5 Temporal (in)stability in English monosyllabic and disyllabic words: Insights on the effect of voicing on vowel duration [Paper II]
Coretta, Stefano. 2020. Temporal (in)stability in English monosyllabic and disyllabic words: Insights on the effect of voicing on vowel duration. OFS Preprints. DOI: https://doi.org/10.31219/osf.io/jvwa8.
Abstract
English is one in the wide range of languages in which the duration of vowels is modulated by the voicing of the following consonant: Vowels are shorter when followed by voiceless stops, and longer when followed by voiced stops. The so-called voicing effect has been attributed to a variety of mechanisms. Temporal compensation between the duration of the vowel and the following stop closure is one of these mechanisms. Based on acoustic data from Italian and Polish disyllabic words, the compensatory mechanism has been proposed to be a consequence of the temporal stability of the interval between the consonant releases flanking the vowel. The timing of the VC boundary within this interval determines the respective durations of the vowel and the stop closure. In this paper, it is shown that the duration of the release-to-release interval is not affected by the voicing of the second consonant in English disyllabic words, but that it is in monosyllabic words. It is argued that the stability of the interval can be derived from the isochronous phasing of the vocalic gestures in the VCV sequence of disyllabic words. The absence of the temporal anchor of a second vowel in monosyllabic words, on the other hand, allows the vocalic and the consonant gesture durations to be modified independently. Other aspects of production and perception behind the voicing effect can coexist with a temporal compensation mechanism and cannot be excluded.
5.1 Introduction
A well-known cross-linguistic tendency is that vowels have shorter durations when followed by voiceless stops and longer durations when followed by voiced stops. This so-called “voicing effect” has been long documented in a wide range of languages across different linguistic families (Maddieson & Gandour 1976; Beguš 2017). Several hypotheses have been proposed as to the origin of this phenomenon, from articulatory mechanisms to perceptual biases; however, no one particular account has gained universal support.
One such hypothesis, the compensatory temporal adjustment account, states that the voicing effect involves a compensatory mechanism between vowel and consonant closure duration. Vowels are shorter when followed by voiceless stops because the latter have longer closure durations, and, vice versa, vowels are longer before voiced stops because the latter have shorter closure durations. However, the compensatory account fails to clearly identify a speech interval within which compensation is implemented. Both the syllable (Lindblom 1967; Farnetani & Kori 1986) and the word (Slis & Cohen 1969a; Slis & Cohen 1969b; Lehiste 1970a; Lehiste 1970b) have been proposed as such intervals, but these have been subsequently criticised on empirical and logical grounds (Chen 1970; Jacewicz, Fox & Lyle 2009; Maddieson & Gandour 1976; Coretta 2019d).
In an exploratory study of acoustic durations in Italian and Polish trochaic CV́CV words, Coretta (2019d) finds that the duration of the interval between the two consonant releases is not affected by the voicing status of the second consonant. The duration of the release-to-release interval in words where the second consonant is voiceless (like /pata/) is not significantly different from that in words where the second consonant is voiced (for example, /pada/). The temporal stability of the release-to-release interval is compatible with a compensatory temporal adjustment account of the voicing effect (Lindblom 1967; Slis & Cohen 1969a; Slis & Cohen 1969b; Lehiste 1970a; Lehiste 1970b), and it offers a resolution to the drawbacks of previous versions of the account.
Given the temporal stability of the release-to-release interval, the timing of the vowel/consonant (VC) boundary (corresponding to the vowel offset and the consonant closure onset) within that interval will determine the respective durations of vowel and consonant closure. Since the VC boundary in voiceless stops is timed earlier within the release-to-release relative to voiced stops, the vowel is shorter and closures is longer when the post-vocalic stop is voiceless than when it is voiced. This interpretation agrees with known differences of closure durations in voiceless vs voiced stops (Lisker 1957; Summers 1987; Davis & Summers 1989; Jong 1991), namely that voiceless closures are longer than voiced ones. Thus, a possible diachronic pathway to the voicing effect in disyllabic words is one in which vowel and closure duration differences emerge from changes in the timing of the VC boundary within the release-to-release interval which affect the voiceless and voiced contexts differently.
