Phonetics and Phonology in Unstressed Vowel Reduction

Phonetics and Phonology in Unstressed Vowel Reduction: Russian and Beyond I. Introduction • Phonological vowel reduction (hereafter VR): neutralization of vowel quality contrasts in unstressed syllables. • Figures in recent theoretical literature (e.g. Barnes 2002, Beckman 1998, Crosswhite 2001, Flemming 2oo1a and b, Steriade 1994) as an example of a pattern of Positional Neutralization with relatively clear phonetic underpinnings. • Understanding the relationship between these phonetic underpinnings and the phonological grammar in patterns such as VR is a central task of current work on the phonetics-phonology interface. 2. Russian Vowel Reduction 2.1. Facts • Contemporary Standard Russian: traditionally claimed to display two distinct patterns or degrees of VR, one for the syllable immediately preceding the tonic (here, first pretonic), and one for all other unstressed syllables. • Five vowels contrast in stressed syllables1. • Mid vowels do not surface in unstressed syllables. (1) Stressed vowel inventory i e o u 1 pretonic s ⇒ i a ⇒ u 2 pretonic, posttonic s i  u a • Focus of this talk: /a/ vs. /o/ • /a/ and /o/ contrast in stressed syllables, but this contrast is neutralized in unstressed syllables as follows: (2) a. /a/, /o/ ⇒ [a]/[] in first pretonic syllables, a.k.a. Degree 1 reduction b. /a/, /o/ ⇒ [] in other unstressed syllables, a.k.a. Degree 2 reduction -1- (3) VR in Contemporary Standard Russian /o/ moldst bol youth pain DEGREE 1 REDUCTION DEGREE 2 REDUCTION malodnkj young (dim.) mladoj young balet to hurt blvoj pain (adj.) /a/ starj razum old (adj.) reason starik razumn STRESSED • old man strna wisely (adv.) rzumet old times to understand Several exceptions to the pattern described above have been recognized: EXCEPTIONAL CONTEXT /odnomu/ /obiajet/ #1: No reduction to schwa in absolute word-initial position [adnamu] [abiajt] EXCEPTION CONTEXT 'one' (dat.sg.) *[dnamu] 'insult, abuse' (3sg.) *[biajt] #2: No reduction to schwa in hiatus before [a] /odnoobraznj/ -> [adnaabraznj] /sootnoenije/ -> [saatnaenije] 'monotonous' ‘relationship’ *[adnabraznj] *[satnaenije] #3: No reduction to schwa in absolute phrase-final syllables (Matusevic 1976: 102, Zlatoustova 1962: 109-139) EXCEPTIONAL CONTEXT 2.2. Previous approaches • Two recent OT accounts: a. Alderete (1995): Faithfulness constraints parameterized to the head foot (i.e. the stressed syllable and first pretonic) rank higher than markedness constraints against [a] in unstressed syllables, which in turn outrank general Faithfulness constraints. b. Crosswhite (2001): Vowels outside the head foot cannot bear moras, but [a] cannot be non-moraic, and so is realized as schwa. • Both accounts assume high-ranked constraints banning mid vowels from unstressed syllables. • Both these approaches must present essentially diacritic solutions to the exceptional patterns described above. EXAMPLE: Crosswhite (2001): Exception #1: Align-µ − The left edge of every word must align with some mora. Schwa is non-moraic. Consonant initial words violate this constraint to avoid creating moraic onsets. -2- Exception #2: *Hiatus([a], []) – The featureless vowel schwa may not occur in hiatus before [a]. Exception #3: Not treated. Align-mora/phrase? CRUCIAL POINT: All previous accounts of VR in Russian treat both Degree 1 and Degree 2 reduction patterns as phonological, i.e. as changing phonological category membership through manipulation of distinctive feature specifications. 3. Vowel Reduction: Phonetics and phonology 3.1. Phonetic Underpinnings of VR • Phonological VR patterns have been claimed to have their source in the reduced durations of vowels in unstressed syllables. • Languages with phonological vowel reduction are almost invariably those languages with a stress accent strongly and reliably cued by a difference in vowel duration between stressed and unstressed syllables (Lehiste 1970)2. (4) Stress accent and vowel reduction in Slavic Russian Belarussian Bulgarian Macedonian Serbo-Croatian Czech Polish DURATION-CUED STRESS Yes Yes Yes No No No No PHONOLOGICAL VR Yes Yes Yes No No No No THE CONNECTION: • Shortening of vowels in