1.1. English versus other languages

Much of the work on voicing-conditioned vowel duration is based on English (e.g., Crystal & House, 1988; Fowler, 1981; Kluender et al., 1988; Luce & Charles-Luce, 1985; Port, 1981). However, English is not necessarily representative. The pervasiveness of voicing-conditioned vowel duration across different languages indicates that strong pressures drive it, but the differences between languages suggest that there is a learned component determining the size of the effect. That is, the vowel durations measured within individual languages might reflect learned language-specific duration targets rather than always reflecting purely phonetic factors.

Languages like English may exaggerate vowel duration as a cue to following consonant voicing (Chen, 1970; Cho, 2015; Solé, 2007). Thus, the voicing-conditioned vowel duration differences in English might be larger than what occurs naturally as an automatic consequence of voicing, in which case it is not an ideal language for testing universal effects of voicing on preceding vowels. Universal articulatory, aerodynamic, or perceptual effects of voicing on preceding vowels are likely to be clearest in a language without language-specific vowel duration targets specified as components of voicing. Typological patterns can provide evidence for evaluating the likely size of mechanically-driven differences in voicing-conditioned vowel duration and, consequently, which languages are most likely to be exhibiting voicing-conditioned vowel duration as it arises automatically rather than being specified as a representational target.

Some factors which interact with voicing-conditioned vowel duration have been observed primarily or solely in English. It is unclear whether the same effects would be observed in other languages. Some of the work which does compare such interactions across languages finds that different languages do not behave the same way, e.g., the ratio of vowel duration before voiced versus voiceless consonants when combined with other factors such as stress, speech rate, and inherent vowel duration (de Jong & Zawaydeh, 2002; Solé, 2007). Potential differences across languages are particularly important for patterns which have been used as evidence in evaluating what causes voicing-conditioned vowel duration, such as the size of the vowel duration difference (Chen, 1970; de Jong, 1991), speed of the articulatory transitions from the vowel to the following consonant (Chen, 1970; Raphael, 1975), and the existence of voicing-conditioned vowel duration differences even when there are intervening sonorants (Chen, 1970; Maddieson, 1999).

Previous work has noted cross-linguistic variation in the size of voicing-conditioned vowel duration effects. However, these descriptions often focus on a divide between languages with a large effect and languages with negligible effects. Some work discusses English as an outlier with a much larger effect of voicing on preceding vowel duration than other languages (e.g., Chen, 1970; Sóskuthy, 2013). Other studies focus on significance, noting that many languages have a significant effect (e.g., Coretta, 2019; Durvasula & Luo, 2014; Sanker, 2020; Schwarz, 2024). The same three languages are often cited as having a small or non-significant effect: Polish (Keating, 1980), Czech (Keating, 1985) and Arabic (de Jong & Zawaydeh, 2002; Flege & Port, 1981; Mitleb, 1984).

1.2. Potential causes

A range of different explanations have been proposed for what causes voicing-conditioned vowel duration differences, i.e., the phonetic pressures or biases which could contribute to how voicing-conditioned vowel duration differences arise. No conclusive evidence has led to a consensus about the cause. Some work proposes that there are multiple contributing mechanisms, which helps resolve these disagreements (e.g., Beguš, 2017; Coretta, 2020).

One proposed account for voicing-conditioned vowel duration is that it arises from compensatory timing in production (e.g., Coretta, 2019; Fowler, 1981; Lehiste, 1970) or in perception (e.g., Kluender et al., 1988), given that voiced obstruents are usually shorter than voiceless obstruents. Proposals that C/V ratio is a cue to voicing (e.g., Kohler, 1979; Port & Dalby, 1982) lend themselves to a similar account based on perceptual compensation. The compensatory timing account has also been explained gesturally: Segmental durations are the result of how gestures align, so a longer coda consonant might start earlier and result in a shorter preceding vowel due to greater overlap (e.g., Lehiste, 1970; Lindblom, 1967; Maddieson, 1999). In some languages, the combined V+C duration is nearly consistent with voiced and voiceless consonants (e.g., Dutch: Slis & Cohen, 1969). One criticism of compensation accounts is that syllable duration is not constant in languages like English, so the differences in vowel duration cannot simply reflect differences in gesture overlap or duration modulation to maintain consistent duration of the entire syllable (e.g., Chen, 1970; de Jong, 1991). However, compensatory timing does not need to target consistent syllable duration. Listeners normalize for speech rate, perceiving sounds as longer when surrounding sounds are shorter (e.g., Maslowski et al., 2019; Mitterer, 2018); similar relative perception of duration could result in voicing-conditioned vowel duration. Fowler (1992) argues against a perceptual compensation account because longer stop closures increase perceived vowel duration rather than decreasing it. That finding might be explained as the result of closure duration being a cue to consonant voicing (e.g., Port, 1979) and listeners identifying vowel duration relative to the duration expected in the perceived voicing context (e.g., Sanker, 2020).

A related possibility is that the glottal opening gesture might occur before the consonant constriction gesture. An early onset of the devoicing gesture would produce pre-aspiration or a devoiced portion at the end of the vowel (Hansson, 2003; Lisker, 1974), which could be reinterpreted as part of the consonant, thus shortening the vowel. Early devoicing might occur as overcompensation to avoid continuing voicing from a preceding vowel into an intervocalic consonant (cf. “shielding” as a source of postoralized nasal consonants, based on preserving the orality of a following vowel, Herbert, 1986, Ch. 7).

Another potential explanation is that more force is required to maintain the constriction in a voiceless obstruent, in which there is more pressure build-up, and that the high level of motor excitation necessary to produce such a closure results in the closure being produced more quickly (e.g., Chen, 1970; Öhman, 1967). The large vowel duration differences in languages like English could be a confound in measuring aspects of timing if they are specified as representational targets rather than directly reflecting aerodynamic requirements. The measurements given by Öhman (1967) come from a Swedish speaker, providing a useful comparison with a language other than English. Explanations based on how the transition into the obstruent is achieved seem to predict that effects of obstruent voicing should be local, i.e., should affect the immediately preceding segment but have little effect on more distant segments. In a sequence of vowel + sonorant + obstruent, articulation of the obstruent does not begin during the vowel (Chen, 1970; Maddieson, 1999), so the transition to the obstruent should primarily affect the adjacent sonorant and not the vowel preceding the sonorant. However, vowels in these environments can exhibit voicing-conditioned duration differences comparable to what is observed in vowels immediately preceding obstruents (e.g., English: Chen, 1970; Van Santen, 1992). One potential explanation is that vowel duration targets in these languages extended from vowels with immediately following obstruents to also include vowels with a sonorant intervening between them and a following obstruent.

A final possibility is that obstruent voicing requires more precision in planning the state of the larynx, so the vowel is longer to accommodate the necessary adjustments (e.g., Chomsky & Halle, 1968; Halle & Stevens, 1967). A limitation of this account is that direct articulatory observations do not provide evidence for laryngeal adjustments occurring in vowels before voiced obstruents (Lisker, 1974). Under this account, lengthening is specifically caused by voiced obstruents rather than voiced consonants more broadly, because voicing in sonorant consonants does not have the same aerodynamic challenges as voicing in obstruents. This lengthening pattern predicts that vowel duration before sonorants should be comparable to vowel duration before voiceless obstruents. However, vowel durations before sonorants are actually more similar to vowel durations before voiced obstruents (e.g., House & Fairbanks, 1953).

