The Uyghur inventory includes at least seven contrastive vowels, /ɑ o u æ ø i y/, which are distinguishable in terms of three features, [back], [high], and [round], shown below in Table 1. In addition to these vowels, /e/ is marginal, typically occurring in non-nativized loans, and in the initial-syllable only (Hahn, 1991, p. 37). As noted by Nadzhip (1971, pp. 48–49) and Yakup (2005, p. 31), [e] is typically a raised variant of /æ/ or /ɑ/, further suggesting the peripheral status of this sound. In comparison with the common Turkic eight-vowel system (e.g., Menges, 1995), there is one notable gap in the Uyghur inventory—there is no [+back] counterpart of /i/.

Backness harmony in Uyghur triggers alternations on the low vowels and the high rounded vowels. In (1a-d), the locative suffix alternates between [-dæ] and [-dɑ] according to the backness of the initial-syllable vowel. Similarly, in (1e-h) the backness of the initial syllable conditions the backness of the vowel of the gerundial suffix, as well as the place of articulation of the preceding dorsal consonant, [ɡ]~[ʁ].

In addition to the low vowels and the high round vowels, two other sets of vowels can be defined for harmony. The first is comprised of the mid vowels /e ø o/, which do not typically occur in non-initial syllables (though see Abdurehim, 2014, pp. 70–75 for the non-initial occurrence of [ø o] in the Lopnor dialect). Since they do not typically occur in alternating contexts, the paper does not focus on these vowels. The second set contains /i/, which previous work reports is transparent to harmony (Lindblad, 1990; Hahn, 1991; Vaux, 2000; Mayer & Major, 2018; Mayer, Major, & Yakup, 2019; cf. Nadzhip, 1971, p. 49; Yakup, 2005, p. 55; Abdurehim, 2014, p. 74). Relatedly, root /i/ triggers [+back] vowel suffixes in some roots (2a,b), but front vowel suffixes in others (2c,d). In polysyllabic roots /i/ is transparent; second-syllable /i/ of disyllabic roots does not affect the realization of following vowels (2e,f).1

In (2), variation occurs across different lexical roots, but the data in (3) demonstrate that a single root may trigger both front and back suffixes. In (3a,b), the roots /iʃ/ ‘work’ and /ʧiʃ/ ‘tooth’ trigger the [+back] variants of the plural and dative suffixes. Yet, in (3c,d), these roots trigger the [–back] variants of the verbalizer suffix. While this variation may be partially conditioned upon the particular suffixes involved, there is also variation that is completely independent of suffix type. Observe in (3e,f) that /ilim/ ‘science’ may trigger either front or back variants of the plural suffix. Similar variation is described in Lindblad (1990, p. 26), and is common in the 2016 Wikipedia Uyghur corpus maintained in the Leipzig Corpora Collection (Goldhahn, Eckart, & Quasthoff, 2012).

Significantly, no other vowels behave like /i/ as a trigger for harmony—the back vowels /ɑ o u/ consistently trigger [+back] suffixes, and the front vowels /æ y ø/ consistently trigger [–back] suffixes. The fact that root-internal /i/ behaves differently from other vowels has been used to support the case that the phonological status of /i/ is fundamentally different from the other vowels in the language. Like root-internal /i/, suffixal /i/ does not alternate for harmony, nor does it impose its own phonetic frontness on subsequent vowels, as seen in (4). In (4a,b), the backness of the roots extends across two transparent /i/ vowels to determine the realization of the locative suffix.2

In addition to underlying high vowels, high vowels may surface via vowel raising or epenthesis in the language. Epenthetic vowels are inserted to break up illicit coda clusters. The first-person singular possessive suffix, /-m/, regularly triggers epenthesis when attached to consonant-final roots. In (5a,b) this suffix attaches to a vowel-final root, surfacing without any epenthetic vowel. In (5c-f) however, this suffix is preceded by an epenthetic high vowel when concatenated to a consonant-final root. The epenthetic vowel is rounded after [+round] roots (5c,d) and unrounded after [–round] roots (5e,f). Lindblad (1990) and Vaux (2000) analyze the epenthetic vowel as copy of preceding backness and rounding, subject to a prohibition on [ɯ], resulting in surface [i] instead of [ɯ] after back vowels (5f).3 When [i] surfaces, the backness of the preceding vowel still controls the realization of subsequent suffixes, like the locative suffix shown in (5).

As for vowel raising, low vowels in medial open syllables raise to [+high]. To help keep track of underlying vowel qualities, both underlying and surface forms are presented in (6). In (6a), the unaffixed root for ‘child’ shows that the word-final vowel is low. The underlying height of the root-final vowel is also evident in (6b,c), where this vowel occurs in a closed syllable. Yet, when this vowel occurs in a medial open syllable, as in (6d), it raises to [i]. The same generalizations hold for all underlying low vowels, as in (6e–j).

