The phonetics and phonology of Uspanteko (Mayan)

3.1 Glottalisation

Glottalised consonants are the most studied aspect of the phonetics of Mayan languages (e.g. Burnett-Deas, 2009; Campbell, 1973; Frazier, 2009a; Herrera Zendejas, 2014; Kingston, 1984; Kuang, 2019; Pinkerton, 1986; Russell, 1997; Shosted, 2009; Wagner & Baker-Smemoe, 2013; see also Bennett, 2016; England & Baird, 2017). One of main takeaways of this literature, as well as the numerous descriptive grammars that consider glottalisation in detail, is that there is extensive variation in the realisation of glottalisation across languages, dialects, speakers, and places of articulation in the Mayan family. Much of this variation is localised in the glottalised labial /$$\textipa{\r*{\!b}}$$/ and uvular /q^ʔ/ stops, which vary between implosive and ejective realisations, /$$\textipa{\r*{\!b}}$$/ tending to be implosive and /q^ʔ/ tending to be ejective. When realised as implosive, both /$$\textipa{\r*{\!b}}$$/ and /q^ʔ/ are typically voiceless [$$\textipa{\r*{\!b}}$$ $$\textipa{\r*{\!{G}}}$$] in Uspanteko, though /$$\textipa{\r*{\!b}}$$/ is sometimes voiced intervocalically (Sections 3.2 and 3.3).

There are some phonetic regularities across glottalised consonants in Uspanteko, be they realised as ejective or implosive. In particular, glottalisation tends to be marked by creakiness, or other kinds of laryngealized non-modal phonation, on adjacent vowels and sonorant consonants (Bennett, 2016). This is evidenced in pairs like [ˈkaːχ] ‘sky’ and [ˈk^ʔaːm] ‘cord, twine’ (Figure 3). In [ˈk^ʔaːm] ‘cord, twine’, the onset of the vowel following [k^ʔ] shows weak and irregular voicing corresponding to creaky voice (Gordon & Ladefoged, 2001; Keating et al., 2015). This creakiness is noticeably absent on the vowel following [k] in [ˈkaːχ] ‘sky’.³

Relatedly, f0 is lowered following [k^ʔ]: the spacing between glottal pulses is wider at the onset of the vowel than at the midpoint, indicating reduced f0 at the CV transition (this is easiest to see in the spectrogram). In contrast, f0 is relatively unperturbed following plain [k]: the spacing between glottal pulses is regular and close at the CV transition, and similar to the spacing at vowel midpoint. Lastly, the amplitude rise time on the vowel is longer following [k^ʔ] than [k], as can be seen by comparing amplitude at vowel onset versus midpoint (e.g. Russell, 1997; Wright et al., 2002). All of these phonetic effects are consequences of coarticulatory creakiness on the vowel, reflecting some degree of additional glottal constriction.

Along with coarticulatory creakiness, [k k^ʔ] are also distinguished phonetically by the intensity and quality of their release phases: the burst for [k^ʔ] is louder in Figure 3, and is followed by a period of silence corresponding to sustained glottal closure after the oral stop release. Additionally, the release phase for [k^ʔ] is overall longer than the release phase for [k] (about 71 vs. 38 ms), though the release noise associated with [k^ʔ] is perhaps shorter (about 28 vs. 38 ms).

The phonetic differences described above reliably distinguish plain versus ejective stops in Uspanteko, at least in careful speech (we discuss implosives below in Sections 3.2 and 3.3). In spontaneous speech, some of these phonetic cues may be less salient (e.g. the intensity of release bursts and duration of release phases may be reduced in ejectives). Given the extensive phonetic variability associated with glottalised stops in Mayan languages, it seems worthwhile to verify the reliability of these phonetic differences across a larger sample of spoken Uspanteko, from a range of speech styles and genres.

Another sporadic difference between plain stops and ejectives is that ejectives are sometimes realised with a brief interval of voicing following the release of glottal constriction (Figure 4). Brief voicing after ejective release is most often observed in word-final position.

We suspect that these short intervals of voicing at stop release owe to mechanical, aerodynamic factors (Westbury & Keating, 1986). If the glottis is tightly constricted during the production of an ejective, air pressure may build-up below the glottis during stop closure (Demolin, 2011). When the oral constriction is released, oral air pressure will begin to drop, provided that the glottis remains closed (see Pinkerton, 1986 and Kuang, 2019 for oral pressure traces illustrating exactly this phenomenon in Mayan languages). At this point, sub-glottal air pressure may be substantially higher than oral air pressure, especially if there is a long lag between the release of the oral constriction and the release of the glottal constriction, as in Figure 4. When the glottal constriction is released, a large disparity between sub-glottal air pressure (high) and oral air pressure (low) will drive rapid airflow through the glottis, encouraging brief passive voicing. In the absence of actual aerodynamic data this proposal remains speculative, but nonetheless strikes us as a plausible explanation for the occasional short bursts of voicing found at ejective release in Uspanteko.⁴

To reiterate, these voicing bursts only sometimes occur after glottal release in ejectives. Ejectives are frequently produced without voiced intervals following release, even in final position (Figure 5). Voiced intervals are never observed for plain stops (Figure 5), which lack the sub-glottal pressure build-up which may be responsible for occasional, transient voicing after glottal release in ejectives.