Note that the release-to-release interval in itself does not have a special status. The proposed account of compensatory temporal adjustment can be understood in relation to the acoustic duration of vowels, hence the scope of compensation can (but need not) be defined in terms of acoustic intervals. The interval found to be temporally stable across voicing contexts in disyllabic words is the release-to-release interval. However, it is desirable to derive the isochrony of this acoustic interval from properties of articulatory coordination. A tentative account of the underlying gestural coordination from which the release-to-release isochrony could be derived is offered here.
5.1.1 A gestural account of the voicing effect
The task-dynamic model (Saltzman et al. 2008) of Articulatory Phonology (Browman & Goldstein 1986; Browman & Goldstein 1988; Browman & Goldstein 1992) states that any two gestures can be implemented according to two modes. Either they are initiated in synchrony or they are implemented sequentially. These modes of gestural phasing (in-phase and anti-phase) can account for a variety of patterns of articulatory timing. Relevant to our discussion is that onset consonants are generally produced in-phase with the following vowel, meaning that the vocalic and consonantal gestures are initiated together. This mechanism gives rise to the so-called C-centre effects observed with onsets, by which the acoustic duration of a vowel depends on the number of onset consonants (Browman & Goldstein 1988; Marin & Pouplier 2010; Hermes, Mücke & Grice 2013; Marin & Pouplier 2014).
According to Öhman (1966); Öhman (1967b), the speech stream is composed by a series of continuous vocalic gestures interrupted by gestures of oral constriction (consonants). Fowler (1983) further proposes that the vocalic gestures of a VCV sequence are characterised by a cyclic pattern of production, so that the temporal distance between the two vowels is constant, independent of the nature of the intervening consonant. However, the temporal distance of the V-to-V interval is modulated by the number of intervening consonants (Smith 1995; Zmarich et al. 2011; Zeroual et al. 2015). Figure 5.1 (a) illustrates this point.
Figure 5.1: Schematics of the gestural phasing of vocalic and consonantal gestures in different contexts. The x-axis is time, while the y-axis can be interpreted as oral aperture for vowels and oral constriction for consonants. (a) shows singleton vs geminate stops, (b) voiceless and voiced stops, and (c) singleton vs tautosyllabic cluster. Note that in (a) the distance between the vowels increases in the geminate context, while it is stable in (b) and (c). The circles in (b) on the consonant gesture lines indicate the time of closure onset.
The schematics at the top of Figure 5.1 (a) shows an abstract representation (based on Marin & Pouplier 2010) of a word-like series of consonants and vowels, pata (nonce words are used as examples to enable the creation of a full set of minimal pairs, as needed). The x-axis represents time, while the y-axis can be interpreted as oral aperture for vowels and oral constriction for consonants. As per C-centre effects, the onset of the consonants /p/ and /t/ are aligned with the onset of the respective following vowels (the consonants are produced in-phase with the vowels). The bottom scheme of Figure 5.1 (a) represents the word patta, syllabified as /pat.ta/. As a representational device, the geminate stop is given as two separate consecutive gestures (the actual details of gestural implementation depend on one’s chosen gestural account and are not directly relevant to the present discussion). In patta, the onset of /p/ and that of the second /t/ are, as in pata, aligned with the onset of the respective following vowels. However, the first /t/ in /pat.ta/ is instead produced anti-phase with the preceding vowel. As it can be seen by comparing the top and bottom scheme of Figure 5.1 (a), the temporal distance between the two vowels differs in pata and patta, as per the results in Smith (1995), Zmarich et al. (2011) and Zeroual et al. (2015).
On the other hand, the distance of the vowels can still be expected to be stable when there is a single intervocalic consonant that alternates in voicing. This is shown in Figure 5.1 (b). In pata and pada, the consonants are in-phase with the respective following vowels, as we have seen in Figure 5.1 (a). Summers (1987) and Jong (1991) show that the closing gesture of voiceless stops has greater velocity than that of voiced stops. Assuming that the closing gesture of both voiceless and voiced stops is initiated in synchrony with that of the following vowel as per the in-phase alignment, full oral closure will be achieved earlier in voiceless than in voiced stops relative to the beginning of the preceding vocalic gesture. Under this scenario, the temporal distance of the vowels does not differ in pata vs pada. The result is, everything else being equal, a shorter vowel and a longer (full) closure in voiceless stops, and a longer vowel and a shorter closure in voiced stops. Warren & Hay (2006) offer evidence from lip data, where the jaw and lip closing gesture duration accounts for about 80% of the vowel duration difference.