Whether underlying, epenthetic, or derived by raising, Lindblad (1990, p. 13) and Vaux (2000, p. 2) suggest that /i/ does not even exhibit small, phonetic effects for harmony. Lindblad (1990, p. 13) writes:

Lindblad’s claim is noteworthy, since it suggests that backed variants of /i/ may surface due to consonantal effects, but not due to harmony. The distribution of the high unrounded vowels is noted most extensively in Hahn (1986, pp. 43–48, 1991, pp. 34–37). Generalizing over Hahn’s (1991, pp. 34–37) complex set of allophonic statements, backed allophones of /i/, including [ə ɨ ɤ ɯ], appear following dorsal obstruents, laryngeals, and before /l/ and /ŋ/. In all other contexts, /i/ surfaces as a true front vowel, usually transcribed as [i] or [ɪ]. To simplify discussion throughout, all front allophones of /i/ are transcribed as [i], while all back allophones of /i/ are transcribed as [ɯ].

In most work on the language, underlying /ɯ/ is preserved, but surface [ɯ] is neutralized with [i], in line with Vago’s (1973) abstract analysis of Hungarian (Hahn, 1991; Lindblad, 1990; Yakup, 2005). The use of a late neutralization rule crucially allows harmony to operate locally during the lexical stratum of the phonology, only to be masked by the later rule mapping |ɯ| to [i]. In Vaux (2000), a different analysis is proposed, which requires harmony to operate over contrastively specified vowels only (see also Calabrese, 1995; Nevins, 2010). Under Vaux’s analysis, since /i/ does not have a [+back] harmonic counterpart, the harmony rule effectively skips the high unrounded vowel.

Problematically, in all previous work the fact that [ɯ] may not arise via harmony is never juxtaposed to the allophonic backing of /i/ to [ɯ] in certain consonantal contexts. For instance, Hahn (1986, 1991) reports that /i/ surfaces as a back vowel before /l/, which suggests that the transcription of the forms in (6e,j) should be [bɑlɯlɑr] rather than [bɑlilɑr] ‘child-PL’ and [sællɯlær] rather than [sællilær] ‘turban-PL’ respectively. When the role of adjacent consonants is considered, the transparency evident in [bɑli-dɑ] ‘child-LOC’ is obscured in [bɑlɯ-lɑr] by the allophonic backing of /i/ before the lateral. The same complication holds for front vowel forms, as well. Given the effects of flanking consonants, transparent, but harmonic /i/ in [sælli-dæ] ‘turban-LOC’ should be transparent and disharmonic [ɯ] in [sællɯ-lær] ‘turban-pl.’ In both sets of forms, the predicted allophonic backing of /i/ to [ɯ] before /l/ masks the harmony pattern. If this is the true state of affairs in the language, with both vowel and consonantal features driving high unrounded vowel backness, transparency appears more complex than other cases of transparency in harmony.

First, in a number of languages with reported transparency, phonetic studies have shown that the ‘transparent’ vowels actually alternate for harmony (Gordon, 1999; Gick et al., 2006; Benus & Gafos, 2007; Ritchart & Rose, 2017). Gordon (1999) examines putatively transparent /e/ and /i/ in Finnish, reporting small (up to 100 Hz) differences in the second formant based on backness context. Similarly, Benus and Gafos (2007) find small articulatory and acoustic differences in Hungarian /i/ based on backness context, arguing that these vowels are not truly transparent at the phonetic level. Unlike the small, decidedly phonetic effects reported for Finnish and Hungarian, Gick et al. (2006) and Ritchart and Rose (2017) report much more salient alternations in Kinande and Moro. In Kinande, the low vowel /a/ is produced with lower F1, and with significantly advanced tongue root in [+ATR] contexts; Gick and colleagues propose that underlying /a/ phonologically alternates with surface [ə]. Even more strikingly, Ritchart and Rose (2017) demonstrate that the vowel once thought to be transparent in Moro, /ə/, is actually two distinct vowels, /ə/ and /ɘ/, which form a contrastive pairing for height harmony, differing both as triggers and targets of harmony. Like in Kinande, the results in Ritchart and Rose’s study support a phonological distinction, and not just a low-level phonetic difference based on harmonic context. In turn, these findings support theoretical claims of ‘strict locality,’ which demand that all segments within a harmony domain alternate phonetically for harmony (Gafos, 1999; Ní Chiosáin & Padgett, 2001).4

Yet, some more recent results cast doubt on the larger claim that all transparency is false. Dye (2015) examines ultrasound and acoustic evidence from Pulaar and Wolof ATR harmony. She finds that both /i/ and /u/ fail to undergo even low-level phonetic alternations based on ATR context in Wolof. Additionally, Szeredi (2016) questions the perceptual significance of the types of low level effects reported in Hungarian and Finnish. His findings suggest that Hungarian speakers do not perceive the small phonetic differences based on vowel backness context (cf. Benus & Gafos, 2007), arguing rather that other sublexical cues support learning transparent and exceptional patterns in harmony.