When a vowel duration difference exists due to phonetic pressures, it might become part of what is specified as a learned target for vowel duration in each environment. Thus, some languages might encode voicing-conditioned vowel duration as part of the representation (e.g., Chen, 1970; Fromkin, 1976; Solé, 2007).¹ Once vowel durations are encoded as targets under speaker control rather than being produced automatically, they could subsequently shift in ways which are not driven by the original phonetic cause(s). One proposed piece of evidence for vowel duration being encoded as a representational target is that some languages have a language-specific consistent proportional size of the effect across different conditions which also impact vowel duration, e.g., stress (de Jong & Zawaydeh, 2002) and speech rate (Solé, 2007). The interpretation of this evidence might depend on what we propose as the phonetic cause of voicing-conditioned vowel duration. However, there are other languages in which the evidence for reflecting a phonological process is clear. In Breton and Friulian, vowel length differences conditioned by underlying voicing of the following consonants are categorically preserved with word-final consonants despite final devoicing (Baroni & Vanelli, 2000; Le Dû, 1986). Cho (2015) proposes three types of languages based on the size of the voicing effect and whether it scales proportionally when interacting with other factors that also impact vowel duration: (1) the mechanically motivated size, e.g., Arabic and Catalan; (2) representationally encoded large effect, e.g., English; (3) representationally encoded lack of effect, e.g., Polish and Czech. While this captures more patterns than a binary divide, it does not account for the wide range of gradient sizes which are observed across languages.

2.1. Typological patterns

There are some observable patterns within the data, which provide information for the typical size of voicing-conditioned vowel duration and how it interacts with other characteristics.

Figure 1 shows the distribution of how many languages fall into each range of vowel duration ratios by position and by consonant manner. Affricates are excluded, given the very small number of studies which report them. Measurements from word-final position in languages with final devoicing are excluded. Each datapoint is an average for the consonant manner and position in each language.

Figure 1: Histograms and overlaid density curves for number of languages with each vowel duration ratio (before voiceless versus before voiced), separated by position and by consonant manner. These are the by-language vowel duration ratios reported in Table 2 through Table 7. A smaller ratio reflects a larger difference in vowel duration based on the following consonant voicing.

In medial position, Breton is clearly separate from the other languages in having a particularly large effect when measured as a ratio, i.e., a small ratio of durations; this is the bin which is an outlier on the left side of each distribution in Figure 1. It has a ratio of 0.54 with stops and 0.57 with fricatives, which is much smaller than the next smallest ratio with stops (Oriya, at 0.69) or fricatives (Turkish, at 0.74).

Figure 2 also shows the distribution of languages, but presents the voicing effect measured as a difference. When voicing-conditioned vowel duration is measured as a difference, the overall distribution is largely similar to the distribution of the effect when measured as a ratio, but the position of individual languages in the distribution is substantially different. English has the largest difference (the most extreme negative value) for word-final stops (–73) and fricatives (–79). The next largest difference with stops is in Ozolotepec Zapotec (–64) and for fricatives, it is in French (–61). Javanese has the largest difference for word-medial stops (–64), followed by Breton (–50). Breton has the largest difference for word-medial fricatives (–60), followed by Friulian (–50).

Figure 2: Histograms and overlaid density curves for number of languages with each vowel duration difference (before voiceless minus before voiced), separated by position and by consonant manner. These are the by-language vowel duration differences reported in Table 2 through Table 7.

Figure 3 shows the relationship between voicing-conditioned vowel duration in word-medial position and word-final position for the 23 languages with data from both positions and without word-final devoicing. Only manners of articulation observed in both positions for a language were included. In the by-language average, each manner of articulation was given equal weight to avoid potential effects based on how many studies examine each manner of articulation in each language: Using duration ratios, r(21) = 0.0925, p = 0.675; using duration differences, r(21) = –0.0956, p = 0.664.

Figure 3: (a) Vowel duration ratio (before voiceless versus before voiced) by position for each language. (b) Vowel duration difference (before voiceless minus before voiced) by position for each language.

Figure 4 shows the relationship between voicing-conditioned vowel duration with fricatives and stops from the 23 languages with data from both stops and fricatives in the same position. For languages with final devoicing, word-final position was excluded from the by-language average. Otherwise, each position attested with both manners of articulation was given equal weight. There is a strong positive correlation: Using duration ratios, r(21) = 0.761, p < 0.0001; using duration differences, r(21) = 0.808, p < 0.0001.

Figure 4: (a) Vowel duration ratio (before voiceless versus before voiced) by manner of articulation for each language. (b) Vowel duration difference (before voiceless minus before voiced) by manner of articulation for each language.

Very few languages are consistently found to have a large effect of voicing on preceding vowel duration when measured as a ratio (ratios below 0.7, i.e., vowels before voiceless obstruents are less than 70% the duration of vowels before voiced obstruents). English is the most consistent in having a large effect word-finally (average of 0.67 with stops, 0.7 with fricatives), across a large number of studies, though the effect is weaker word-medially (average of 0.89 with stops, 0.76 with fricatives). Kabyle (Taqbaylit Berber), Korean, and Portuguese have a relatively large effect in some studies but a much weaker effect in other studies, sometimes even for the same manner of articulation within the same position in the word; see the range of vowel duration ratios in Table 2 through Table 7. Large effects of voicing-conditioned vowel duration have also been found in Breton, Hebrew, Oriya, and Ozolotepec Zapotec, but each language is only represented by a single study. When measured as a difference, the patterns are similar; the largest differences are not consistent across studies within a language or evidence for the large difference within a language only comes from a single study. Many of the particular languages exhibiting a large duration difference are also languages with the largest average absolute vowel durations: English (word-finally), Farsi, French (word-finally), Friulian, and Javanese (word-medially).

There are also relatively few languages which are consistently found to have very little effect of voicing on preceding vowel duration when measured as a ratio (ratios above 0.9). For several languages with small observed voicing effects, evidence comes from a single study each (see Table 1 and the Appendix): Argobba, Chamoru, Degema, Gujarati, Hausa, Hindko, Ibibio, Igbo, Kapampangan, Kanashi, Kazakh, Lezgian, Oromo, Romanian, Tagalog, Tarifit Berber, Telugu, and Wolaytta Doonaa. Given that results for a language can vary substantially across studies, it is difficult to interpret the results of an individual study. Some languages exhibit a small effect in certain positions or manners of articulation but larger effects in other conditions (e.g., Japanese, Hungarian, Polish, Tashlhiyt Berber) or have a small effect in certain studies but a larger effect in other studies with the same position and manner of articulation (e.g., Arabic, Czech, English, Farsi, French, German, Hungarian, Russian). Armenian is the language with the most consistent evidence for a small voicing effect: There are two studies, both on word-final stops, where each finds a small voicing effect (Maddieson, 1977; Seyfarth & Garellek, 2018). When measured as a difference, the patterns are similar.

Do any characteristics of a language predict the degree of voicing-conditioned vowel duration? Regression models were calculated using the lme4 package (Bates et al., 2015) in the R software program (R Core Team, 2021). The lmerTest package (Kuznetsova et al., 2015) was used to calculate p-values for the factors in these models. The models were built up from an intercept-only model. Adding factors was tested incrementally with model comparison; each factor was added if it significantly improved the model. Due to the relatively small amount of data and the large amount of variation within it, interactions were not considered.