3.2 Uvular stops /q q^ʔ/

The uvular stops /q q^ʔ/ are sometimes realised with frication noise during the transition from a preceding vowel (Figure 6). This is particularly true for plain /q/, which is commonly realised with a noisy, affricate-like release in coda position as well (Figure 6). Even in pre-vocalic position, plain /k q/ are often produced with fairly long and noisy releases (Figures 6 and 7; see also Figure 33 below).⁵

These noisy transitions probably reflect the fact that the dorsum is a slow-moving articulator: the formation and release of dorsal stops may involve extended phases in which oral constriction is incomplete, but narrow enough to produce frication noise. Glottalised [k^ʔ q^ʔ] lack these noisy transitions, likely because glottal closure during [k^ʔ q^ʔ] inhibits frication by reducing airflow through the oral tract.⁶

As noted above, the glottalised uvular /q^ʔ/ is often realised as ejective [q^ʔ], but also as the voiceless implosive [$$\textipa{\r*{\!{G}}}$$]. This variation can occur in the speech of a single speaker, and seems unconditioned in that [q^ʔ] and [$$\textipa{\r*{\!{G}}}$$] variants occur in essentially the same phonetic environments (Figure 8).⁷

Ejective [q^ʔ] is produced with a clear release burst, followed by a period of silence corresponding to glottal closure. Implosive [$$\textipa{\r*{\!{G}}}$$] lacks an egressive release burst (Clements & Osu, 2002). Both [q^ʔ] and [$$\textipa{\r*{\!{G}}}$$] allophones can occur with coarticulatory creakiness on adjacent vowels and sonorants.

When implosive, the glottalised uvular /q^ʔ/ → [$$\textipa{\r*{\!{G}}}$$] is sometimes auditorily quite similar to [ʔ]. Figure 9 illustrates: the formant transitions out of /q^ʔ/ → [$$\textipa{\r*{\!{G}}}$$] and into the following vowel are relatively flat, as found for glottal stop (e.g. Borroff, 2007 and Figure 18 below); this contrasts with the more dynamic formant movements (particularly F2 lowering) observed during the transition from the vowel into the following uvulars [χ] and [q^ʔ] (e.g. Alwan, 1986, Reetz & Jongman, 2011:Ch. 10).

In certain dialects of other K’ichean-branch Mayan languages, historical /q^ʔ/ is described as pharyngeal or pharyngealised (e.g. England, 2001; Larsen, 1988; Patal Majzul et al., 2000:25-6). These innovative pronunciations may be related to the [ʔ]-like realisation of implosive /q^ʔ/ → [$$\textipa{\r*{\!{G}}}$$] seen in Figure 9.

Ejective realisations of /q^ʔ/ may also have flat, steady formant transitions into a following vowel, as in for example, Figure 14 below. However, ejective [q^ʔ] is more easily distinguished from [ʔ] by its clear release burst, which may include acoustic cues to its uvular articulation (e.g. the duration, intensity, and spectral shape of the burst; Alwan, 1986; Cho & Ladefoged, 1999; Raphael, 2005).

The flat formant transitions sometimes found after ejective [q^ʔ] and implosive [$$\textipa{\r*{\!{G}}}$$] probably reflect the fact that the tongue body can move towards the posture for a following vowel as soon as the uvular constriction is released, before the release of the glottal constriction. Formant transitions associated with tongue body movement may not be audible (or visible) if produced with simultaneous glottal closure, since a tight glottal constriction will sharply restrict airflow through the vocal tract.

3.3 Implosive /$$\textipa{\r*{\!b}}$$/

The glottalised labial in Uspanteko is normally produced as a voiceless implosive [$$\textipa{\r*{\!b}}$$] (Figure 10).

Implosive [$$\textipa{\r*{\!b}}$$] usually lacks anything resembling a release burst, though sometimes a clear negative impulse can be seen in the waveform at the transition between stop closure and a following vowel. This negative impulse plausibly corresponds to ingressive airflow associated with implosion at the release of the oral constriction. A negative impulse of this type can be seen in Figure 10 (right panel), and in Figure 9 above for implosive /q^ʔ/ → [$$\textipa{\r*{\!{G}}}$$] (left panel).

As with all other glottalised stops in Uspanteko, the glottalised labial /$$\textipa{\r*{\!b}}$$/ may condition creakiness on adjacent vowels and sonorants (Figure 11, and also Figure 6 above).