In addition to pseudo-transparency, phonological evidence supports an additional, distance-based restriction on transparency. In Hungarian, a single /i/, /iː/, /e/, or /eː/ is transparent (with, as noted above, small phonetic effects) to backness harmony. This is demonstrated in (7). In (7a,b), a [+back] root is followed by each of the two diminutive suffixes, which both surface as [i] irrespective of root backness. In (7c,d) both suffixes are followed by the regularly alternating dative suffix, which agrees with the initial-syllable vowel for [back]. Thus, when a single /i/ occurs, it is transparent to harmony. However, in (7e,) we see a different pattern when harmony must span two consecutive transparent vowels. When a regularly alternating suffix is preceded by two transparent vowels (7e), that suffix is far more likely to surface as [–back], regardless of initial vowel backness (Vago, 1980; Ringen & Kontra, 1989; Siptár & Törkenczy, 2000; Hayes & Londe, 2006; Gafos & Dye, 2011).

This effect is distinct from the low-level phonetic alternations reported in Benus and Gafos (2007). In (7f), the final syllable is output as [ɑ] or [e] (phonetically [ɔ] or [ɛ]), not slight variants of a single category, which suggests a decidedly phonological effect. If these vowels were entirely transparent, it should not matter how many of them intervene between a trigger and target. However, the fact that their number does matter, suggests that they are not entirely transparent, a point made by Ringen and Kontra (1989), Hayes and Londe (2006), as well as Gafos and Dye (2011).

While some previous work argues that all ‘transparent’ vowels alternate for harmony (Gafos, 1999; Ní Chiosáin & Padgett, 2001), Dye’s (2015) findings undermine any simple reanalysis of putative transparency. Further, in some cases these alternations appear to be purely phonetic (Finnish), while in others they are allophonic (Kinande), or even evidence for an additional, unreported vowel contrast (Moro). Thus, the range of actual phonetic and phonological variation discernible among reportedly transparent vowels is not at all clear. Additionally, data from Hungarian suggest that it may be overly simplistic to categorize elements as categorically transparent or not, since the invisibility of the front unrounded vowels in Hungarian is constrained by distance. For these reasons, data from more languages are necessary to evaluate extant theoretical claims, and to further understand the amount and types of variation attested in these types of harmony patterns.

In addition to the general need to collect more data, Uyghur provides a promising opportunity for experimental work. Previous research on the language, while cognizant of claims concerning strict locality (Vaux, 2000, footnote 2), argues that the realization of /i/ is unaffected by flanking vowel backness. Thus, Uyghur may present another case like Wolof, where certain vowels are invisible to harmony. Second, reports indicate that transparency persists across multiple vowels, although strings of three or more medial /i/ vowels are never discussed. Given Uyghur’s agglutinatve morphology and morphotactics, it is possible to produce words with longer sequences of medial /i/ vowels, which offers a fuller comparison with a language like Hungarian, where phonological transparency is sensitive to distance.

Existing work on transparency in Uyghur depends entirely on textual and impressionistic data. This paper, however, investigates the pattern with acoustic data from a production study. In addition, previous work has not examined the reportedly variable behavior of lexical items as a way to examine the Uyghur harmony system. This study uses acoustic data paired with counts of [±back] suffix selection to provide a fuller picture of harmony in the language. Specifically, four questions are addressed. First, do different members of a near-minimal pair select different classes of suffixes? Second, does F2 of root-internal [i]-[ɯ] differ between members of a near-minimal pair? These two questions probe whether there is a relationship between the backness of root-internal [i]-[ɯ] and the backness of following vowels? Third, do non-initial [i] and [ɯ] alternate for harmony? Fourth, do low vowels alternate for harmony when preceded by medial high unrounded vowels? The first and second questions address the behavior of high unrounded vowels in trigger positions. The third and fourth questions address the behavior of high unrounded vowels in target positions.

Preliminarily, results indicate that the status of [i] and [ɯ] differs across speakers, particularly within roots. In Hungarian, /i/ does not behave as entirely transparent or alternating. Similarly, the behavior of /i/ in Uyghur does not fit neatly within any single category. For at least one speaker, there is evidence for contrastive /i/ and /ɯ/ while for others production data support a non-contrastive relationship between [i] and [ɯ]. More generally, the data discussed in the following sections suggest that the possible types of phonological relationships may be more numerous than the standard categories discussed in the literature.