Table 13 presents the summary of a mixed effects linear regression model for the vowel duration ratio (before voiceless versus before voiced).² The fixed effects were position in the word (final, medial), manner of articulation (stops, affricates, fricatives), the status of gemination in the language (non-contrastive, contrastive), and the average duration of the vowels before voiceless consonants. This last factor was centered. There was a random intercept for language. The model included all of the data for studies which separated measurements by position and manner of articulation, excluding data for word-final consonants in languages with word-final devoicing: 275 datapoints from 71 languages.

Position in the word is a significant predictor of the extent of voicing-conditioned vowel duration. The overall effect of voicing on preceding vowel duration is smaller word-medially than word-finally, and is also observed individually in English, Farsi, Hungarian, Kabyle, Kanashi, Q’eqchi’, Tagalog, and Tashylhiyt Berber. However, not every language exhibits the same tendency. Among the languages with contrastive voicing both word-medially and word-finally for which there is data in both positions, some languages exhibit similar effects in both positions (Bengali, French, Hindi, Lezgian, Serbian, and Swedish). In Arabic, Assamese, Gujarati, Javanese, Kapampangan, Marathi, Nepali, and Yalálag Zapotec, voicing-conditioned vowel duration is a larger effect word-medially than word-finally. Among languages with word-final devoicing, the effect of voicing-conditioned vowel duration is usually smaller word-finally (based on underlying voicing) than word-medially, though not always, e.g., Friulian exhibits a larger effect word-finally than word-medially.

Manner of articulation is a significant predictor of the extent of voicing-conditioned vowel duration. Overall, the effect of voicing on preceding vowel duration is larger with fricatives than with stops. A larger effect of voicing-conditioned vowel duration with fricatives than with stops is also found in most individual languages in this study (Arabic, Dutch, English, French, Hungarian, Italian, Kazakh, Polish, Portuguese, Russian, Tashlhiyt Berber, Turkish), though sometimes the interaction is small or only in certain environments. On the other hand, Breton, Bulgarian, German, Japanese, and Ozolotepec Zapotec have a difference in the opposite direction; the effect of voicing-conditioned vowel duration is larger with stops than with fricatives. What is being evaluated here is how the effect of voicing on preceding vowel duration varies by manner—recall that the dependent variable is the duration ratio by voicing environment. One of the factors which might contribute to this difference in voicing-conditioned vowel duration is the absolute effect of manner on vowel duration: Vowels tend to be longer before fricatives than before stops, as has been established in previous work (Crystal & House, 1988; Laeufer, 1992; Peterson & Lehiste, 1960; Tanner et al., 2019).

The existence of a gemination contrast in a language is a significant predictor of the extent of voicing-conditioned vowel duration in that language. The effect of voicing on preceding vowel duration is smaller in languages which have gemination contrasts. One potential confound in experimental design might contribute to why the existence of a gemination contrast in a language would predict voicing-conditioned vowel duration. Several of the studies in languages with geminates were aimed at testing effects of gemination on preceding vowel duration and had minimal pairs for gemination, while the comparisons for voicing were less closely matched. The more limited control of additional influences on vowel duration may weaken the evidence for voicing-conditioned vowel duration.

Despite the language-level effect of gemination contrasts predicting weaker voicing-conditioned vowel duration, there was no strong pattern for gemination in a particular stimulus impacting voicing-conditioned vowel duration, i.e., no clear difference between the size of the voicing effect with singletons versus with geminates. As shown in Table 11, the interaction between gemination and the voicing effect varies substantially across studies, and the average size of the voicing effect is only slightly weaker with geminates than with singletons, both when measured as a ratio (0.87 vs 0.84) and when measured as a difference (14 ms vs 18 ms). This factor was not tested in the model because relatively few of the studies separated measurements based on gemination. While some studies find different results for geminates versus singletons, there is little evidence for consistency of that difference across manners of articulation (e.g., Tashlhyt Berber: Ridouane, 2007; Farsi: Zirak & Skaer, 2013; Polish: Malisz, 2013) or across different studies (e.g., Italian stops: Dian et al., 2024, Dipino & Celata, 2018, and Shigemori & Vietti, 2019; Arabic stops: Alamri, 2022, Al-Tamimi & Khattab, 2018, and Ferrat & Guerti, 2015). The one clear tendency with gemination is that vowels are generally longer before singletons than geminates, as has been noted previously (e.g., Al-Tamimi & Khattab, 2018; Di Benedetto & De Nardis, 2021; Ridouane, 2007).

The absolute duration of the vowel before voiceless consonants is a significant predictor of the extent of voicing-conditioned vowel duration. The effect of voicing on preceding vowel duration is smaller with longer vowels. One potential interpretation of this result is that at least some of the driving factors tend to produce consistent absolute differences rather than proportional differences. Within a language, phonological vowel length interacts with voicing-conditioned vowel duration (Tables 9 and 10), though how the interaction behaves depends on how the effect of voicing is measured. If the size of voicing-conditioned vowel duration is measured as a difference, the effect word-medially is larger with long vowels than with short vowels (17 ms vs. 12 ms) and word-finally is slightly smaller with long vowels than with short vowels (19 ms vs. 20 ms). If the size of the effect is measured as a ratio, the effect word-medially is slightly smaller with long vowels than with short vowels (0.91 vs. 0.88) and word-finally is also smaller with long vowels than with short vowels (0.9 vs. 0.85). Most studies did not use minimal quadruplets, so some differences in voicing-conditioned vowel duration for long vowels and short vowels might reflect confounding factors rather than differences inherent to the length contrast. The factor of phonological vowel length was not tested in the model because relatively few of the studies separated measurements based on phonological vowel length.

Some previous work has suggested that the existence of the phonological vowel length contrast in languages like Swedish can explain why the effect of voicing-conditioned vowel duration is weak (Buder & Stoel-Gammon, 2002). However, whether or not the language has contrastive vowel length did not improve the model (χ² = 0.085, df = 1, p = 0.77). If there are constraints on using the same cue for multiple contrasts, it is not absolute. Vowel duration can be a cue to multiple contrasts within the same language, e.g., stress and vowel length in Arabic (de Jong & Zawaydeh, 2002), coda voicing and stress in English (de Jong, 2004).

One additional factor was tested but not included because it did not improve the model: Whether additional laryngeal contrasts such as aspiration exist in the language (χ² = 0.056, df = 1, p = 0.81).

Table 14 presents the summary of a mixed effects linear regression model for the vowel duration difference (before voiceless minus before voiced) rather than vowel duration ratios.³ The fixed effects were position in the word (final, medial), manner of articulation (stops, affricates, fricatives), the status of gemination in the language (non-contrastive, contrastive), and the average duration of the vowels before voiceless consonants. This last factor was centered. There was a random intercept for language. The model included all of the data for studies which separated measurements by position and manner of articulation, excluding data for word-final consonants in languages with word-final devoicing: 275 datapoints from 71 languages.

There is one major difference between this model for duration difference and the preceding model for duration ratio. When measuring the voicing effect as a duration ratio, longer overall vowel durations predicted ratios closer to 1 (i.e., less of an effect). When measuring the voicing effect as a duration difference, longer overall vowel durations predicted more negative difference values (i.e., more of an effect).

The overall effect of position in the word is the same in this model as in the model using duration ratios as the dependent variable: The voicing effect is smaller word-medially than word-finally. However, some individual languages pattern differently depending on whether the effect is measured as a ratio or as a difference. When measured as a ratio, Arabic and Kapampangan exhibit a larger effect word-medially than word-finally, but when measured as a difference, they exhibit similar effect sizes in both environments. When measured as a ratio, French and Lezgian exhibit similar effect sizes in both environments, but when measured as a difference, they exhibit a larger effect in word-final position.