Additionally, brief periods of voicing can sometimes be found at or just before the release of implosive [$$\textipa{\r*{\!b}}$$] (Figures 10-12). We again assume that this transitory voicing is a mechanical by-product of increased transglottal airflow, which occurs when the glottal closure for the implosive is released, and the compressed air below the glottis begins to flow outward again.⁸

Implosive /$$\textipa{\r*{\!b}}$$/ is sometimes realised as ejective [p^ʔ], particularly in word-final position (Figure 13). Even in word-final position, ejective [p^ʔ] realisations of the glottalised labial vary with implosive [$$\textipa{\r*{\!b}}$$] (Figure 14).

Glottalised /$$\textipa{\r*{\!b}}$$/ is occasionally produced as something closer to a plain voiced stop [b] (Figure 15). This tends to occur between voiced sounds, particularly voiced consonants, and is more likely in rapid or casual speech.

Plain [b] renditions of /$$\textipa{\r*{\!b}}$$/ are relatively infrequent, and seem best analysed as an example of lenition or hypoarticulation. Indeed, [b]-like realisations of /$$\textipa{\r*{\!b}}$$/ seem to be on a cline of reduction that also includes voiced [ɓ] variants produced in essentially the same environments during running speech (Figure 16). So while Figure 15 includes a weak [b]-like release burst absent from Figure 16, both examples show weak, irregular voicing during stop closure, suggesting a similar laryngeal articulation.

Like /q^ʔ/, /$$\textipa{\r*{\!b}}$$/ is sometimes auditorily similar to [ʔ] (Section 3.2). This can be seen in Figure 17: again, the formant transitions from /$$\textipa{\r*{\!b}}$$/ to the following vowel are relatively flat, resembling the transitions for [ʔ].

In a number of Mayan languages both /q^ʔ/ and /$$\textipa{\r*{\!b}}$$/ have sometimes merged with [ʔ], at least sporadically: Comalapa Kaqchikel, for example, has [ˈʃʔe] and [nu.ˈq^ʔaʔ] for historical [ˈʃ$$\textipa{\r*{\!b}}$$e] ‘(s)he went’ and [nu.ˈq^ʔa$$\textipa{\r*{\!b}}$$] ‘my hand’ (e.g. Chacach Cutzal, 1990; Patal Majzul et al., 2000:25–6; García Matzar & Rodríguez Guaján, 1997, p. 30). We speculate that mergers between /q^ʔ $$\textipa{\r*{\!b}}$$/ and [ʔ] may have been facilitated by [ʔ]-like realisations of both /q^ʔ/ and /$$\textipa{\r*{\!b}}$$/ (Figures 9 and 17).

3.4 Glottal stop /ʔ/

Glottal stop /ʔ/ is phonemic in Uspanteko, as evidenced by pairs like (2).⁹

Along with phonemic glottal stop, an epenthetic glottal stop also occurs at the beginning of words which are underlyingly vowel-initial (see Bennett, 2016, 2018; England & Baird, 2017). Epenthesis can be diagnosed by alternations like those in (3), which show that word-initial glottal stops often disappear under prefixation. Figure 18 illustrates both phonemic /ʔ/ and epenthetic word-initial [ʔ].

Some word-initial phonetic glottal stops do not alternate with zero, suggesting that they are underlying and phonemic rather than inserted. For example, aab’ [ˈʔaː$$\textipa{\r*{\!b}}$$] ‘hammock’ retains its initial glottal stop under possession, as in in’aab’ [ʔin.ˈʔaː$$\textipa{\r*{\!b}}$$] ‘my hammock’. This implies that [ʔ] is underlying rather than inserted in this noun.

Supporting evidence for this claim comes from prefixal allomorphy. Possessive prefixes take different forms when attaching to vowel-initial versus consonant-initial stems. The noun aab’ [ˈʔaː$$\textipa{\r*{\!b}}$$] ‘hammock’ takes possessive allomorphs like [ʔin-] ‘my’ that otherwise only occur with consonant-initial stems, not vowel-initial stems (e.g. inb’aatz’ [ʔin-ˈ$$\textipa{\r*{\!b}}$$aː$$\stackrel{{\frown}}{\mathrm{ts}}$$^ʔ] ‘my thread’ vs. [ˈw-íʃim] ‘my corn’ (3)). This is consistent with treating the non-alternating [ʔ] in aab’ [ˈʔaː$$\textipa{\r*{\!b}}$$] ‘hammock’ as underlying. See Bennett (2016, 2018) and Kaufman (2015) for more discussion.