Participants were presented a set of pictures corresponding to the Uyghur nouns containing the three uncontroversial harmonic pairings in the language, /ɑ-æ, o-ø, u-y/. Pictorial prompts were used to avoid an orthographic confound, since Uyghur orthographies do not represent [ɯ]. Target words were derived from monosyllabic and disyllabic roots, as shown in (8). Monosyllabic roots ended either in a sibilant or a liquid (8a–g), and disyllabic roots contained two vowels that agreed for the feature [high] (8h–n).

Additionally, stimuli with putative /i/ were drawn from lexical items that either, one, exhibit variation in Hahn (1991) or Lindblad (1990), or two, are cognates with /ɯ/ in closely related languages that maintain contrastive /ɯ/, Kyrgyz and Kazakh. The full set of root-internal /i/ stimuli is shown in Table 2. These particular lexical items were selected for study because their status in closely related languages differs. Hahn (1991:47) indicates that the behavior of Uyghur /i/ reflects the historical status of *i and *ɯ. Thus, selecting stimuli with differing backness in related languages that have maintained a clear contrast between these two vowels increases the likelihood of detecting a contrast, if one exists, in contemporary Uyghur. Two monosyllabic stimuli were selected that are reported to trigger [+back] harmony and have [+back] cognates in Kyrgyz and Kazakh, /qiʃ/ and /jil/. Two monosyllabic stimuli were selected that are reported to trigger [+back] harmony but have [–back] cognates in Kyrgyz and Kazakh, /ʧiʃ/ and /pil/. Finally, two disyllabic target words were selected that reportedly trigger [–back] harmony and correspond to [–back] cognates in Kyrgyz and Kazakh /ilim/ and /ʃilim/. As far as I know, there is no correlation between the length of the root and its tendency to trigger [±back] suffixes. In Lindblad (1990), a number of monosyllabic verbs as well as functional items with /i/ that select for [–back] suffixes are listed, but among nouns, the only roots that are lexically marked as [–back] are disyllabic.5

Table 2 also includes predicted surface quality for each word based on Hahn’s (1991) detailed discussion of /i/ allophony. For the present study, it is relevant to note that Hahn predicts that /i/ is fronted around (alveo)palatals and is backed around dorsals and laterals. The elicited words thus exemplify a range of surface qualities, varying in backness from [i] to [ɤ]. Also, the elicited contexts pit Hahn’s various descriptions against one another. For instance, he predicts fronting before /ʃ/ but backing after /q/. Thus, in a word like /qiʃ/ it is not at all clear which allophonic pattern prevails. One benefit of the present study is the ability to assess the relative dominance of conflicting pressures on the realization of /i/. Since the topic of discussion is backness harmony, all front vowel phones related deriving from /i/ are transcribed as [i] while all central and back phones deriving from /i/ are transcribed as [ɯ].

Observe that the stimuli in Table 2 form three near-minimal pairs. Previous work on Uyghur has focused over larger sets of lexical items, devoting almost no discussion to the potential value of pairs such as those in Table 2. In structuralist and post-structuralist linguistic research, the use of minimal and near-minimal pairs is a primary heuristic for discovering phonological contrasts. As such, evaluating a small number of pairings, their acoustic characteristics, and their behavior with respect to harmony offers potential insight to complement existing work on the language.

Each session was divided into training and recording phases. During the training phase, participants, all of whom are native Uyghur speakers, were taught a small set of pictorial-lexical correspondences. As an example, a photo of a flower prompted the word /gyl/ ‘flower’ while a photo of lake prompted the word /køl/ ‘lake.’ In addition, participants were also taught a set of pictorial-grammatical correspondences involving number, case, and possession. For instance, a downward red arrow indicated locative case while two outward pointing red arrows indicated ablative case. The training phase typically lasted less than five minutes. After participants completed training, the recording phase began. Throughout each session, participants were presented images on a laptop computer screen that included both a picture representing a lexical item and a grammatical prompt from the training phase. Thus, a picture of a single flower with a downward pointing red arrow would prompt the word /gyl-dæ/ ‘flower-loc.’ When speakers were unable to guess the target word from the prompt, they were given either the equivalent Russian word or a paraphrase in the target language. To be clear, participants did not read or hear any of the target words. Instead, they inferred them from pictorial prompts or discussion with the researcher.

Roots were elicited in four cases (nominative, accusative, locative, and ablative), singular and plural numbers, and in first- and third-person possessive forms. Example inflected forms from the roots /bæl/ ‘waist’ and /bɑl/ ‘honey’ are shown below in (9). Participants were prompted to produce each word only once. If a speaker produced a particular word several times though, all repetitions were analyzed.

I examined three types of high vowel targets, underlying, epenthetic, and raised. Underlying high vowels are present in three of the suffixes elicited, the ablative, accusative, and third-person possessive (9b,d,h). An epenthetic high vowel occurs preceding the first-person singular possessive suffix after consonant-final roots (9g). Third, I examined raised vowels, which surface via the reduction of low vowels in medial open syllables, shown in (6). These raised vowels were either the second syllable of a disyllabic root, or the plural suffix.

