A CROSS-VARIETAL CONTINUUM OF UNSTRESSED VOWEL REDUCTION: EVIDENCE FROM BULGARIAN AND TURKISH Mitko Sabev, Elinor Payne University of Oxford [email protected], [email protected] ABSTRACT reduction (UVR). This is not surprising, for it has been argued that Turkish is a pitch accent language We compare speech production data from three [9] and shows no significant stress-dependent spec- Bulgarian and two Turkish varieties with respect to tral differences [10], and that reduction cannot coex- spectral and durational reduction of unstressed non- ist with vowel harmony (VH) in a language [11]. front unrounded vowels, and ensuing height neutral- We consider spectral and durational data from isation. Istanbul Turkish lies at one end of a reduc- five varieties expected to exhibit different degrees of tion continuum, with only non-neutralising, gradient UVR: East Bulgarian (EB), West Bulgarian (WB), F1 frequency undershoot that correlates with dura- bilingual Bulgarian (BB) and Turkish (BT) as spo- tion. Monolingual East Bulgarian lies at the opposite ken by a long-standing bilingual community from end: unstressed, underlyingly non-high vowels raise the same region as the EB monolingual group (Tǎr- considerably and merge with their high counter- govište), and Istanbul Turkish (IT). Based on previ- parts. The Bulgarian speech of bilingual Turkish– ous research and assumptions, we expect any spec- Bulgarian speakers from the same region of eastern tral UVR in IT to be gradient and correlated with Bulgaria shows less reduction and neutralisation; duration, as could result from gestural undershoot perhaps surprisingly, it resembles the reduction pat- under temporal pressure. For WB and EB categori- tern of West (Standard) Bulgarian, while at the same cal UVR is expected, with much less or no correla- time also being gradient, probably under the influ- tion with duration. EB should show more UVR, and ence of Turkish. The bilinguals’ Turkish speech, on greater merger of high and non-high unstressed the other hand, exhibits more neutralisation than Is- vowels. It is difficult to predict where exactly BT tanbul Turkish, but less than their own Bulgarian, and BB will lie on a reduction scale, as these varie- which in turn suggests prosodic transfer from these ties have not been studied before. If we are to be- speakers’ Bulgarian to their Turkish. lieve that VH and UVR are typologically incompati- ble, there should be a definitive watershed point with Keywords: vowel reduction, vowel merger, incom- BT and BB logically lying on different sides, or one plete neutralisation, Bulgarian, Turkish of them having crossed that point as a result of L1– L2 transfer. [11] argues that VH and UVR do not co- 1. INTRODUCTION occur because VH neutralises backness and round- ness while preserving height contrasts, whereas The phonological literature often cites Bulgarian as UVR does precisely the opposite. The watershed having a system of six contrastive stressed vowels point should then be defined in terms of both the that neatly contracts into a three-vowel unstressed difference between stressed and unstressed allo- inventory through reduction and height neutralisa- phones, and – perhaps more importantly – the degree tion in each of the pairs /i-ɛ, ə-a, u-ɔ/ [1–4]. Howev- of height contrast neutralisation, i.e. merger. er, phonetic descriptions make clear that while this For space limitations, we focus here on Bulgarian may be true of certain eastern dialects, it is not so in /ə-a/ (ъ-а), and Turkish /ɯ-ɑ/ (ı-a). As well as Standard (western) Bulgarian, where only /ə-a/ and being phonologically matched (both pairs are non- /u-ɔ/, but not /i-ɛ/, merge in unstressed position. Two front/back and non-round), these vowels are suffi- claims have traditionally been made about the pho- ciently phonetically similar to map onto one another netic realisations of the merged vowels, both of in loan adaptation and foreign language learning. which we challenge in this paper: (1) that unstressed high and non-high vowels merge into intermediate 2. METHODS realisations, such as [ɐ] and [o], respectively, and (2) that immediately pretonic vowels are more open 2.1. Participants, procedure and recording than other unstressed vowels. [5–8] Turkish, too, has a system that can be symmetri- Native speakers of each variety, aged 18–27, were cally divided into height-contrasting pairs, /i-ɛ, y-œ, recorded in return for nominal payment (Table 1). ɯ-ɑ, u-ɔ/, but is not known for unstressed vowel The bilinguals’ Bulgarian and Turkish recordings were made on separate days. Participants read out termined by the presence of clear formant structure utterances from a computer screen, displayed one at and sharp changes in intensity. A Praat script was a time, in a quasi-random order. The recordings used to extract vowel duration