Word-initial glottal stop can be realised as a full stop and/or creakiness on adjacent sonorants, as in Figure 18 and other examples in Section 3.3 above. Still, word-initial glottal stop is not always phonetically salient, especially in running speech. However, epenthetic word-initial glottal stop is sometimes retained under prefixation, and in these cases the glottal stop is quite clear (Figure 19).¹⁰

After consonants, glottal stop is typically realised with full closure, often with creakiness on neighbouring vowels and sonorants, as in Figure 19. Between vowels, glottal stop may be realised as a full stop, but is more commonly realised as creakiness on the vowels themselves, or as an interval of creaky voicing with significantly reduced amplitude reflecting glottal constriction between the two vowels (Figure 20). Note again that creakiness often involves lowering of f0, as is apparent from the wide spacing of voicing striations in the spectrograms in Figure 20, along with reduced amplitude.

Glottal stop is commonly found in two other environments, /Vʔ#/ and /VʔC#/. Three different phonetic outcomes for /ʔ/ are typical in these environments. First, /ʔ/ may be realised as a true stop, as in Figures 18, 19 and 21. In final position, this variant of /ʔ/ includes a clear release burst, along with possible creaky voice on the preceding vowel (Figure 21; see also Figure 11).

Second, /ʔ/ may be realised primarily as extensive creakiness on the preceding vowel, and/or any following sonorant consonant. This is very common for /Vːʔ#/ and /V(ː)ʔC#/ sequences, as in Figure 22 (see also Figure 2 above). The glottal stop itself may or may not have an audible release in these cases.

Impressionistically, extensive creakiness of this sort is less common following short vowels in final position, /Vʔ#/. It may be that glottal stop is phonetically and phonologically more like a vowel feature in /Vːʔ#/ and /V(ː)ʔC#/ sequences, and more like an independent consonant in /Vʔ#/ sequences (see also Sections 5.1.5 and 5.2 on the fact that /ʔC#/ coda clusters only occur in stressed final syllables, much like long vowels). If this is correct, it seems plausible that /ʔ/ is realised as a vowel feature rather than a consonant in /V(ː)ʔC#/ due to a restriction on the size, weight, and/or composition of syllable rimes in Uspanteko. For further discussion of the consonantal versus featural status of [ʔ] in Mayan languages, see Bennett and Henderson (2013), Bennett (2016, 2018), England and Baird (2017) and references there. This is a clear topic for future investigation.

Lastly, /Vʔ#/ and /VʔC#/ sequences may be realised with a ‘rearticulated’ or ‘broken’ vowel. Auditorily, these sound like a modal-voiced vowel that has been interrupted by a glottal stop. Phonetically, the glottal interruption is usually just creaky voice, though true stop realisations do occur as well (Figure 23).

There are some cases where laryngealization associated with /ʔ/ is audible, but not readily apparent as creak in the corresponding audio recording (Figure 24). Such weak laryngealization may involve dips in intensity and/or f0, without any of the other potential correlates of non-modal phonation (see Gerfen & Baker, 2005; Keating et al., 2015).

Much of what we have said about the phonetics of [ʔ] concerns stressed syllables. In unstressed syllables, [ʔ] may be substantially weakened, sometimes to the point of apparent deletion (Figure 25).

We are unsure whether coda glottal stop is phonologically deleted in unstressed syllables, or simply phonetically reduced. Understanding the phonetics and phonology of word-medial [ʔ], especially in coda position and in unstressed syllables, is an area of future research.

On the phonetics of glottal stop in other Mayan languages, see Frazier (2009a,b, 2013), Baird (2011), Baird and Francisco Pascual (2011), Bennett (2016), England and Baird (2017), Sobrino Gómez (2018), and references there. On glottal variability more generally, see Borroff (2005, 2007), Garellek (2013, 2014), Keating et al. (2015), Whalen et al. (2016), Davidson (2021), Garellek et al. (to appear), and references there.

3.5 Other phonetic observations

For reasons of space, we do not discuss the phonetics of other consonants or consonant types in Uspanteko in this paper. On the phonetics of rhotics in Mayan languages, see Romero (2009), Bennett (2016), England and Baird (2017), and references there, as well as Solé (2002). On the phonetics of the dorsal fricative /x/ ∼ /χ/, see Redmon & Jongman (2018), whose phonetic observations about such fricatives accurately characterise the phonetic properties of these sounds in Uspanteko and K’ichean-branch Mayan languages.

4.1 Vowel length and vowel duration

The phonemic vowel system /a(ː) e(ː) i(ː) o(ː) u(ː)/ is common in the Mayan family (Bennett, 2016; England & Baird, 2017). Still, the phonetic realisation of this vowel inventory varies even among K’ichean-branch Mayan languages. For example, in closely related K’ichean languages like Tz’utujiil (Dayley, 1985), K’iche’ (Baird, 2010), and Q’eqchi’ (Berinstein, 1991), long vowels are about twice as long as their short counterparts (see also Herrera Zendejas, 2014; Sobrino Gómez, 2010 on other Mayan languages). But in Uspanteko — a language which may have diverged from other K’ichean languages fairly early on (e.g. Campbell, 1977) — long vowels are not as widely separated from short vowels in terms of their duration.