There are two basic questions that motivate the paper—what is the relationship between the high unrounded vowels root-internally, and how do the high unrounded vowels participate in harmony. These two questions are, in turn, framed in more specific, testable terms in the four questions in (10). Vowel backness is operationalized as variation in the second formant (F2), since this is the primary acoustic manifestation of varying tongue body backness. Since back vowels exhibit lower F2 than front vowels, for any alternating pair, F2 of a given vowel should be lower in a [+back] context and higher in a [–back] context.

The lexical items used in the study form three near-minimal pairs, qiʃ-ʧiʃ, jil-pil, and ilim-ʃilim. Using these pairs, it is possible to probe the nature of contrast and the role of adjacent consonants on root-internal [i]-[ɯ]. I evaluate the acoustics of root-internal high unrounded vowels to determine if there exist spectral differences, specifically F2 differences that are not attributable to consonantal context. In addition, using pairs allows one to connect potential differences in suffix allomorph selection and the acoustics of root-internal [i]-[ɯ].

The first two questions, which concern the nature of root-internal [i] and [ɯ], yield four logically possible interpretations, which are shown in Table 3. If each member of the near-minimal pair triggers the same set of suffix allomorphs, then there is no evidence for a contrast between pair members, and the relationship between [i] and [ɯ] is at most allophonic. If, however, different roots trigger different suffix allomorphs, this would suggest either a transparent (with lexical idiosyncrasy; as in Hungarian) or contrastive relationship between the high unrounded vowels (as in Moro). Previewing the results to be discussed in Section 5.1, data from some speakers show evidence for a non-contrastive relationship with potential allophonic differences between [i] and [ɯ], data from other speakers show evidence for an idiosyncratic relationship between near-minimal pairs, and data from at least one speaker suggest a contrastive relationship between the two high unrounded vowels.

In addition to their contrastive status and behaviors as triggers, Section 5.2 discusses the behavior of the high unrounded vowels as targets of harmony. Do these vowels alternate for backness harmony, and do subsequent vowels alternate for harmony? Four possibilities are noted in Table 4. If these vowels fail to alternate, but allow root backness to propagate to more peripheral suffixes, as reported in previous research, then the high unrounded vowels are transparent. If these vowels fail to alternate, but stop [back] spreading, then they are blockers. As a third possibility, both the high unrounded vowel and subsequent suffixes may alternate for harmony. This state of affairs would suggest that these vowels are not transparent, but alternate for harmony. Finally, it is possible that these vowels alternate for harmony but prevent backness from spreading to subsequent suffixes. Jurgec (2011) calls this type of vowel an “icy target.” As a preview of findings from Section 5.2, both F2 of non-initial high unrounded vowels and F2 of following low vowels is predictable based on root backness. This suggests that [i] and [ɯ] phonetically vary in accordance with the larger harmony pattern.

To examine the behavior of non-initial high unrounded vowels, elicited words fell into one of two conditions, exhibiting either a short- or long-distance dependency (11). The short-distance condition involved only a single high unrounded vowel between trigger and alternating target vowel. In contrast, the long-distance condition involved three high unrounded vowels intervening between trigger and alternating target vowel. If the alternation of low vowel suffixes varies across these two conditions, then transparency is distance-based, as in Hungarian. Speaker 1 did not participate in the short- and long-distance conditions, producing data only for the root-internal analysis in Section 5.1.

Sessions were conducted in a quiet room. Participants wore a Shure-SM10A unidirectional head-mounted microphone, and all data were recorded to a Marantz PMD 661 MKII digital recorder at a sampling rate of 44.1 kHz. Each session lasted between 45 and 90 minutes.

Participants were recruited through existing relational networks in Chunja, Kazakhstan. Speaker participation and informed consent were obtained in accordance with University of California San Diego Linguistic Fieldwork IRB protocol #141520. Nine Uyghur speakers (five females; mean age: 44.4 years; range: 19–63 years) participated in the study. All speakers were from Chunja or immediately surrounding villages. All participants reported native fluency in Uyghur, as well as fluency in Kazakh and Russian. When asked which of the three languages they speak best, Speakers 1, 2, 4, and 9 reported Uyghur; Speaker 6 reported Russian; and Speakers 3, 5, 7, and 8 reported that they speak all three equally well. This is unsurprising in Chunja, the seat of the Uyghur district in the Almaty region. In Chunja, Uyghur is spoken and written in public, and used in more domains than elsewhere in Kazakhstan, where proficiency in Uyghur is often secondary to proficiency in another language. According to the 2009 national census, 83.8% of Uyghurs indicated the Uyghur was their mother tongue. For comparison, 93.7% of Uyghurs reported fluency in Kazakh, and 95.8% reported fluency in Russian (Smailov, 2011). The proficiency of Kazakhstani Uyghurs in Kazakh and Russian differentiates them from those living in China, who are often bilingual in Uyghur and Mandarin, with Mandarin encroaching in many linguistic domains.