unstressed syllables leads to undershoot of more open target articulations of non-high vowels (à la Lindblom 1963). • Raising of these creates a compressed vowel space in which contrasts are likely to be misperceived and to collapse over time (See Barnes 2002, Flemming 2001a for discussion). • Accounts for a strong, if not exceptionless, typological generalization: in VR systems, vowel height contrasts are the first to go3. When this happens, the vowels to go are almost invariably the mid and low vowels. High vowels stay put. 2 Exceptions, of course, exist: Shimakonde (Bantu), for example, removes mid vowels from unstressed syllables even when they are bimoraic (and surface as such). It cannot be accidental that this system would have developed out of a system already implementing vowel height harmony. 3 The other two contrast types commonly targeted in stress-based positional neutralization systems are length and nasalization, both negatively impacted by decreased duration (see Barnes 2002 for details)s. -3- Common Stressed-Unstressed Mapping (5) i e u o i ⇒  u a (6) Non-occurring Stressed-Unstressed Mapping i e u o e ⇒ a • a o VR systems specifically targeting other commonly-neutralized dimensions of contrast for vowels, such as palatality, rounding, or ATR, are spottily attested at best. No vertical vowel systems in unstressed syllables only. No VR analogue of Turkish palatal harmony4. (7) Non-occurring Stressed-Unstressed Mapping i e u o   a ⇒ a 3.2. VR in the Grammar: Three Approaches (8) Positional Faithfulness and Positional Markedness (Alderete 1995, Beckman 1996, 1998, Crosswhite 2001, Zoll 1997, inter alia) • Phonological strong licensing results from Faithfulness or Markedness constraints parametrized to refer to the certain positions. a b • (9) • >> Ident[hi]/σ@ *MidV/unstressed σ >> *MidV Ident[hi] >> >> Ident[hi] *Mid PROBLEM: In principle, any feature or marked structure may be paired with any strong or weak position to derive attested patterns. Licensing-by-Cue and Direct Phonetics (Steriade 1997 et seq., Flemming 1995, 1997, 2001, to appear, Kirchner 1998, Zhang 2001) Phonetic cues, not arbitrarily listed positions, license the realization of contrasts EXAMPLE: Vowel contrasts are licensed in stressed syllables not by stressedness per se, but by the additional duration accorded vowels realized there. Durational pressures in unstressed syllables force neutralization. PAYOFF: Accounts for typology given above. 4 Though see below on Sosva Mansi. -4- POTENTIAL PROBLEMS: • Predicts necessary co-presence of PN patterns and phonetic patterns giving rise to them. • Must assign "unnatural" or synchronically arbitrary patterns a radically separate grammatical implementation. (10) Grounding Conditions (Archangeli and Pulleyblank 1994, Hayes 1997, Smith 2002) • Smith 2002: Schema/Filter model of CON; position-specific markedness and faithfulness constraints freely constructed from constraint schemas (IDENT , ALIGN , C/str) and arguments. • Results submitted to "substantive filters" which remove phonetically or psycholinguistically problematic constraints from consideration. • Accounts for the typology of positional neutralization patterns without allowing the phonology direct access to phonetic information. 4. One experiment, begetting another... • In order to determine more precisely the role of phonetic duration in the synchronic VR system of Russian, the following experiment was designed. 4.1. Methodology • Experiment #1 was designed to determine what, if any, active role phonetic duration plays in the realization of vowel reduction. • Test words: 72 Russian trisyllables 12 stressed /a/ 12 1 pretonic /a/ 12 2 pretonic /a/ 12 stressed /o/ 12 1 pretonic /o/ 12 2 pretonic /o/ • All target vowels were located in open syllables followed by unpalatalized voiced or voiceless stops and preceded by unpalatalized voiced stops, voiceless stops, or laterals. • Tokens were embedded in frame sentences of the form Mashka X skazala, 'Mashka said X'. Each token appeared twice for a total of 144 sentences. • Sentences were randomized and arranged in blocks of 12. • Subjects, one male from St. Petersburg, one female from Ufa, student-age