Figure 27 reports average vowel duration in Uspanteko, grouped by vowel length and accent type (see Section 2 on data collection for this analysis). Vowels longer than 300 ms were excluded from this analysis (7 tokens, 0.3% of the data). Here we focus on how vowel length and stress influence duration, and return to the effect of lexical high tone on duration in Section 5.3.4. In this data set, phonemic long vowels, which are always stressed (Section 5), have an average duration of 167 ms. In contrast, stressed short vowels have an average duration of 105 ms. Duration thus clearly distinguishes phonemic long and short vowels (p < 0.001, by two-sided t-test).

Unstressed short vowels are reduced relative to stressed short vowels, with an average duration of 74 ms (p < 0.001). This is particularly true for short vowels in post-tonic, unstressed syllables (e.g. ínp$$\underline{i}$$x [ˈʔímpiʃ] ‘my tomato’), which average 60 ms in our data compared to the 75 ms average of pre-tonic, unstressed short vowels (e.g. k$$\underline{i}$$naq’ [k$$\underline{i}$$.ˈnaq^ʔ] ‘bean’; p < 0.05). For related discussion on pre-tonic versus post-tonic position, see Bennett et al. (ms.).

4.1.1 Vowel duration in a Mayan context

Vowel length contrasts in Mayan involve durational differences of various sizes. The surveys in Bennett (2016) and England and Baird (2017) report durational ratios ranging from 1.25:1 to 2:1 for long versus short vowels in Mayan languages. Proportionally, stressed long vowels in Uspanteko seem to be about 62% longer than stressed short vowels, at least in non-tonal, word-final syllables (Section 5). This amounts to a ratio of about 1.6:1, somewhere in the middle of the durational ranges reported for other Mayan languages.

Berinstein (1979) examines the correlates of stress in two other K’ichean languages, Kaqchikel and Q’eqchi’. She reports that vowel duration is a correlate of stress in Kaqchikel, but not in Q’eqchi’. She attributes this difference to the fact that Q’eqchi’ has true vowel length contrasts, which may inhibit the use of duration as a cue to stress. Kaqchikel makes use of centralisation (‘tense-lax’) contrasts in its vowel system instead, such as /a e i o u/ versus /ə ɛ ɪ ɔ ʊ/, though such contrasts did develop historically from earlier length contrasts (e.g. Bennett, 2019; Campbell, 1977; see also Vogel et al., 2016; Lunden et al., 2017; van Heuven & Turk, 2021 for critical discussion).

Uspanteko does not quite follow the predictions of Berinstein’s (1979) work, as stressed short vowels do seem to be longer than unstressed short vowels (there are no unstressed long vowels, so nothing can be said about cues to stress for long vowels specifically; Section 5). However, the phonetic differences between stressed and unstressed short vowels may reflect the fact that unstressed short vowels are often heavily reduced in Uspanteko, sometimes to the point of deletion (Figure 28; see the right panel of Figure 20 for another example, and Bennett et al. (ms). for extensive discussion).

4.2 Vowel quality as a function of vowel length and stress

It common for length contrasts to be augmented by a quality difference, with short vowels being more centralised than their long vowel counterparts. This correlation between length and centralisation has been observed in a number of Mayan languages, belonging to several different major subgroups (e.g. Baird, 2010; Barrett, 1999; Dayley, 1985; Du Bois, 1981; Edmonson, 1988; England, 1983; see also Bennett, 2016; England & Baird, 2017).

However, it has also been reported that some Mayan languages implement vowel length contrasts almost entirely by means of duration, without any significant differences in quality between long and short vowels (e.g. England, 2001; England & Baird, 2017). Uspanteko appears to be a language of this type, with only very limited centralisation of most short vowels (Figures 29 and 30). The phonetic results presented here are in keeping with our own auditory impressions as fieldworkers: the qualities of short /e i o u/ are very similar to the qualities of long /eː iː oː uː/, while short /a/ shows a tendency to raise and/or centralise relative to long /aː/.

To assess vowel quality for short and long vowels, formants were measured by averaging values for F1, F2, and F3 over the middle 20% of each vowel, using a custom Praat script (Boersma & Weenink, 2020). We used a vowel-intrinsic normalisation method — F3-normalisation (F1/F3 and F2/F3 for each vowel) — in order to pool formant measurements across speakers (Monahan & Idsardi, 2010 and references there). The F3-normalised data is highly correlated with the output of two alternative formant normalisation methods, Lobanov’s z-score normalisation (F1: r = 0.95, F2: r = 0.95; Lobanov, 1971) and Barreda-Neary log-additive regression normalisation (F1: r = 0.96, F2: r = 0.96; Barreda & Nearey, 2018) (see too Adank et al., 2004, Hillenbrand et al., 1995, Johnson, 2020). It is also highly correlated with the original data as measured in Hz (F1: r = 0.96, F2: r = 0.94), Bark (F1: r = 0.96, F2: r = 0.95), and ERB (F1: r = 0.96, F2: r = 0.95), suggesting that our speakers had similar vowel spaces and vocal tract lengths. To remove outliers and potential measurement errors, F3-normalised formant values were converted to z-scores, and tokens were excluded from analysis if they had z-scores for F1/F3 or F2/F3 which were greater than 2.5 z-units from the mean value for vowels of the same quality (pooling over length, stress, and tone). This procedure trimmed 88 tokens (3.6% of the data; Figures 29 and 30).