Since almost all theoretical work on Uyghur phonology has focused on the standard variety spoken in much of Xinjiang, it is important to note that the variety spoken in Kazakhstan is part of the central dialect group, upon which the standard language is based (Kaidarov, 1970; Hahn, 1986, pp. 36–42, 1991, pp. 5–6; Yakup, 2005, pp. 8–22). Further, previous research indicates that one of the traits that Kazakhstani Uyghur shares with the standard variety, which has been the focus of most theoretical work, is the behavior of /i/ (Baskakov, 1970; Kaidarov, 1970; Yakup, 2005, p. 8). Therefore, results reported below should be generally relevant for existing theoretical proposals concerning harmony and transparency in Uyghur.

All sound files were segmented in Praat (Boersma & Weenink, 2015). The beginning and end of each vowel were set to the onset and offset of the second formant. In cases where the second formant persisted across flanking consonants, abrupt changes in energy or formant frequencies were used to indicate vowel onset and offset.

After segmentation, the second formant was measured at vowel midpoint. To facilitate across-speaker comparisons, the data were z-score normalized (Lobanov, 1971). Four tokens of /ɑ o u æ ø y/ were used for normalization. In addition, eight tokens of the high unrounded vowels, four tokens with relatively high F2 and four with relatively low F2 were included for normalization.

Outliers were inspected for measurement errors. A number of errors were found with [u], where the formant tracker in Praat failed to distinguish the first two formants. In these cases formant frequencies were hand measured at the approximate vowel midpoint. The data were analyzed in R (R Core Team, 2017), using the lme4 package (Bates, Machler, Bolker, & Walke, 2015). A mixed effect linear regression was used to predict normalized F2 at vowel midpoint for both high unrounded vowels and subsequent alternating vowels. Significance was assessed using likelihood ratio tests.

The first question to address is whether members of each near-minimal pair exhibit different patterns of suffix allomorph selection. Mean (and standard deviation percent) [+back] suffix selection (n = 296 words) aggregated across speakers is shown in Table 5. Interestingly, roots with [+back] cognates in Kyrgyz and Kazakh, /jil/ and /qiʃ/, triggered [+back] suffixes 100% of the time, whereas roots with [–back] cognates in those related languages exhibited significant variation. Note however that every root tended to trigger [+back] suffixes, with /ilim/ being the most likely to precede a [–back] suffix. Observe also that there is no obvious correlation between the surface forms predicted in Hahn (1991) and the backness of the following suffix.

As noted by a reviewer, if there is a phonological contrast, one would expect some roots to consistently trigger [+back] suffixes while other roots consistently trigger [–back] suffixes. The amount of variation, and the absence of any roots that always trigger [–back] suffixes across all nine speakers suggests a more complex state of affairs. Comparisons within each pair in Table 5 do, however, suggest some potential phonological differences—/jil/ always behaves like a [+back] vowel, while /pil/ does so only 79% of the time; /ʃilim/ behaves like a back vowel 82% of the time, while /ilim/ does so only 57% of the time. Given the variation in Table 5, it is necessary to examine patterns of suffix selection for each participant. Table 6 presents suffix selection counts for each speaker and root, demonstrating even more clearly that there is noteworthy between-speaker and between-lexeme variation. For example, Speakers 1, 2, 8, and 9 produce only [+back] suffixes after /ilim/, while speakers 4, 6, and 7 produce only [–back] suffixes after this root. Observe also variation for /ʧiʃ/, /pil/, and /ʃilim/, which always precede [+back] suffixes for some speakers, while these roots may precede [+back] or [–back] suffixes for some speakers. Observe also that the number of tokens produced by each speaker varies. Some speakers produced certain words multiple times, and although no repetitions were prompted, all repetitions were analyzed. In addition, some speakers used Russian words in place of the target Uyghur lexemes. For instance, no tokens occurred for Speaker 3 because she always produced the Russian word /kl^jej/ rather than /ʃilim/ ‘paste.’

Results in Table 6 suggest that the phonological behavior of these lexical items is speaker-dependent. For Speakers 8 and 9, there is no evidence for any phonological difference between these six roots. Comparisons between each near-minimal pair for Speaker 2 also fail to suggest any phonological distinctions—all pairs show the same basic tendency for roots to trigger [+back] suffixes. However, results for Speaker 1 and Speakers 3–7 suggest a different state of affairs for these speakers.

For Speakers 1 and 5, /ʧiʃ/ triggers [–back] suffixes more often than [+back] suffixes, in contrast to /qiʃ/, which always triggers [+back] suffixes. Despite a difference in suffix patterning here, previous research has argued that the dorsal consonants may trigger harmony, causing any root with a uvular to behave as phonologically [+back] (Hahn, 1991; Mayer & Major, 2018). As such, it is not possible to determine whether the vowel or the uvular consonant acts as the trigger of harmony here.