native speakers of Russian, were instructed to read through each block of sentences once slowly and then once again as quickly as possible to achieve maximum variation in vowel durations. • Sessions were recorded directly to CD at a sampling rate of 44.1 KHz. Duration and LPC formant measurements were taken using Praat 4.1.11 speech analysis software (Copyright@1992-2003 by Paul Boersma and David Weenink). -5- 4.2. Results 4.2.1. Mean values for duration and F1 (11) Speaker 1 5 Duration F1 (12) Speaker 2 Duration F1 STRESSED /a/ 86 ms. 722 Hz /o/ 73 ms 466 1 PRETONIC 84 ms. 553 Hz 2 PRETONIC 41 ms 462 Hz 1 PRETONIC /a, o/ 68 ms. 726 Hz 2 PRETONIC /a, o/ 25 ms. 492 Hz • Neither speaker had any significant difference between unstressed /a/ and /o/ in any position. This contrast is completely neutralized. • Speaker 1 shows a mean F1 for 1 pretonic /a, o/ dramatically lower than a typical value for stressed /a/. Reflects traditional transcription of the former as [] (e.g. Matusevic¬ 1972, Avanesov 1968). Probably more like [], see Barnes 2002, chapter 2, for details. • For Speaker 2, no significant difference was found between F1 values for 1 pretonic /a, o/ and stressed /a/. 4.2.2. Vowel height as a function of phonetic duration The following displays the results of linear regression analyses of the relationship between vowel duration and vowel height for Speakers 1 and 2 in Experiment 1. • (13) Speaker 1 a. R (92) = .179, p > .05 b. R (60) = .602, p <.0001 First Pretonic /a/ and /o/ Second Pretonic /a/ and /o/ 800 600 700 500 400 500 Series1 400 Linear (Series1) 300 R 2 = 0.0145 F1 in Hz F1 in Hz 600 Series1 300 Linear (Series1) 200 R 2 = 0.3981 200 100 100 0 0 0 20 40 60 80 100 120 0 Duration in ms. 20 40 60 80 Duration in ms. 5 Speaker 1 in fact read a somewhat earlier version of the same experiment containing fewer words and not testing vowels in stressed syllables. -6- Speaker 1: unstressed /a/ and /o/ 800 700 F1 in Hz 600 500 Series1 400 Series2 300 200 100 0 0 20 40 60 80 100 120 Duration in ms. (14) Speaker 2 a. R (146) = .142, p > .05 b. R (78) = .644, p <.0001 Second Pretonic /a/ and /o/ First Pretonic /a/ and /o/ 700 1000 900 600 800 500 Series1 500 Linear (Series1) 400 F1 in Hz 600 2 R = 0.0297 300 400 Series1 300 Linear (Series1) R2 = 0.4204 200 200 100 100 0 0 0 20 40 60 80 0 100 10 20 30 40 Duration in ms. Duration in ms. Speaker 2: unstressed /a/ and /o/ 1000 900 800 700 F1 in Hz F1 in Hz 700 600 Series1 500 Series2 400 300 200 100 0 0 10 20 30 40 50 60 Duration in ms. -7- 70 80 90 100 50 4.3. Discussion • Neutralization of /a/ and /o/ in unstressed syllables is categorical. No significant difference between these underlying categories is preserved. This is true regardless of position relative to stress. • Reduction to schwa in Degree 2 reduction environments has a clear gradient character. Vowel height is strongly dependent on vowel duration. A cline of F1 values is recorded between more [a]-like values on longer vowels and more schwa-like values on shorter vowels. • Categorical vs. gradient application is argued to be a defining distinction between symbolic phonological processes and quantitative phonetic processes (see e.g. Keating 1988c, Cohn 1990, Zsiga 1993, Myers 2000, Barnes 2002, inter alia) • Hypothesis: given the duration-dependent cline of F1 values produced in Degree 2 reduction contexts, there is no reason to think the phonology has any role in (or access to) reduction of unstressed /a, o/ to schwa. The merger of /a/ and /o/ in unstressed syllables, on the other hand, is clearly categorical and phonological. 5. Experiment #2: Hyperarticulation • F1 values for Degree 2 reduction in Experiment 1 approach, but fail to clearly attain canonical [a]-like qualities. Experiment 2 seeks to draw out this result more clearly. 