (2003) Reinterpreting transparency in Nasal Harmony. In The phonological spectrum, Part I: Segmental structure (Jeroen van de Weijer, Vincent van Heuven, and Harry van der Hulst, eds.) (Current Issues in Linguistic Theory No. 233) Amsterdam: John Benjamins. 37-72. Walker, Rachel and Geoffrey K. Pullum (1999) Possible and impossible segments. Language 75 (4). 764-780. Whitney, William D. (1865) On the relation of vowels and consonants. Journal of the American Oriental Society 8. 357-373. (Reprinted in William D. Whitney, Oriental and linguistic studies, 277-300. SRI Satguru Publications: Delhi, 1987)

261 Whitney, William Dwight (1879) Sanskrit Grammar. (Reprinted 1913, 4th Edition, Boston: Ginn & Company.) Whitney, William Dwight (1885) The roots, verb-forms, and primary derivatives of the Sanskrit language. (Reprinted 1963, Delhi: Motilal Banarsidass). Yavas, Mehmet (1980) Vowel and Consonant harmony in Turkish. Glossa 14. 189-211. Yun, Yungdo (2003) Korean vocoids lexically marked and unmarked for syllabicity. University of Washington Working Papers in Linguistics 22. 101-120. Zimmer, Karl and Orhan Orgun (1992) Turkish. JIPA 22. 43-45. Zoll, Cheryl (1993) Directionless Syllabification and ghosts in Yawelmani. ROA 280000.



Susannah Levi was born in St. Louis, Missouri. Her undergraduate degree is in Mathematics and French from Washington University in St. Louis. She received her Master of Arts in Linguistics at the University of Washington in 2000. In 2004 she earned her Doctor of Philosophy degree in Linguistics from the University of Washington.


Microsoft Word - Levi_diss_shortversion.doc

282 pages

Report File (DMCA)

Our content is added by our users. We aim to remove reported files within 1 working day. Please use this link to notify us:

Report this file as copyright or inappropriate


Notice: fwrite(): send of 212 bytes failed with errno=104 Connection reset by peer in /home/ on line 531