Moreover, the combined action of the isochrony of the vowel-to-vowel interval and the in-phase alignment of the onset consonant would also be responsible for the isochrony of the release-to-release interval in CVCV words. The first consonant and vowel are produced in-phase with each other, and these are sequentially followed by the second consonant and vowel, again produced in-phase with each other. Then, the differential duration found in the voicing effect would be a consequence of the different velocity of the closing gesture in voiceless vs voiced consonants. Assuming that the closing gesture of both voiceless and voiced stops is initiated in synchrony with that of the following vowel (as per the in-phase alignment), full oral closure is achieved earlier in voiceless stops relative to its timing in voiced stops.
Finally, further evidence for a vowel-based rhythmic gestural implementation of speech comes from work by Farnetani & Kori (1986) and Celata & Mairano (2014). These studies investigate the relation between vowel duration and syllable structure in Italian. In the first study, it was found that vowels followed by a singleton stop (for example in /la.da/) are longer than vowels followed by a cluster belonging to the following syllable (/la.dra/). This pattern can easily be derived from a scenario in which the distance between the vowels is the same in the two contexts (/la.da/ and /la.dra/), and the onset consonants follow a C-centre alignment. This is represented in Figure 5.1 (c). Celata & Mairano (2014) also show that the duration of the consonant/consonant cluster is negatively correlated with the duration of the preceding vowel (although the magnitude of the correlation is low to moderate).
5.1.2 The voicing effect in English
English is one of the most investigated languages in relation to the voicing effect (Meyer 1904; Heffner 1937; House & Fairbanks 1953; Belasco 1953; Peterson & Lehiste 1960; Halle & Stevens 1967; Chen 1970; Klatt 1973; Lisker 1974; Laeufer 1992; Fowler 1992; Hussein 1994; Lampp & Reklis 2004; Warren & Jacks 2005; Durvasula & Luo 2012; Ko 2018). English is also the language in which the voicing effect has the greatest magnitude relative to that of other languages. This special status of English is traditionally attributed to the phonologisation of the voicing effect in this language (Sharf 1964; Jong 2004). Vowel duration and the vowel-to-consonant duration ratio are considered to be among the most stable cues to consonantal voicing (Peterson & Lehiste 1960; Raphael 1972; Port & Dalby 1982). Kluender, Diehl & Wright (1988) proposed that the difference in vowel duration before voiceless vs voiced stops could have been enhanced and exploited to cue the voicing contrast. This could explain the greater effect of English compared for example to the effect in Italian, in which voicing is most robustly cued by vocal fold vibration during closure (Pape & Jesus 2014).
Indeed, previous studies on English report a difference in vowel duration before voiceless vs voiced stops which ranges between 20 and 150 ms, while the values for the effect in Italian are lower, between 15 and 25 ms (Magno Caldognetto et al. 1979; Farnetani & Kori 1986; Esposito 2002; Coretta 2019d). A Bayesian meta-analysis of the voicing effect (see B) returned a 95% credible interval for the effect of voicing in English monosyllabic words between 55 and 95 ms, with a meta-analytical mean of 75 ms. In other words, we can be 95% confident that the effect is between 55-95 ms. On the other hand, the meta-analytical estimate of the voicing effect for disyllabic words is lower, at about 25 ms (around 50 ms less than in monosyllabic words). This estimate is closer to the effect sizes reported for Italian. Note also that the Italian values refer to the effect as observed in disyllabic words.
However, it is possible that the alleged differences in magnitude between English and other languages are a product of the different contexts under examination (Laeufer 1992). Ko (2018), in a more recent investigation of the voicing effect in English monosyllabic words, finds a substantially lower difference in vowel duration (35 ms). The Bayesian meta-analysis (see B) further suggests a potential for publication bias, which means that the meta-analytical estimate (75 ms) could be an overestimation. Finally, the surveyed studies have a very low number of participants (mean = 3.4, SD = 2.5), which can lead to so-called Type M errors (estimate magnitude errors) and overestimation of the effect (Kirby & Sonderegger 2018; Roettger 2019). In sum, it is generally assumed that the voicing-driven differences in vowel duration are greater in English than in other languages, although the empirical foundation of this conception is not entirely straightforward. Although not the focus of this study, arguments based on differences in effect size will become relevant when discussing the results.