5.1. Methodology • Test words were 20 trisyllables of Russian, 10 with /a/, and 10 with /o/ in open second pretonic syllables. All target vowels were followed by unpalatalized voiced or voiceless stops and preceded by unpalatalized voiced stops, voiceless stops, or laterals. • Target words were framed in sentences of the form Po-moemu on X skazal, 'It seems to me he said X'. • Speakers were four student-age native speakers of Russian: three females raised in Ufa, Saint Petersburg, and Moscow respectively, and one male, from Moscow. • Sentences were embedded in a longer list of similar sentences and randomized. • Following an experimental paradigm used in, e.g. Johnson, Flemming and Wright 1993, subjects read each sentence, and then were asked twice to repeat themselves by an experimenter simulating incomprehension (speakers were informed this would occur before the experiment began): (15) Sample dialogue Sp: Po-moemu on X skazal. JB: Ah? Ne ponjal. Sp: X on skazal! JB: Sµtoetoonskazal??!! Sp:X! 'It seems to me he said X' 'Huh? I didn't get that.' 'He said X!' 'He said what??!!' 'X!' -8- • Each word received three repetitions, the second and third with substantial hyperarticulation. • All speakers reported finding this experimental methodology deeply exasperating. • Sessions were recorded directly to CD at a sampling rate of 44.1 KHz. Duration and LPC formant measurements were taken using Praat 4.1.11 speech analysis software (Copyright@1992-2003 by Paul Boersma and David Weenink). 5.2. Results (16) Vowel height as a function of phonetic duration Speaker 1 Speaker 2 R (54) = .618, p < .0001 R (58) = .934, p < .0001 Speaker 1: second pretonic /a/ and /o/ Speaker 2: second pretonic /a/ and /o/ 700 1200 600 1000 Series1 300 Linear (Series1) 2 R = 0.371 200 F1 in Hz 400 800 Series1 600 Linear (Series1) 400 R2 = 0.8741 200 100 0 0 0 10 20 30 40 50 60 0 50 100 Duration in ms. 150 200 250 Duration in ms. c. Speaker 3 d. Speaker 4 R (58) = .718, p < .0001 R (57) = .84, p < .0001 Speaker 4: second pretonic /a/ and /o/ Speaker 3: second pretonic /a/ and /o/ 1000 1000 900 900 800 800 700 700 600 Series1 500 Linear (Series1) 400 2 R = 0.5154 300 F1 in Hz F1 in Hz F1 in Hz 500 600 Series1 500 Linear (Series1) 400 R2 = 0.7052 300 200 200 100 100 0 0 0 10 20 30 40 50 60 70 0 80 20 40 60 80 Duration in ms. Duration in ms. -9- 100 120 140 160 (17) All speakers, raw scores: r (228) = .737, p < .0001 Second Pretonic /a/ and /o/: Raw Scores 1200 1000 F1 in Hz 800 600 Series1 Linear (Series1) 400 R2 = 0.5405 200 0 0 50 100 150 200 250 Duration in ms. • To escape problems of interspeaker comparison of vowel formant values, raw duration and F1 values for all speakers were converted to a standardized z-score format6, yielding the following results: (18) All speakers, z-scores: R (228) = .803, p < .0001 Second Pretonic /a/ and /o/: standardized 5 4 3 F1 2 Series1 Linear (Series1) 1 2 R = 0.6456 0 -3 -2 -1 -1 0 1 2 3 4 5 6 -2 Duration 5.3. Discussion • All speakers show a highly significant correlation between vowel duration and vowel height. Reduction of /a, o/ to [] applies gradiently, with syllables afforded sufficient duration by hyperarticulation showing little or no vowel raising. 6 zx = ( x - µx ) / σx, , where µ is the mean for a given set of values, and σ the standard deviation. -10- • Crucially, no amount of lengthening is sufficient to restore an underlying /o/ in an unstressed syllable. The neutralization of /a/ and /o/ in unstressed syllables is categorical and phonological, displaying no sensitivity whatsoever to phonetic duration. • Were the non-application of reduction to schwa in clear speech in any sense a paralinguistic effect, e.g. a spelling pronunciation, we would expect restoration of /o/ in unstressed syllables as well. This never occurs. 6. Implications of the Russian experiments • There are indeed two types of vowel reduction in Russian. • Only one is accomplished in the phonology. The other is a gradient process accomplished by the phonetics. // is not a phonological category in Russian. • Optimality-Theoretic interpretation of this conclusion: *// remains ranked above all Faithfulness constraints. • This conclusion simplifies Russian phonology greatly, while requiring nothing of the phonetics other than the capacity to implement fine-grained durational patterns conditioned by prosodic structure, a capacity it must have under any circumstances. 