Speakers 3 and 4 produce [–back] suffixes after /pil/ in a majority of cases, but always produce [+back] suffixes after /jil/. Most intriguing, though, is the comparison between /ilim/ and /ʃilim/. Speakers 4 and 7 produce only [–back] suffixes after /ilim/ and only [+back] suffixes after /ʃilim/. Speaker 6 also produces only [–back] suffixes after /ilim/, but suffix backness is variable after /ʃilim/. These results suggest that these roots are treated differently by these speakers. In sum, counts of suffix backness support a possible contrast between the vowels in /pil/ and /jil/ for Speakers 3 and 4, and a possible contrast between the vowels in /ilim/ and /ʃilim/ for Speakers 4, 6, and 7.

If these patterns of suffix backness selection are evidence for contrast, acoustic differences should exist between these particular roots for these particular speakers. Before examining speaker-specific acoustic data, observe the aggregate patterns for F2 for each lexical root in Figure 1. In this violin plot, the vertical length and shape of each violin represent the distribution of F2 for each lexical root. Horizontal lines within each violin indicate median and interquartile range for each root. F2 values range from less than -1z to more than 2z. Impressionistically, some tokens of putative /i/ are consistent with a very back [ɯ], while others are consistent with a very front [i]. As seen below, most high unrounded vowels surface with acoustic qualities somewhere in between these two extremes. While the range of effects is similar to Hahn’s description of high unrounded vowel allophony, a more detailed picture is evident below. Regarding consonantal influence, the realization of these vowels is more affected by the following consonant than the preceding consonant, with a following lateral resulting in a more backed vowel, even in a root like /jil/, with a preceding palatal.7 In addition to the global distribution of F2, there is noteworthy variation between lexical roots. Observe that the highest F2 is associated with the root, /ʧiʃ/ ‘tooth’ while the lowest F2 is associated with /jil/ ‘year,’ /pil/ ‘elephant,’ and /ʃilim/ ‘paste.’

Figure 1 presents aggregate patterns of suffix backness selection. If a general contrast existed for all six lexical items for all speakers, then we would expect the leftmost roots to exhibit the highest percentage of [+back] suffixes, and the rightmost roots to exhibit the highest percentage of [–back] suffixes. No obvious phonetic correlation between F2 and suffix backness is evident in the aggregated data, suggesting the absence of any general between-speaker relationship between F2 of /i/ and suffix backness. This is likely due to the influence of consonants, in particular the following consonant. The roots with the lowest F2 values have a following lateral while the roots with the highest F2 values have a following alveopalatal.

By-speaker boxplots showing variation across the six lexical items are presented in Figure 2. Since no between-speaker comparisons are made below, F2 is presented in Hertz. Observe across all nine speakers the relative paucity of tokens from /qiʃ/ ‘winter’ and /ʧiʃ/ ‘tooth.’ Initial-syllable high vowels often elide before voiceless fricatives in the language. For some speakers, the high vowel was frequently elided (e.g., Speakers 3, 4, and 9; others have also reported devoicing in this context), while for others the vowel was produced with slightly more consistency (e.g., Speakers 7 and 8). Speaker age and gender are noted in all by-speaker plots.

In order to determine which vowels of the three near-minimal pairs exhibit significant F2 distinctions for each speaker, a model was constructed to predict F2 (Hz) from speaker, lexical root, and their interaction, along with random intercepts for speaker and root. From this model, post-hoc comparisons were implemented in the emmeans package (Lenth, Singmann, Love, Buerkner, & Herve, 2018) with Bonferroni correction for multiple comparisons. The comparisons noted above—/pil/ and /jil/ for Speakers 3 and 4, and /ilim/ and /ʃilim/ for Speakers 4, 6, and 7—were tested. The roots /ʧiʃ/ and /qiʃ/ were not considered due to the fact that /q/ actively triggers harmony in the language, as well as the frequent elision of these vowels before /ʃ/. If these pairings are truly contrastive for these speakers, then I predict an acoustic difference between these items, specifically higher F2 for those items more frequently triggering [–back] suffixes.

As for /pil/ and /jil/, the difference in F2 for Speaker 3 trends toward a significance, but in the opposite direction of predictions. If the vowels in /pil/ and /jil/ behave like front and back vowels respectively, one would expect F2 of the vowel in /pil/ to show higher F2. Yet for Speaker 3, F2 of /pil/ is actually lower than F2 of /jil/ [β = –299, z = –2.15, p = .03]. The acoustic realization of the vowels in these two roots does however conform to predictions for Speaker 4. For Speaker 4, /pil/ triggers [–back] suffixes 60% of the time, while /jil/ always triggers [+back] suffixes. Based on Table 7, the vowel in /pil/ is produced with around 250 Hz greater F2 than the vowel in /jil/, although this difference does not reach significance [β = 259, z = 153, p = .13]. Note that this phonetic difference runs counter to very general tendencies we would expect from coarticulation, since /p/ would likely depress F2 while /j/ would increase it, all else being equal. No other speakers produced /pil/ with a more anterior vowel quality than /jil/. In sum, there is potential evidence for a backness contrast between the vowels in /pil/ and /jil/ for Speaker 4, but not for Speaker 3.