Overall, vowel quality does not vary widely with differences in stress or phonemic vowel length. Long, stressed vowels [ˈVː] are somewhat more peripheral and more tightly clustered than short vowels, which show a tendency towards slight centralisation regardless of stress. The short low vowel /a/ is often quite audibly centralised, sounding something like [ɐ] or [ə] when unstressed (Bennett & Henderson, 2013). For some speakers, stressed short /a/ is also quite centralised, so that for example, pach /pa$$\textipa{\t{tS}}$$/ ‘friend, partner’ is pronounced as [ˈpa$$\textipa{\t{tS}}$$] by many speakers, but as something closer to [ˈpɐ$$\textipa{\t{tS}}$$] or [ˈpə$$\textipa{\t{tS}}$$] by others.

5.1 Syllable structure

Syllable structure in Uspanteko is typically CV(C), with a number of important caveats and exceptions we outline below. An example of this basic template is shown in (5).

5.1.6 Syncope

Syncope, which is widespread in Uspanteko, variably targets unstressed short vowels in specific positions (Bennett & Henderson, 2013, Bennett et al. ms.). When stress is final, syncope frequently targets the immediately pre-tonic syllable (17).

• 

When stress is penultimate, syncope instead targets the post-tonic syllable (18).

• 

Post-tonic syncope in kúmb’al [ˈkúm$$\textipa{\r*{\!b}}$$al] ∼ [kúm$$\textipa{\r*{\!b}}$$l] ‘medicine’ can be seen in Figure. 15 above.

Bennett and Henderson (2013) and Bennett et al. (ms.) analyse vowel deletion in terms of foot structure (Section 5.2): under the assumption that final stress involves an iamb […($$\sigma '\sigma $$)], and penultimate stress involves a trochee […$$\stackrel{\check{'}}{\sigma }\sigma $$], then syncope can be said to consistently target the weak branch of the foot.

Syncope appears to create complex consonant clusters (19), Figure 31.

• 

However, Bennett et al. (ms.) argue that syncope does not involve actual vowel deletion. Instead, it is a species of extreme phonetic vowel reduction, which leads to acoustic forms that sound like the vowel has been deleted (on vowel reduction, see Sections 4.1 and 4.1.1 above). Their evidence for this claim comes, in part, from the observation that weak vocal fold vibration associated with ‘deleted’ vowels is sometimes observable in electroglottographic recordings of Uspanteko even when those vowels are seemingly absent from the acoustic recording. Phonologically, then, the consonant clusters derived by syncope (19) may in fact be [CVC] sequences after all, but with a highly reduced vowel that is essentially inaudible, though weakly articulated.

5.2 Stress

Primary stress occurs by default on the rightmost syllable of the word in Uspanteko. There is no evidence of secondary stress. In these respects, the Uspanteko stress system resembles the stress systems of related K’ichean-branch Mayan languages, which also strongly tend towards fixed final stress (apart from loanwords, and a few lexical and morphological exceptions; e.g. Baird, 2014b, Bennett, 2016, Berinstein, 1979, Henderson, 2012, England & Baird, 2017, Can Pixabaj, 2017, and references there).

There is also phonological evidence for final stress: long vowels, and thus vowel-length contrasts, are limited to word-final stressed syllables (Bennett et al., 2022). This again has parallels in other K’ichean-branch Mayan languages (e.g. Bennett, 2016, 2019; Can Pixabaj, 2017; Dayley, 1985). This restriction is clearly illustrated by vowel length alternations that occur under suffixation: when an underlying lexical long vowel occurs outside the stressed final syllable, it is systematically shortened (21).

Major intonational contours also tend to align with final syllables, again implying that these are prominent positions (e.g. Hayes, 1995; for Mayan languages, Gussenhoven & Teeuw, 2008; Baird, 2014a; Bennett, 2016; England & Baird, 2017; Adell, 2019; DiCanio & Bennett, 2021). Lastly, native speakers do have the intuition that stressed final syllables are more prominent than other syllables in the word.

Stress can also occur on the penultimate syllable, but only in words bearing lexical high tone, which we now turn to.

5.3 Lexical tone

5.3.2 The synchronic system of lexical tone

Long vowels only occur in word-final stressed syllables (Section 5.2). In final position, stressed long vowels may be toneless, or may bear a high tone [h] (22). As Figure 32 illustrates, tonal long vowels have higher pitch than non-tonal long vowels, and may also have more dramatic pitch excursions (rises and falls). The slightly raised pitch on the final stressed syllable of non-tonal qaxoot [qa.ˈʃoːt] ‘our comal’ (Figure 32, lower-left) arguably represents intonational prominence rather than any f0 effects associated with stress as such; see Bennett et al. (2022) and Section 5.3.3 below for discussion.