and midpoint F1 and were made on a MacBook with an iRig Mic HD ex- F2 frequencies. ternal microphone, digitised at a sampling rate of Outliers, defined as values 1.5 times beyond the 44,100 Hz and a 16-bit resolution, stored as PCM- interquartile range, and devoiced vowels were ex- encoded single-channel WAV files. cluded from analysis. Formant values were normal- ised using [13]’s adaptation of Neary’s vowel- Variety Sp M F Vowel tokens extrinsic, formant-intrinsic log-mean procedure, as EB 12 5 7 1216 implemented in the vowels package for R [14,15]. WB 12 4 8 1344 A combination of repeated-measures ANOVA, BB 14 5 9 1064 MANOVA and t-tests were performed to measure BT 14 5 9 1232 the effect of stress and syllable (σ , σ , σ́ , σ ) on F1, IT 14 6 8 1176 1 2 3 4 F2 and duration. Pearson product-moment correla- Table 1: Number of M(ale) and F(emale) tion between duration and F1 within σ1–4 allophones Sp(eakers) and vowel tokens per experiment. was used as an indicator for gradient undershoot. F1/F2/duration distribution overlap was calculat- 2.2. Test items ed between (1) stressed and unstressed allophones for each vowel, lesser overlap indicating greater A set of nonsense words were designed to test the acoustic displacement in unstressed position; (2) effects of various accentual and segmental contexts high and low vowels in unstressed position as a on the three variables of interest: F1, F2 and dura- measure of merger, greater overlap corresponding to tion. Real words were also included to verify that the a higher degree of merger. Two overlap measure- nonsense words were reliable indicators for speak- ment procedures were used. In the first, allophones ers’ natural linguistic behaviour. Comparisons of are plotted as 2-SD best-fit ellipsoids in the real- and nonsense-word vowels revealed mostly F1/F2/duration space. The fraction of overlap be- non-significant differences in formant frequencies. tween the two ellipsoids is computed, which can The nonsense words had the shape range from 0 (no overlap) to 1 (complete overlap) (C)VCVˈCVC(C)V(C), usually keeping vowel and [16]. The second metric used is Pillai’s trace, part of consonant phonemes constant throughout the word, the output of MANOVA performed with F1, F2 and and conforming to Bulgarian/Turkish phonotactics. duration as dependent variables. A higher value in- As far as possible, the nonsense words had compa- dicates a greater difference between the two distribu- rable segmental and accentual structure across the tions with respect to the dependent variables. For a two languages. In Bulgarian, stress was indicated discussion of these and other vowel overlap metrics, with a grave accent on the vowel. Turkish stress is see [17]. generally word-final, so in order to elicit penultimate stress, the enclitic /lɑ,lɛ/ ‘with’ (or /diɾ,dyɾ,dɯɾ,duɾ/ 3. RESULTS ‘is’) was attached to a trisyllabic nonsense stem. Multiple unstressed syllables per word were neces- The ANOVA outputs showed significant main ef- sary to test the repeated claim that pretonic vowels fects in all cases except for F1 in IT /ɯ/. To estab- in Bulgarian are more open than other unstressed lish which comparisons are likely to have produced vowels. Consonants varied for: place of articulation, these effects, dependent t-tests with Bonferroni cor- /ptk/; voicing, /p b/; degree of stricture, /b v/; C ~ ∅ rection (α = 0.0125 for Bulgarian, and 0.0167 for word-initially and finally. In some words stressed Turkish) were performed to compare the three pa- and unstressed vowels differed in phonological rameters across the three/four syllables (Table 2). height. Except for vowels being consistently longer The mean values in Table 3 interpreted in the in all varieties when flanked by voiced consonants, context of significant differences (Table 2) reveal a segmental context did not yield any conclusive re- number of clear patterns. In all varieties, duration is sults. The words appeared in the carrier sentences longest in the stressed (third) syllable, and in Bulgar- /ˈkazax … ˈpak/ ‘I said … again’ (Bulgarian) and ian the second longest syllable is the fourth and final /ɑhˈmɛt … dɛˈdi/ ‘Ahmet said …’ (Turkish). one. This pattern is only broken by /ə/ in BB, where [ə4] is even longer than [ə́ 3], and in EB, where [ə́ 3] = 2.3. Analysis [ə4]. In all Bulgarian dialects the second and pretonic vowel tends to be the shortest, while in Turkish V The vowels were manually segmented in Praat [12], 1 and V often have equal durations. F1 is highest in on the basis of the synchronised spectrogram, wave- 2 stressed position across the board for the low vowel, form and audio signal. Vowel boundaries were de- and also for the high vowel in WB and BT, while in Duration–F1 Duration–F2 EB [ə4] > [ə1,2,3], in BB [ə4] > [ə1] > [ə2,3], and in IT σ1 σ2 σ́3 σ4 σ1 σ2 σ́3 σ4 EB n.s.