The effect of manner on the voicing effect is the same in this model as in the model using duration ratios as the dependent variable: The voicing effect is larger with fricatives than with stops. However, some individual languages pattern differently depending on whether the effect is measured as a ratio or as a difference. When measured as a difference, more languages exhibit a larger voicing effect with fricatives than with stops; some of these languages exhibited little variation by manner of articulation when measured as a ratio (Friulian, Greek), or exhibited a larger effect with stops than with fricatives when measured as a ratio (Breton).

2.2. Limitations for interpretation

A wide range of variables differ across studies, many of which can impact measurements of vowel duration. This means that no individual result can be interpreted as a definitive measurement of voicing-conditioned vowel duration within that language. Some studies also only have a small number of tokens or a small number of speakers (e.g., Chen, 1970; Maddieson, 1977). See the Supplementary Materials for the summary data from each study along with information about their methods.

Many characteristics which varied across these studies have been shown to interact with voicing-conditioned vowel duration. Some of these factors are: vowel height (Mack, 1982; Tanner et al., 2019), vowel tenseness (Luce & Charles-Luce, 1985; Peterson & Lehiste, 1960; Sharf, 1962), phonological vowel length (Elert, 1964; Maddieson, 1977; Sokolović-Perović, 2012), manner of articulation of the consonants (Crystal & House, 1988; Laeufer, 1992; Peterson & Lehiste, 1960; Tanner et al., 2019), stress (Davis & Van Summers, 1989; de Jong & Zawaydeh, 2002), position within the word (Ridouane, 2007), word length⁴ (Umeda, 1975), position within the phrase (Crystal & House, 1988; Laeufer, 1992; Luce & Charles-Luce, 1985; Umeda, 1975), and speech rate (Port, 1981; Solé, 2007; Tanner et al., 2019).

Other factors which influence vowel duration could also interact with voicing-conditioned vowel duration, even if previous work has not directly tested the interaction. Some of these factors are lexical frequency (Gahl et al., 2012; Munson & Solomon, 2004), use of real words versus nonce words (Conwell & Barta, 2018; Koob & Shea, 2024), the number of repetitions of each word (Bard et al., 2000; Fowler & Housum, 1987), and characteristics of the preceding consonant (Crystal & House, 1988; Fischer-Jørgensen, 1964). In some languages, the ratio of vowel duration before voiced versus voiceless consonants remains relatively consistent when combined with other factors like stress, speech rate, and inherent vowel duration, while other languages exhibit little proportional scaling (de Jong & Zawaydeh, 2002; Solé, 2007). Language-specific patterns like this mean that different phonological conditions could produce very different estimates of how much the size of voicing-conditioned vowel duration differences varies across languages.

Some of the studies reported here use real words (e.g., Mack, 1982 on English and French; Port et al., 1980 on Arabic; Lousada et al., 2010 on Portuguese), while others use a fixed structure which produces mostly or entirely nonce words (e.g., House & Fairbanks, 1953 on English; Beguš, 2017 on Georgian; Coretta, 2019 on Italian and Polish). Some use mostly real words with a few nonce words (e.g., Maddieson, 1977 on Armenian, Bengali, Hindi, and Marathi; Mohaghegh, 2011 on Farsi).

Most studies have vowels immediately preceding the obstruents of interest, but a few use a vowel+sonorant sequence (e.g., Mărdărescu-Teodorescu, 1995 on Romanian; Bárkányi & Kiss, 2009 on Hungarian), or a mix of words with immediately preceding vowels and words with intervening sonorants (e.g., Chen, 1970 on English; Port & Crawford, 1989 on German).

Many studies use words produced in a frame sentence (e.g., Laeufer, 1992 on English and French; Pape & Jesus, 2015 on German, Italian, and Portuguese; Beguš, 2017 on Georgian), but others use words in meaningful sentences (e.g., Davis & Van Summers, 1989 on English; Aguilar et al., 1997 on Catalan) or words in isolation (e.g., Mack, 1982 on English and French; Warner et al., 2004 on Dutch; Yu, 2004 on Lezgian). Some studies compare multiple contexts (e.g., Barry, 2003 on Russian; Sokolović-Perović, 2009 on Serbian).

Depending on the design of the study, any factors which can influence vowel duration might substantially influence the results if they are not consistent across stimuli. For example, Crystal & House (1988) demonstrate that the proportional size of the voicing effect is similar for tense and lax vowels in English, in contrast to their earlier study (Crystal & House, 1982). They note that their earlier study did not control for whether the words were prepausal or not and found little evidence for voicing-conditioned vowel duration in lax vowels.

Some studies use only exact minimal pairs (e.g., Sharf, 1962 on English; Port & Crawford, 1989 on German; Keating, 1980 on Czech and Polish; as well as most studies using nonce words). Others include some non-minimal pairs (e.g., Chen, 1970 on French, Russian, and Korean; Braunschweiler, 1997 on German), or systematically look at vowels with matched preceding and following consonants (e.g., House & Fairbanks, 1953 on English). Some of the studies do not describe the stimuli in enough detail to make clear whether they were exact minimal pairs or not (e.g., Savithri, 1986 on Kannada). Given the range of factors which contribute to vowel duration, non-minimal pairs might create the appearance of a larger or smaller difference in vowel duration than is actually caused by coda voicing. For example, the French items used by Chen (1970) are about half minimal pairs and half pairs which differ in additional ways; the duration ratio among the minimal pairs is 0.83, while the duration ratio in the other items is 0.9. However, using minimal pairs does not guarantee comparability across studies, given the range of factors which influence the size of the voicing effect.

An additional consideration in interpreting measurements of voicing-conditioned vowel duration is keeping in mind what exactly the “voicing” contrast is in different languages. Contrasts labelled as voicing or fortis/lenis often involve a bundle of several acoustic cues, e.g., Lisker (1986) identifies 16 correlates of intervocalic voicing in English stops. Phonetic evidence sometimes suggests that the fortis/lenis contrast is primarily a voicing contrast (e.g., Russian and Hungarian: Petrova et al., 2006), an aspiration contrast (e.g., German: J. Beckman et al., 2013; English: Keating, 1984), or might not neatly fall into either of these categories (e.g., Ozolotepec Zapotec: Leander, 2008). Phonetic details also vary among languages within each broad categorization, e.g., in the duration of the prevoicing (Souganidis et al., 2024) or the duration of aspiration (Cho & Ladefoged, 1999). The phonetic correlates of a fortis/lenis contrast can vary by position; in languages where unaspirated stops are not consistently voiced word-initially, intervocalic unaspirated stops are often partially voiced (e.g., German: J. Beckman et al., 2013) or even completely voiced (e.g., Farsi: Bijankhan & Nourbakhsh, 2009). Languages were only included in this typology if voicing during the consonant constriction seems to be a correlate of the laryngeal contrast in the environments being considered. This meant that some studies with vowel duration measurements before fortis and lenis consonants were not part of the analysis due to consonant information that indicated that voicing during the consonant constriction is not a consistent correlate of the fortis/lenis contrast: Lule Sami (Engstrand, 1987), Mongolian (Rogers, 2000), and Musey (Shryock, 1995).