There are very few minimal pairs for tone on long vowels in Uspanteko. Further, as a result of lexical variation in the Uspanteko community, some speakers do not have any minimal pairs at all for tone on long vowels (Bennett et al., 2022). Still, there are various near-minimal pairs like Figure 32 which show lexically-specific pitch differences that cannot be reduced to conditioning by the segmental or morphological environment (see Snider, 2014; Herrera Zendejas, 2014:Ch. 9 for related discussion). Such pairs firmly establish that Uspanteko has a system of lexical tone on long vowels, even if the functional load of tonal contrasts is low (Hyman, 2006, 2009).

In words with final short vowels, the interaction of tone and stress is more complex. When the final vowel is short, tone can only occur on the penultimate syllable. Stress then retracts to coincide with tone (22).

• 

Stress retraction is evident in alternations like Figure 33: the position of stress is associated with greater duration and intensity; and high tone (right panel) is associated with raised f0. (See Figures 31, 38 for other similar examples.)

Bennett and Henderson (2013) analyse the distribution of tone in Uspanteko by appealing to moras, a prosodic unit below the level of the syllable which distinguishes short vowels (1 mora) from long vowels (2 moras) (Trubetzkoy, 1939, Hyman, 1985, etc.). They propose that high tone [h] only occurs on the penultimate mora of the word. It follows directly that [h] tone will occur on final long vowels (which contain two moras), or will occur on the penultimate vowel when the final vowel is short.

Tone is associated with morphology in Uspanteko, as documented by Can Pixabaj (2007), Bennett and Henderson (2013), and Bennett et al. (2022). For example, the plural suffix /-V$$\textipa{\r*{\!b}}$$/ introduces high tone on the preceding syllable (23) (capital /V/ here stands for a vowel of varying quality; see also Figure 33 for tone introduced by a possessive prefix).¹³

• 

Note that penultimate stress does not licence long vowels (23), unlike final stress (Section 5.2). If stress is analysed with reference to foot structure in Uspanteko, this may reflect well-known quantitative asymmetries between iambic and trochaic feet: final, iambic stress allows both […(LˈL)] and […(LˈH)] footing; while penultimate, trochaic stress allows only […(ˈLL)] footing (Hayes, 1995; this assumes that coda consonants do not contribute to syllable weight in Uspanteko, following Bennett & Henderson, 2013).

5.3.3 Phonetic analysis of lexical tone

Pitch contours on target items were analysed in Praat (Boersma & Weenink, 2020). The recordings were downsampled to 16 kHz, then pitch values were automatically extracted in Hz with a script specifying by-speaker pitch ranges following the recommendations of De Looze & Rauzy (2009) and Evanini et al. (2011). Along with measurements of mean and maximum f0 for each vowel, time-normalised pitch measurements were produced by averaging pitch values over 1/7 intervals of the duration of each vowel.

F0 measurements were z-score normalised for each speaker, so that the data could be pooled for analysis. Pitch measurements more than 2.5 z-units away from each speaker’s mean were treated as outliers and removed from the data. This resulted in the elimination of 0.4% of the mean f0 measurements (8 tokens), 1.2% of the maximum f0 measurements (26 tokens), and 1.2% of the time-normalised f0 interval measurements (175 measurements). For all three measures, the z-score normalised data was closely correlated with the results of two other normalisation methods: semitone transformation relative to each speaker’s mean pitch in Hz (Zhang, 2019), and range normalisation using 2% and 98% estimates of floor and ceiling values in semitones (Bardiaux & Mertens, 2014) (lowest correlation r = 0.96; see also Ladd, 2008:192–202).

In our wordlist data, tonal vowels are not sharply distinguished by mean or maximum f0 (Figures 34 and 35). This is particularly true for short tonal vowels [ˈ$$\acute{\mathrm{V}}$$], which have f0 distributions which largely overlap those of non-tonal, stressed short vowels [ˈV] (that said, f0 differences between short [ˈ$$\acute{\mathrm{V}}$$] vs. [ˈV] are statistically significant by two-sided t-test: mean f0: Δ = 0.15 z-units, p < 0.05; maximum f0: Δ = 0.17 z-units, p < 0.05).¹⁴ (There are just 15 tonal long vowels in this dataset, so we avoid making any quantitative claims about the phonetics of tone on long vowels here, though we do include this data in plots and statistical models.)