Moving on to /ilim/ and /ʃilim/, recall that Speakers 4, 6, and 7 always produced [–back] suffixes after /ilim/. Speakers 4 and 7 always produced [+back] suffixes after /ʃilim/, and Speaker 6 did so 50% of the time. There is a significant difference between the vowels in /ilim/ and /ʃilim/ for Speaker 4, and the difference for Speaker 6 trends toward significance [Speaker 4: β = 398, z = 3.05, p = .002; Speaker 6: β = 301, z = 2.51, p = .01]. Finally, for Speaker 7 there is no meaningful acoustic difference between the vowels of /ilim/ and /ʃilim/ [β = 54, z = 0.44, p = .66]. Thus, the combination of acoustic data and suffix backness selection suggests a contrast between the vowels in /ilim/ and /ʃilim/ for Speaker 4, and possibly for Speaker 6, but not for Speaker 7.

The results above suggest that the status of the high unrounded vowels varies significantly between speakers. Speakers 8 and 9 appear to treat all high unrounded vowels equivalently, triggering [+back] suffixes across all roots tested despite the fact that many of these roots are produced with a relatively high F2. Speaker 4 however, shows a pattern that supports a phonological contrast for /ilim/ and /ʃilim/, which would be better transcribed as /ilim/ and /ʃɯlɯm/, since they trigger front and back suffixes, respectively, and are marked by significantly different F2. Data from Speaker 6 are also suggestive in this regard, although /ʃɯlɯm/ optionally triggers [–back] suffixes for this speaker, and the acoustic differences between the vowels of these two words only trend toward significance. Importantly, differences in F2 correspond to phonological patterning for these speakers—roots that trigger [–back] vowels are produced with higher F2. Other speakers, however, produced data that are consistent with previous descriptions, specifically claims of lexical idiosyncrasy. For instance, Speaker 7 produces [–back] suffixes after /ilim/ but [+back] suffixes after /ʃilim/ with no reliable acoustic differences for the vowels in these two roots. It is reasonable to tentatively conclude that for those speakers each lexical item is specified for a particular harmonic value independent of surface [i]-[ɯ]. If there is only one contrastive high unrounded vowel, then perhaps the significant variation in F2 is a byproduct of more freedom along the F2 dimension. The extent of variation seen in Figure 1 is potentially consistent with a general lack of and /i/-/ɯ/ contrast, although this variation appears slightly less rampant in the by-speaker plots in Figure 2. As an example, even for Speaker 4, whose production data best support a contrast between /i/ and /ɯ/, /qiʃ/ and /ʧiʃ/ have high F2 but behave like [+back] roots.8

Evidence thus far points to three types of systems: (1) those where /i/ and /ɯ/ contrast, (2) those where [i] and [ɯ] do not contrast, and roots exist with idiosyncratic harmonic specifications independent of surface [i]-[ɯ], and (3) those where [i] and [ɯ] do not contrast and all suffixes after [i] and [ɯ] are always [+back]. As a final note before moving on, if one posits that there is only one high unrounded vowel phoneme in the language, it is not at all clear that it should be /i/ (with a default surface quality of [ɪ]) and not /ɯ/. Central and back variants are much more common, occur with a much freer distribution than [i] (see also Hahn, 1986 for discussion), and [i]-[ɯ] most often trigger [+back] suffixes. As support for relative backness of most high unrounded vowels, observe the vowel plot in Figure 3, which shows mean F2 of the vowels in these roots in the context of the larger vowel inventory. Observe that F2 for most of these six roots falls in between the front and back rounded vowels. Although Hahn (1991) suggests the default realization of /i/ is [ɪ], for these roots the default is more likely a more central vowel, like [ə] or [ɘ].9 Additionally, if one posits /ɯ/ rather than /i/, then the tendency to trigger [+back] suffixes receives a more straightforward explanation, and is no less consistent with the relative backness of these roots in Figure 3.

This subsection investigates whether or not high unrounded vowels alternate for harmony in target positions (e.g., in suffixes and raised vowel contexts). To that end, the subsection is broken up into two parts. The first part examines results from the short-distance condition, where a single high vowel intervenes between a trigger and alternating target low vowel. The second examines results from the long-distance condition, where three high vowels intervene between a harmony trigger and alternating low vowel target.