Classification as a “true voicing” contrast is generally based on how consistently voicing is present in the lenis consonants (e.g., J. Beckman et al., 2013; Gao & Arai, 2019) and whether these lenis consonants trigger voicing assimilation (e.g., Jansen, 2004; Nicolae & Nevins, 2016). Different consonants within a language will not necessarily exhibit the same voicing behavior. The phonological patterns of laryngeal contrasts within a language can vary based on manner of articulation, e.g., word-initial English /z/ triggers voicing assimilation in preceding stops, whereas /d/ does not (Jansen, 2004). It is not clear whether surface voiced consonants would be expected to behave the same way as underlyingly voiced consonants in how they affect preceding vowel duration. The other side of the question of underlying versus surface voicing is what happens with devoiced consonants. In the current analysis, the voicing effect was evaluated separately for the neutralizing environment in languages which are described as having (nearly) categorical final devoicing. However, many languages have some degree of final devoicing. If voicing-conditioned vowel duration depends on phonetic voicing rather than phonological voicing, even gradient devoicing patterns might be expected to decrease the effect of voicing-conditioned vowel duration.

The phonetic realization of voicing contrasts can also vary across speakers of a language (e.g., English: Flege, 1982). Because of this variation, the broad characterization of a language is likely to be less informative than voicing measurements for the same tokens as the vowel duration measurements. Many studies do not report the degree of voicing in phonologically voiced versus voiceless consonants. Among those which do include this information, there is variation in how they report it. Some studies report the average proportion of voicing in the consonant (e.g., Russian: Kulikov, 2012; Spanish: Campos-Astorkiza, 2019; Turkish: Kopkallı, 1993), while others quantify voicing in other ways, such as what percentage of the consonants are voiced through at least half of their duration (e.g., Arabic: Flege & Port, 1981) or have voicing present at all (e.g., Lio: Miatto & Wivell, 2022).

Voicing during the consonant constriction is not the only correlate of voicing that might be relevant for voicing-conditioned vowel duration. For example, constriction duration is likely to be relevant. Some proposed explanations for the effect of voicing on vowel duration are based on compensatory timing (e.g., Coretta, 2019; Fowler, 1981; Kluender et al., 1988; Lehiste, 1970); such accounts would predict that languages with little difference in the duration of voiced versus voiceless consonants would also exhibit little difference in the duration of preceding vowels.

If some aspects of study design or the nature of voicing within a language have systematic effects, they might be apparent in a model. Table 15 presents the summary of a mixed effects linear regression model for the vowel duration ratio (before voiceless versus before voiced), excluding positions in which voicing is neutralized.⁵ The fixed effects were position in the word (final, medial), manner of articulation (stops, fricatives, affricates), status of gemination in the language (not contrastive, contrastive), the average duration of the vowels before voiceless consonants, and the language’s contrast type (true voicing, not true voicing). There was a random intercept for language. The model was built up from the model in Table 13. Each factor was added if it significantly improved the model, as tested with model comparison between models differing only in the presence of that factor. Studies were omitted if the value for any model factor was unclear. In the final model, there were 250 datapoints from 55 languages.

There is limited evidence here for any methodological factors creating systematic effects in measurements of voicing-conditioned vowel duration. The lack of evidence does not necessarily mean that these factors have no effect; most crucially, some of them might introduce variability rather than consistently increasing or decreasing voicing-conditioned vowel duration. Null results might be due to insufficient data or confounding factors, particularly given that this is not a controlled dataset designed to test these effects.

All of the methodological factors listed for each study in the Supplementary Materials were tested as potential factors in the model if they exhibited sufficient variation across studies to be testable. Most of them did not improve the model and were not included in the final model: number of speakers in the study (χ² = 2.9, df = 1, p = 0.089); number of tokens in the study (χ² = 0.47, df = 1, p = 0.49); number of distinct words in the study (χ² = 0.08, df = 1, p = 0.78); whether the vowel immediately preceded the obstruent with the voicing contrast (χ² = 2.2, df = 2, p = 0.33); whether the vowel was stressed (χ² = 0.54, df = 2, p = 0.76); number of syllables in the word (χ² = 4.2, df = 3, p = 0.24); the broader context of the stimuli, e.g., isolation, frame sentence, meaningful sentences (χ² = 5.3, df = 4, p = 0.26); whether the stimuli were real words or nonce words (χ² = 4.5, df = 2, p = 0.1); and whether the stimuli were exact minimal pairs (χ² = 0.066, df = 1, p = 0.8). Whether the stimuli were elicited or occurred in natural conversation was not tested as a factor because the vast majority of studies used elicited items. Whether the segment following the word-medial consonant of interest was a consonant or a vowel was also not tested, given that very few studies tested pre-consonantal environments.

The vowel duration ratio was significantly predicted by the way that the fortis/lenis contrast in the language is typically categorized. When compared to languages known to have a true voicing contrast, the effect of voicing-conditioned vowel duration was likely to be larger in languages which have been described as having a fortis/lenis contrast that is not primarily dependent on voicing. One important caveat here is that the categorization of languages is typically made at the language level, which has been followed here, e.g., English is categorized as not true voicing. However, phonetic evidence indicates that English word-final and word-medial lenis stops exhibit substantially more voicing than word-initial stops (Davidson, 2016). Another important caveat is that only seven of the languages in the dataset are in the “not true voicing” category, which is not enough for a clear generalization.

One possible interpretation is that this result reflects a tradeoff in cue weighting across languages: If voicing is a weaker cue to a fortis/lenis contrast, preceding vowel duration is likely to be a stronger cue. Friulian is described as having fully shifted from a voicing contrast in word-final obstruents to a length contrast in the preceding vowels (Baroni & Vanelli, 2000). A similar process has been proposed for varieties of English in which word-final stops are devoiced (Holt et al., 2016). This interpretation depends on all of these contrasts having historically been voicing contrasts, i.e., the initial development of vowel duration differences was triggered by voicing. Note that Friulian only exhibits a moderate effect of voicing on preceding vowel duration word-internally, where voicing of the consonants is maintained (for stops, an average ratio 0.78 and average difference 43 ms, vs. word-final ratio 0.47 and difference 136 ms).

A second possibility is that voicing-conditioned vowel duration is actually more directly influenced by characteristics other than voicing itself, e.g., duration of the consonant. Under this interpretation, some of the contrasts might have never consistently involved voicing, and the factor(s) driving the vowel duration differences might still be actively influencing it in languages without true voicing contrasts.

About half of the studies have consonant duration measurements, so it was added as a factor. This model is presented separately because including this factor requires excluding the studies which did not provide consonant duration, resulting in 151 datapoints from 46 languages. The much smaller dataset reduces or eliminates some of the effects which were clear in the models which did not require this subsetting. Table 16 presents the summary of a mixed effects linear regression model for the vowel duration ratio (before voiceless versus before voiced), excluding positions in which voicing is neutralized.⁶ The fixed effects were position in the word (final, medial), manner of articulation (stops, fricatives, affricates), status of gemination in the language (not contrastive, contrastive), the average duration of the vowels before voiceless consonants, the language’s contrast type (true voicing, not true voicing), and the consonant duration ratio. The continuous factors were centered. There was a random intercept for language.

The ratio of constriction duration in voiceless versus voiced consonants is a significant predictor of the ratio of preceding vowel duration. Languages with a larger consonant duration ratio have a smaller vowel duration ratio. Note that because both ratios have the voiceless consonant (or consonant environment) as the numerator and voiceless consonants are typically longer than voiced consonants, a larger consonant duration ratio and smaller vowel duration ratio both reflect a larger difference in duration.

Previous studies have found a similar relationship between vowel duration and following consonant duration within an individual language (e.g., Beguš, 2017; de Jong, 1991). Some authors propose that this relationship reflects a tendency for consistency in word duration (Port et al., 1980, 1987).