The modest f0 increase associated with tone in Figures 34 and 35 arguably reflects the fact that intonational prominences may also occur on toneless vowels, leading to some degree of phonetic neutralisation between tonal and non-tonal syllables (Bennett et al., 2019, 2022). Rising (L)H% contours often occur utterance-finally in K’ichean-branch Mayan languages, including Uspanteko (e.g. Baird, 2014a,b; Bennett, 2016; Bennett et al., 2022; Berinstein, 1991; England & Baird, 2017). These intonational rises occur in both declarative sentences and questions, and may be associated with the right-edge of clause-sized prosodic units (e.g. Intonational Phrases, or IPs; for detailed discussion of K’iche’, see Nielsen, 2005; Henderson, 2012; Burdin et al., 2015; Baird, 2018, and citations there; for more distantly related Mayan languages, see e.g. Clemens, 2021; Royer, 2022).

Final intonational rises are common in our wordlist data, likely as a result of eliciting words in a repeated frame sentence. In such tasks, the target word is pragmatically or metalinguistically focussed, and may be produced with intonation characteristic of a full, independent utterance or IP (Himmelmann, 2006; Himmelmann & Ladd, 2008; Hyman, 2014; Jun & Fletcher, 2014; Pike, 1948; Snider, 2014; Yu, 2014). A particularly dramatic example is shown in Figure 36.

These intonational rises plausibly obscure f0 differences between tonal and non-tonal vowels in Uspanteko, both in our specific wordlist data, and in the language more generally (Bennett et al., 2022).

Further evidence for an IP/utterance-final H% boundary tone in Uspanteko comes from the analysis of f0 on unstressed vowels. Fig. 37 shows f0 trajectories across vowels, grouped by stress, length, tone, and position in the word. These plots again show that f0 is somewhat higher for tonal than non-tonal vowels, though not by a wide margin. The key observation here is that unstressed short vowels $$\left[\breve{\mathrm{V}}\right]$$ have a substantially raised pitch in final syllables (i.e. in words with penultimate accent, e.g. kútz’$$\underline{i}$$j [ˈkú.$$\stackrel{{\frown}}{\mathrm{ts}}{\textipa{?}} \underline{i}$$χ] ‘flower’). This is consistent with the presence of an H% boundary tone in final position, corresponding to the large final rise in examples like Figure 36. An example illustrating an H% target on an unstressed, final vowel is shown in Figure 38. (The rising f0 on tonal and non-tonal long vowels in Figure 37 may also corroborate the presence of an H% boundary tone in much of our data.)

To assess the effect size of lexical tone on Hz in our study, we fit a linear-mixed effects model to predict mean vowel f0 using the lme4 package in R (Bates et al., 2020). This model included fixed effects for stress (ˈ$$\sigma $$ vs. $$\breve{\sigma }$$), tone ([h] vs. ∅), vowel length ([V] vs. [Vː]), vowel height (low vs. mid vs. high), vowel position (final σ in the word vs. non-final σ), and a tone × vowel length interaction. The vowel position predictor was intended to control for the H% boundary tone observed in Figure 36. The vowel height predictor was a control for the fact that high vowels are often produced with higher pitch than non-high vowels (e.g. Sapir, 1989 and references there). This initial model also included a random intercept for speaker, and a by-speaker random slope for tone. A random effect for word could not be included because some words only occur once in our corpus. Adding a by-speaker random slope for stress led to convergence errors during model comparison, so this random effect was not included in the model.

Step-down model criticism using the log-likelihood test with a threshold of α = 0.1 led to the omission of the stress predictor from the final model.¹⁵ No further model simplification was possible, largely because the simple vowel length predictor could not be dropped from the model while the higher-order interaction tone × vowel length was retained. A summary of this model is given in Table 2; p-values were estimated from the t statistic using an upper-bound 2386 degrees of freedom (2393 observations less the 7 fixed-effect parameters in the final model; Baayen, 2008, p. 297). The final model has very low collinearity between predictors (κ = 4.37). Very similar results emerge if maximum f0 is used as the dependent variable rather than mean f0, as maximum and mean f0 on vowels are highly correlated in our data (r = 0.94).

TABLE 2. Final linear mixed-effects model for mean f0 in Hz

Predictor	β (in Hz)	se(β)	|t|	p <
Intercept	180	12.4	14.51	0.001
tone ($$V{\check{'}} $$)	54	3.98	13.51	0.001
v position (final)	51	1.47	34.33	0.001
tone ($$V{\check{'}} $$) × v length (long)	−38	7.79	4.94	0.001
v height (high vs. low)	24	1.3	18.67	0.001
v height (mid vs. low)	6	1.49	4.24	0.001
v length (long)	−7	1.40	4.74	0.001

The main takeaway from Tab. 2 is that tone does have a substantial effect on f0 (particularly for short vowels) once the intonational H% tone observed on isolation forms is controlled for. On the other hand, the effect of stress on mean f0 is essentially nil — the stress predictor was dropped from the final model. The apparent correlation between stress and raised f0 seems to be reducible to the fact that word-final stress often coincides with an intonational H% tone.