6.1. Exceptional Contexts Again • Additional payoff: the seemingly-arbitrary exceptions to Degree 2 reduction detailed in 2 above now make perfect sense. EXCEPTION #1: • Absolute word-initial position is associated cross-linguistically with articulatory strengthening and/or durational enhancement of consonants and vowels (Cho 2001, Cho and Jun 2000, Byrd 2000, Cho and Keating 1999, Keating, Cho, Fougeron and Hsu 1999, Dilley, Shattuck-Hufnagel and Ostendorf 1996, Fougeron and Keating 1996, inter alia Byrd 2000, Dilley, Shattuck-Hufnagel and Ostendorf 1996, Fougeron and Keating 1996, Oller 1973). • Consequences for positional neutralization elsewhere: a. Nawuri (Casali 1995, Kirchner 1998): absolute word-initial vowels fail to undergo an assimilation process involving roundness of which they would otherwise be targets. Long vowels and vowels in phrase-final position likewise avoid this assimilation. b. Luganda: the contrast between long and short vowels is neutralized in onsetless word-initial syllables. Additionally, only three vowels contrast in this position (in fact, initially in any morpheme): /e,o, a/ (Hubbard 1994: 161-165). c. Runyambo, in which initial /i, u/ are lowered to [e, o] after pause (Larry Hyman, p.c.). -11- EXCEPTION #2: • No reduction of /a/ to schwa before /a/ in hiatus. If reduction to schwa is essentially duration-dependent articulatory undershoot, a sequence of two such vowels should provide adequate duration for articulators to reach targets. EXCEPTION #3: • Absolute phrase-final position is associated cross-linguistically with additional duration for vowels, a phenomenon known as domain-final lengthening (see e.g. Oller 1973, Klatt 1975, Beckman and Edwards 1987, Wightman et al. 1992, Keating, Wright and Zhang 1999, inter alia). • Final-syllable vowels are resistant to reduction or assimilation processes which should otherwise target them in Belorussian, Ukrainian dialects, Brazilian Portuguese, Eastern Mari, Uyghur, Hausa, Catalan dialects, English, Yakan, Maltese, Nawuri, Shimakonde, Bonggi, and Timugon Murut (see Barnes 2002 for typological details)7. 6.2. Theoretical Consequences • Any model of the phonetics-phonology interface must be equipped to handle not only the distinction between categorical and gradient processes, but also the complete lack of sensitivity of categorical processes to their phonetic environments. • Best sort of model for degree 1 reduction: /a/, /o/ > /a/ Ident[hi]/σ@ • >> *MidV >> Ident[hi] Best sort of model for degree 2 reduction: /a/ -> [] for Speaker 18 f(d) = 52.187 + 112.758 * ln(d) where d = duration in ms, and f(d) = F1 in Hz • This is an immediate problem for the Direct Phonetics approach. • Technically, of course, gradience can be modeled in a categorical system by proliferating discrete stages along a continuum (Flemming 2001, Kirchner 1999). Flemming's approach to VR in particular allows two categories to approach one another in acoustic space up to a certain threshold, after which contrast can no longer be maintained, and neutralization occurs. • At issue though is not the way categorical and gradient processes apply, but rather the nature of the processes themselves. 6.2.1. VR in Seediq • Seediq is an Atayalic language still spoken in Taiwan (Asai 1953, Holmer 1996, Li 1977, Li 1980) 7 See Zhang 2001 for related facts involving the licensing of contour tones. A logistic growth model would actually be more appropriate here, avoiding unrestricted growth of F1 for large values of d. 8 -12- • Stressed syllables contrast five vowels: /i, e, a, o, u/. • Pretonic syllables allow [u] only, realized variably, and analyzed by Li as epenthetic. • Posttonic syllables realize [i, a, u], with mappings between stressed vowels and their unstressed counterparts as follows: (19) Posttonic vowel realizations in Seediq /i/ /a/ /e, o, u/ (20) -> -> -> [i] [a]/[] [u] Vowel reduction in Seediq /haed/ -> -> /heyeg/ -> -> ‘cook’ h-m-aed infixation haed-an suffixation -> hmaets d -> ts -> [humauts], (Actor Focus) epenthesis and VR -> hedan a -> ø -> [huedan], (Locative Focus) epenthesis ‘stand’ m-heyeg -> mheyew -> mheyuw10 -> mheyu -> prefix g->w VR w -> ø [meheyu] (AF) epenth., translar. harmony h-n-eyeg -> infix [hunuyean] (LF, Pret.) epenthesis hneyeg-an suffix -> hnyegan -> e -> ø • In some respects, Seediq is typologically ordinary: surface distribution of unstressed vowels in Seediq obeys a simple generalization banning mid vowels. • Still, we would presumably not want to generate reduction of /e/ to [u] through a gradual raising, backing, and rounding of [e], yielding intolerable perceptual proximity with [u], and ultimately neutralization11. • Seediq merger of /e/ and /u/ is not phonetically-natural, and we do not want a model of the phonetics-phonology interface which makes it appear otherwise. • Phonetics aside, however, the real question is whether there is anything wrong with Seediq vowel reduction grammatically. The direct phonetics approach to VR must conclude that there is. • Russian Degree 1 reduction and Seediq posttonic reduction would need to be generated by distinct grammatical mechanisms, but are they really so different? 9 /a/ and /i/ are said by to retain their stressed qualities “to a certain extent” (Holmer 1996: 26). I assume this means that they are somewhat centralized. 10 Final -Vw simplified to -V is independently attested and is not the source of [-u] here. Other dialects retain /w/. 11 See Barnes 2002 for a possible diachronic account of this system. -13- • Similar problems arise for approaches deriving typological regularities in phonology from grounding conditions on constraint construction. 6.2.2. VR in Sosva Mansi • Recall that while VR systems routinely target vowel height contrasts, VR systems targeting other features, such as palatality, seem not to occur. • This is the type of generalization grounding conditions such as Smith's Substantive Filters were devised to express. • Constraints barring non-high vowels from unstressed syllables should be freely constructed, while constraints barring, e.g., back vowels from unstressed syllables, should be prevented from entering CON. • In the Sosva dialect of Mansi (Ob-Ugrian), however, only /i, e, , a/ surface outside the initial, stressed syllable (Honti 1988: 335). Contrast of front and back vowels is neutralized12. • Again, typologically, Sosva Mansi is an aberration. But is it an aberration grammatically as well? 7. Conclusions QUESTION : Is there any evidence to suggest that from the point of view of synchronic phonology, there is anything wrong with the reduction systems of Seediq or Sosva Mansi? Are systems like these harder to learn, less stable over time, or in any other way more costly to maintain than other, more common systems? Absent positive empirical evidence that this is so, our assumption should be: • UG is capable of generating and sustaining VR systems like those of Seediq or Sosva Mansi just as well as it does Russian or any other VR system, regardless of whether they are phonetically natural or grounded synchronically. • It follows from this that the typology of VR (and other positional neutralization) systems cannot be (solely) a consequence of constraints on possible grammars imposed by UG. So where does the typology come from? • Barnes 2002 finds the source in phonologization: the influence of phonetics on phonology is indirect. The gradient phonetic component affects the structure of the categorical phonological component through sound change. • Sound change is driven by phonetics, such that sound changes are by definition phonetically natural. "Unnatural" systems can arise through the complex interaction of a set of changes, but are predicted to be uncommon (see e.g. Ohala passim, Blevins and Garrett 1998, Blevins, in press, for related views). 12 That this system exists at all is probably due to the fact that Proto-Ob-Ugric had a system of palatal harmony in place. While both front and back vowels would have surfaced in non-initial syllables, contrast was already prohibited. -14- • Once they do arise, however, typologically disfavored patterns are predicted to be no less robust or phonologically serviceable than any others. • Typological patterns surely sometimes are the result of constraints on possible grammars present in UG. This should not, however, be our null hypothesis. • As far as the phonological grammar is concerned, it is enough for speakers to know that their languages contain a given phonological pattern. They need not also know why. -15-