There is no evidence that voicing-conditioned vowel duration is mediated by the extent of actual voicing during the consonant. Relatively few studies include measurements for proportion of the consonant that is voiced. Using the subset of studies which include these measurements, adding the ratio of proportion voicing in voiceless versus voiced consonants as a factor does not improve the model (χ² = 0.33, df = 1, p = 0.57).

Table 17 presents the summary of a mixed effects linear regression model for the vowel duration difference (before voiceless minus before voiced), testing the same factors considered above for duration ratio.⁷ The fixed effects were position in the word (final, medial), manner of articulation (stops, fricatives, affricates), status of gemination in the language (not contrastive, contrastive), the average duration of the vowels before voiceless consonants, the language’s contrast type (true voicing, not true voicing), and the consonant duration difference. There was a random intercept for language. The continuous variables were centered. In the final model, there were 151 datapoints from 46 languages.

The effects of contrast type and consonant duration in this model are the same as observed above when measuring the voicing effect as a ratio.

The other factors which were tested were not included because they did not improve the model: number of speakers in the study (χ² = 1.7, df = 1, p = 0.2); number of tokens in the study (χ² = 0.12, df = 1, p = 0.73); number of distinct words in the study (χ² = 1.8, df = 1, p = 0.18); whether the vowel immediately preceded the obstruent with the voicing contrast (χ² = 1.8, df = 2, p = 0.42); whether the vowel was stressed (χ² = 0.32, df = 2, p = 0.85); number of syllables in the word (χ² = 3.7 df = 3, p = 0.29); the broader context of the stimuli, e.g., isolation, frame sentence, meaningful sentences (χ² = 4.1, df = 4, p = 0.39); whether the stimuli were real words or nonce words (χ² = 0.58, df = 2, p = 0.75); whether the stimuli were exact minimal pairs (χ² = 0.45, df = 1, p = 0.5); and difference of proportion voicing in voiceless versus voiced consonants (χ² = 2.3, df = 1, p = 0.13).

There remains an imbalance in how much of the data on voicing-conditioned vowel duration comes from Indo-European languages, which means that patterns common in Indo-European languages (particularly in the Balto-Slavic, Germanic, Indo-Iranian, and Italic branches) might have a disproportionate effect on the cross-linguistic patterns. Of the 76 languages described here, 39 are Indo-European and 37 come from other language families. Most of the Indo-European languages are represented by multiple studies, while most of the non-Indo-European languages are represented by a single study each.

3.1. Implications for underlying causes

Several different explanations have been proposed to explain voicing-conditioned vowel duration; the lack of consensus about what drives the effect is in large part due to the evidence not converging towards a single explanation. The patterns described above also provide mixed evidence. These varied results seem most consistent with the existence of multiple contributing factors (as proposed by e.g., Beguš, 2017; Coretta, 2020), as well as the duration differences becoming specified as controlled targets in some languages and increasing in size because of their usage as major cues to the voicing contrast.

One of the patterns in support of voicing-conditioned vowel duration being specified as controlled targets is a strong correlation in the size of the effect between manners of articulation. This relationship is consistent with preceding vowel duration being associated with the voicing feature in each language. If the differences between languages are mechanically driven, other language-specific differences would need to explain why the size of the effect differs by language, e.g., in how gestures are coordinated.

The effect of voicing on preceding vowel duration tends to be larger with fricatives than with stops. Voiced fricatives are more aerodynamically challenging than voiced stops; fewer languages have voiced fricatives than voiced stops (J. Ohala, 1983). Frication requires high airflow at the constriction, but the cyclic glottal closure during voicing decreases airflow. When airflow at the beginning of a voiced fricative is not sufficient to achieve frication, it could result in an approximant which may be indistinguishable from long formant transitions in the preceding vowel.

A few languages exhibit a larger effect with stops than with fricatives. There are five languages in which the voicing effect is larger before stops both when measured as a ratio and as a difference: Bulgarian, German, Japanese, Romanian, and Ozolotepec Zapotec. This result might reflect confounds in the stimuli, because many studies focused on effects of voicing do not have exactly parallel stimuli for stops and fricatives. For three of these languages, measurements come from a single study each. For German, the results vary in the three studies which include both stops and fricatives; two find a slightly larger voicing effect with fricatives (Malécot, 1971; Pape & Jesus, 2015) and one finds a much larger voicing effect with stops (Piroth & Janker, 2004). In Breton and Friulian, this direction of relative effect size is observed when measuring the voicing effect as a ratio but not when it is measured as a difference. In these languages, the larger voicing effect ratio with stops than with fricatives is likely to reflect manner of articulation also having a direct effect on vowel duration. Vowels tend to be longer before fricatives than before stops (Crystal & House, 1988; Laeufer, 1992; Peterson & Lehiste, 1960), and the proportional size of the voicing effect tends to be smaller with longer vowels, as shown in Table 13.

There are a few potential explanations for why the size of the voicing effect when measured as a ratio tends to be smaller with longer vowels (i.e., why vowel durations in voiced and voiceless environments are more similar when the absolute duration of the vowel is longer). One possibility is that durational effects of local articulatory pressures (e.g., Chen, 1970; Chomsky & Halle, 1968; Halle & Stevens, 1967; Öhman, 1967) are relatively consistent in absolute size and thus have a larger impact on vowel duration ratios when the vowels are shorter. For example, a 10 ms duration difference due to articulatory demands of transitioning to a voiceless versus voiced obstruent would be a much larger proportional difference in short vowels than in long vowels. Another possibility is that compensatory timing effects (e.g., Fowler, 1981; Lehiste, 1970) are based on approximate isochrony in syllable duration, e.g., a 10 ms longer consonant will result in a 10 ms shorter preceding vowel. This absolute trade-off in duration would result in smaller proportional effects in longer vowels.

On the other hand, the opposite patten is observed when the voicing effect is measured as a difference, i.e., vowel duration differences are smaller when the absolute duration of the vowel is shorter. If compensatory timing is based on expectations of a consistent ratio of vowel duration to consonant duration (cf. Port & Dalby, 1982), then long vowels and short vowels would be expected to exhibit voicing effects of equivalent proportional size but with a larger absolute size with long vowels than with short vowels.

The existence of these opposite patterns in whether long vowels or short vowels exhibit more of a voicing effect is consistent with multiple mechanisms driving voicing-conditioned vowel duration. The opposite patterns are also apparent across languages in the varied ways that the voicing effect interacts with other factors which influence vowel duration. Some languages are relatively consistent in the absolute size of voicing-conditioned vowel duration, while others are more consistent in the relative size (i.e., scaling proportionally when other factors influence vowel duration). Some analyses propose that proportional scaling exists when voicing-conditioned vowel duration is specified as a controlled representational target (Cho, 2015; de Jong & Zawaydeh, 2002; Solé, 2007). However, based on comparisons between phonologically long and short vowels, a larger effect does not correlate with scaling proportionally. Some languages have a small effect that is nonetheless proportionally consistent (e.g., Telugu, with a ratio of 0.94 both with short vowels and long vowels, Reddy, 1988), and other languages have a moderately sized effect that is nearly consistent in absolute size rather than scaling proportionally (e.g., Dutch, with a difference of 22 ms in long vowels and 20 ms in short vowels, Warner et al., 2004). The patterns within this data are potentially consistent with a moderate effect being automatic, while small effects are targets specified by the representation. However, there are also languages with voicing-conditioned vowel duration patterns which are neither consistent in absolute size nor in size relative to the duration of the vowel (see Tables 9 and 10).

Vowels are longer before singleton consonants than before geminates. Proposed explanations for this effect are sometimes similar to proposed explanations for voicing-conditioned vowel duration. Some accounts are based on compensatory timing, particularly in perception. If C/V ratio is a major cue to the contrast between geminates and singletons, then changing the vowel duration helps to cue the length of the consonant (Pickett et al., 1999). Another account is that speakers balance amount of effort across syllables, with longer durations involving more effort (Ridouane, 2007). An alternative account is that the relationship between gemination and vowel duration is the result of closed syllable shortening, with geminates being ambisyllabic (Maddieson, 1985). This explanation would not have implications for voicing-conditioned vowel duration. However, the existence of vowel duration differences conditioned by word-final geminates versus singletons poses a problem for closed-syllable shortening as an explanation (Ridouane, 2007).

Potential explanations for what drives voicing-conditioned vowel duration make different predictions about how voicing and gemination should interact in their effects on preceding vowel duration. However, none of these accounts entirely align with the observed lack of evidence for a consistent interaction. Slowing caused by the precise articulatory adjustments required for voiced obstruents (Chomsky & Halle, 1968; Halle & Stevens, 1967) might predict that the effect of voicing on vowel duration should be greater with geminates than with singletons, because needing to maintain the voiced obstruent for a longer period of time adds an additional challenge which might be expected to increase the time needed for making the articulatory adjustments. Because vowels tend to be shorter before geminates, many other explanations would also predict a larger effect with geminates than with singletons. This includes explanations based on local changes as the vocal tract transitions to voicelessness (e.g., Chen, 1970; Hansson, 2003; Öhman, 1967), as well as secondary developments via other characteristics which influence perceived duration (Javkin, 1976; Sanker, 2020). Compensatory timing (e.g., Fowler, 1981; Lehiste, 1970) might predict a smaller effect with geminates than with singletons. The relative consonant duration differences between voiced versus voiceless geminates is smaller than the relative differences between voiced versus voiceless singletons (Al-Tamimi & Khattab, 2018; Ghosh, 2015; Kawahara, 2006; Rojczyk & Porzuczek, 2019; Shigemori & Vietti, 2019), so vowel duration differences based on those consonant durations should also be proportionally smaller with geminates.

The lack of clear interaction between voicing and gemination further supports the existence of multiple driving factors. Compensatory timing would need to be one of the mechanisms, given that it is the only one which predicts larger proportional voicing effects with singletons than with geminates. Some studies find a larger voicing effect with singletons than with geminates (e.g., Chamoru: Santos, 2021; Kabyle: A. Elias, 2020), but many find little difference (e.g., Arabic: Alamri, 2022; Pajak, 2013; Bengali: Ghosh, 2015; Italian: Dian et al., 2024; Di Benedetto & De Nardis, 2021) or different behavior based on manner of articulation or position (e.g., Bulgarian: Maneva, 1997; Farsi: Zirak & Skaer, 2013; Polish: Malisz, 2013; Tashlhiyt Berber: Ridouane, 2007). Given that most of these languages are represented by a single study each, some of the variation might be due to characteristics of the particular stimulus items rather than a general pattern within the language.

In contrast to the correlation across manners of articulation, there is no evidence for a correlation in the size of the voicing effect across position in the word. The size of voicing-conditioned vowel duration differences word-finally is not predictive of the size of the effect word-medially in the same language. This distinction is important for making comparisons across studies both within a language and across languages; the pattern in one position in the word is not informative as a point of comparison for a different position, and pooling across positions can obscure what patterns exist in a language.

Differing by environment can be explained easily if vowel duration is specified in the representation; position in the word can be part of the environment conditioning vowel duration. In most studies, the word-medial environment was intervocalic, so position in the word generally also aligns with position in the syllable: word-final coda consonants versus word-medial onset consonants. In many languages, there is some degree of intervocalic lenition (e.g., Katz, 2021; Kirchner, 1998), which might also contribute to differences based on position. Some languages exhibit more of an effect word-finally, though other languages exhibit more of an effect word-medially (see Figure 3). The existence of multiple contributing factors might explain how different word-medial and word-final patterns could arise.

Word-final consonants are codas, while word-medial consonants are likely to be onsets, which will produce different gestural coordination patterns. A coda consonant is more directly coordinated with a preceding vowel than is an onset consonant. Because of these differences, accounts of voicing-conditioned vowel duration based on gesture timing (e.g., Lehiste, 1970; Lindblom, 1967; Maddieson, 1999) predict larger effects of consonant voicing on preceding vowel duration for codas and subsequently larger effects of word-final consonants than intervocalic consonants. Languages with larger voicing effects word-finally are consistent with an origin in gesture timing.

In general, word-medial vowels are shorter than vowels in word-final syllables due to final lengthening (e.g., Chen, 1970; Dutta & Kenstowicz, 2018; Elert, 1964; Gráczi, 2012; Maddieson, 1977; Ridouane, 2007; Sharf, 1962). A mechanism which would produce an effect with a consistent absolute size would predict larger proportional differences in shorter vowels, and thus would predict a larger effect word medially when measured as a ratio. The languages with larger effects word-medially when measured as a ratio are consistent with explanations based on local changes as the vocal tract transitions to voicelessness (e.g., Chen, 1970; Hansson, 2003; Javkin, 1976; Öhman, 1967; Sanker, 2020). A compensatory timing mechanism based on relative duration of consonant duration to vowel duration (e.g., Port & Dalby, 1982) could predict a larger voicing effect word-finally than word-medially when measured as a difference, at least in languages where word-final lengthening impacts vowels more than consonants.

The duration of consonants in different positions might also contribute to differences in voicing-conditioned vowel duration by position. In English, vowels in word-final syllables are much longer than word-medial vowels, regardless of stress, while consonants exhibit smaller effects of position (Oller, 1973) and exhibit different patterns depending on whether or not they are in a stressed syllable (Stathopoulos & Weismer, 1983). Across languages with longer word-final consonants than word-medial consonants, the size of the difference varies (e.g., Bengali: Dutta & Kenstowicz, 2018; Kanashi: Saxena et al., 2022; Lezgian: Yu, 2004; Tashlhiyt Berber: Ridouane, 2007). Some languages do not exhibit consistent differences in consonant duration by position (e.g., Swedish: Helgason et al., 2013). Some studies do not control for stress position or number of syllables, which could create a confound.

The difference in duration of voiced versus voiceless consonants does not align with the overall duration patterns by position, though some languages do have a proportionally similar effect in both positions and a larger absolute difference in the position where consonants and vowels are longer (e.g., Tashlhiyt Berber: Ridouane, 2007). It is most common for languages to exhibit a larger effect of voicing on consonant duration word-finally, both as a difference and as a ratio (e.g., Bengali: Dutta & Kenstowicz, 2018; Lezgian: Yu, 2004; English: Stathopoulos & Weismer, 1983; Swedish: Helgason et al., 2013). However, some languages exhibit a larger effect of voicing on consonant duration word-medially both as a difference and as a ratio (e.g., Kanashi: Saxena et al., 2022). Following a compensatory timing account, a larger difference in consonant duration in one position could predict larger voicing-conditioned vowel duration effects in that position. However, the by-position consonant duration differences associated with voicing do not seem to explain the by-position variation in voicing-conditioned vowel duration. As shown in Tables 16 and 17, including consonant duration ratio or difference as a predictor of the voicing effect does not eliminate the effect of position.

The transparency of intervening sonorant consonants as observed in English and several other languages is a challenge for many potential accounts of voicing-conditioned vowel duration, because most mechanisms would not extend for such a long distance. Compensatory timing is the explanation which best accounts for this longer-distance effect. However, if the voicing effect exists purely automatically in some languages, we might find that intervening sonorants block the effect in these languages and that transparency of sonorants only exists when voicing-conditioned vowel duration is specified as a representational target.

3.2. Implication for study design

Studies on the same language sometimes obtain substantially different results. Unusually large or small effects for a language in the summary data here might be due to methodological design more than characteristics of the language, particularly for languages which are represented by a single study. The variable results within a language highlight the importance of considering a range of factors which can influence vowel duration.

Some studies pool results across different positions in a word or different manners of articulation. Given that both factors have a substantial impact on the degree of voicing-conditioned vowel duration, the results here demonstrate the importance of controlling for these factors. Some factors are so frequently pooled that they could not be examined as potential variables here, e.g., vowel height.

The size of voicing-conditioned vowel duration differences in each language might be biased by tendencies in what methods are used to investigate those languages, which might focus on conditions in which the voicing effect is largest. For example, there is a large amount of work on effects of word-final stops in English but much less on word-medial consonants. As noted above, the effect in English is large with word-final stops (a ratio below 0.7 in most studies) but smaller with word-medial stops (average ratio 0.89). In other languages, most work investigating voicing-conditioned vowel duration looks at word-medial consonants, usually because of restrictions on word-final consonants (e.g., Italian). In the collection of data presented here, some of the studies were not investigating voicing-conditioned vowel duration and only incidentally included the necessary measurements; such studies are particularly useful for establishing a more representative distribution of how voicing-conditioned vowel duration behaves.

The results here do not provide any clear evidence for preferring to measure voicing-conditioned vowel duration as a ratio between vowel durations in each voicing environment or as a difference. The main consideration in this choice is whether a particular method is better at capturing how the voicing effect interacts with other influences on vowel duration. As demonstrated here and in previous work, languages vary in whether voicing-conditioned duration differences or duration ratios are more consistent across vowels which also differ in duration based on other factors (de Jong & Zawaydeh, 2002; Solé, 2007), and neither method produces completely consistent measurements within a language. In the models tested here, the one factor which behaves differently based on how the voicing effect is measured is how overall vowel duration predicts the size of the voicing effect. Longer vowel durations predicted ratios closer to 1 (i.e., less of an effect with longer vowels) but larger differences (i.e., more of an effect with longer vowels). This means that methodological differences in factors like vowel height and speech rate will have different effects on estimates of the voicing effect depending on how it is measured.

Ideally, studies should always give the raw duration measurements for each voicing category, to facilitate comparisons across studies. Because the choice of how to measure the size of the voicing effect produces different results depending on the absolute vowel durations within a particular study, this choice can substantially influence how a particular language appears to behave. For example, Lio has a small voicing effect for word-medial stops when measured as a difference (9 ms), but is at the cross-linguistic average when measured as a ratio (0.86). Javanese has the largest voicing effect for word-medial stops when measured as a difference (64 ms), but is much more similar to the cross-linguistic average when measured as a ratio (0.8).

Degema

Data comes from 10 speakers. Eight produced 12 words (6 minimal pairs) in isolation: /àdá, àtá, míìdà, míìtà, mʊ̀dá, mʊ̀tá, mɔdá, mɔtá, mɛ̀dá, mɛ̀tá, màdá, màtá/. Two produced 10 of those words: /míìdà, míìtà, mʊ̀dá, mʊ̀tá, mɔdá, mɔtá, mɛ̀dá, mɛ̀tá, màdá, màtá/. This produced a total of 116 tokens.

The speakers participated as a group; each speaker said each item in turn. Most of these words are part of the same two verb paradigms, and were elicited adjacent to each other.

Gujarati

Data comes from 4 speakers. One produced 6 words (3 minimal pairs) in isolation twice: /əndər, əntər, adro, atro, kaɡədri, kakədri/. The minimal pairs were elicited adjacent to each other. Two produced 72 words (36 minimal pairs) in isolation, once at fast speech rate and once at slow speech rate. The items were C₁VC₂ with each combination of: C₁ as voiced, voiceless unaspirated, and voiced aspirated labial, alveolar, and velar stops and postalveolar affricates; V as /a, i, u/; and C₂ as /p/ and /b/. One produced 4 words (2 minimal pairs) in isolation: /vad, vat, naɡ, nak/. The minimal pairs were elicited adjacent to each other. This produced a total of 304 tokens.

Ibibio

Data comes from 3 speakers. One produced 2 words (1 minimal pair), /àdá, àtá/, 6 times in isolation and 2 words (1 minimal pair), /díbé, dípé/, twice in isolation. One produced 2 words (1 minimal pair), /àdá, àtá/, 4 times in isolation and 2 words (1 minimal pair), /díbé, dípé/, twice in isolation. One produced 8 words (4 minimal pairs) twice in isolation: /ádú, átú, àdû, àtû, ádá, átá, ádâ, átâ/. This produced a total of 44 tokens.

Igbo

Data comes from 5 speakers. One produced 6 words (3 minimal pairs) in isolation: /ébà, épà, édà, étà, égà, ékà/. Two produced 2 words (1 minimal pair), /íba, ípa/, twice in isolation and 2 words (1 minimal pair), /íga, íka/, once in isolation. Two produced 2 words (1 minimal pair) twice in isolation: /ideꜜ, iteꜜ/. This produced a total of 26 tokens.

Kapampangan

Data comes from 3 speakers. One produced 4 words (2 minimal pairs) in isolation twice, once alongside the translation and once as a list: /atab, atap, lumod, lumot/. One produced 4 words (2 minimal pairs) in isolation once: /imbis, impis, abugadu, abukadu/. One produced 2 words (1 minimal pair) in isolation twice: /atab, atap/. The minimal pairs were elicited adjacent to each other. This produced a total of 16 tokens.

Q’eqchi’

Data comes from 11 speakers. All of them produced 6 words (3 minimal pairs) in a frame sentence: /tobok, topok, tob, top, hob, hop/. However, one of the speakers did not produce both words of the medial stop minimal pair, so the paired item was also omitted for that speaker. This produced a total of 64 tokens.

Tagalog

Data comes from 4 speakers. One produced 10 words (5 minimal pairs) in isolation: /ibon, ipon, tabon, tapon, baga, baka, bagal, bakal, sukoʔ, sugoʔ/. One produced 8 words (4 minimal pairs) in isolation: /bəɡah, bəkah, talob, talop, bɯkod, bɯkot, usoɡ, usok/. One produced 4 words (2 minimal pairs) twice in isolation: /ibon, ipon, sugat, sukat/. One produced 2 words (1 minimal pair) in isolation: /abo, apo/. This produced a total of 28 tokens.

Toda

Data comes from 13 speakers. Six produced 20 words (10 minimal pairs) both in isolation and in a frame sentence: /poːb, poːp, ed, et, ud, ut, kab, kap, pæːd, pæːt, teg, tek, tog, tok, toːg, toːk, kjeb, kjep, nib, nip/. Seven produced 12 words (6 minimal pairs) in isolation: /poːb, poːp, ed, et, kab, kap, teg, tek, toːg, toːk, kjeb, kjep/. This produced a total of 324 tokens.

The speakers participated as a group; each speaker said each item in turn.