5.1.3 Research hypotheses
One of the aims of this study is to test whether the same temporal stability observed for the release-to-release interval in Italian and Polish disyllabic words can also be observed in English. While the temporal stability of the release-to-release interval is expected in English disyllabic words, monosyllabic words are predicted not to show such stability, for the following reasons.
As discussed above, an essential component of the release-to-release temporal stability in disyllabic words is the presence of a direct relation between the two vowels in these words. Since monosyllabic words don’t have a second vowel, there is no direct vowel-to-vowel relation to derive the release-to-release stability from. If coda consonants are produced anti-phase with the preceding vowel and the closing gesture of the consonant starts at a specific time after the production of the vowel independent of consonant voicing, then the greater velocity of the closing gesture in voiceless stops would result in a shorter vowel and a longer closure relative to voiced stops. However, the articulatory data in Jong (1991) suggests that the timing of the onset of the stop closing gesture differs depending on voicing in English monosyllabic words.
Furthermore, Jacewicz, Fox & Lyle (2009) report that, in American English, monosyllabic words are longer when the second consonant is voiced. Based on this finding, it is expected that the release-to-release duration should be longer when C2 is voiced. Jacewicz, Fox & Lyle (2009) attribute the difference in monosyllabic word duration to the difference in vowel duration before voiceless vs voiced stops. Thus, we can expect the magnitude of the difference in release-to-release duration in monosyllabic words to be close to the difference in vowel duration. This hypothesis also fits with the reported greater effect of voicing on vowel duration in monosyllabic than disyllabic words. 5.4.3 will discuss a possible solution based on perceptual mechanisms that can reconcile the expected absence of release-to-release temporal stability in English monosyllabic words with the expected presence of the voicing effect.
The data in Coretta (2019d) suggests that the intrinsic duration of vowels and consonants can contribute to the duration of the release-to-release interval. In particular, release-to-release intervals containing a high vowel have shorter durations than those with a low vowel. This is not surprising, given the well-known tendency of high vowels to be shorter than low vowels (Hertrich & Ackermann 1997; Esposito 2002; Mortensen & Tøndering 2013; Toivonen et al. 2015; Kawahara, Erickson & Suemitsu 2017). As for the consonantal place of articulation, the release-to-release is shorter in Italian and Polish when the second consonant is velar compared to when it is coronal. This could be a consequence of the fact that the closure of velar stops is shorter than that of other stops. For example, Sharf’s (1962) data on closure duration in English suggests that the closure of labial stops (60-90 ms) is about 10 ms longer than that of velar stops (55-75 ms). Thus, we can expected that release-to-release intervals with a labial stop in English will be about 10 ms longer than intervals with a velar stop.
Another set of objectives concerns the effect of voicing on vowel durations. A conceptual replication of previous studies’ effect sizes is sought, with special attention to differences between monosyllabic and disyllabic words. Only a few studies directly compare the effect in different syllabic positions (for example, Sharf 1962; and Klatt 1973). The reported effects are in the range of 50-55 ms in word-final (closed-syllable) position and 20-25 in word-medial (open-syllable) position. The Bayesian meta-analysis of the voicing effect mentioned above suggests a mean difference of 50 ms (75 ms in word-final position vs 25 ms word-medially).
The data in Sharf (1962) and Luce & Charles-Luce (1985) indicate that voiced stop closures in English are 10-13 ms shorter than voiceless stop closures. Other surveyed studies do not report effect sizes for stop closure durations (Port & Rotunno 1979; Summers 1987; Jong 1991). For this study, no specific hypothesis was set in relation to the effect of number of syllables on the effect of voicing on closure duration. Closure duration was nonetheless investigated as a conceptual replication of previous work. Moreover, future studies on closure duration will be able to build Bayesian priors using the estimated posterior distributions obtained in this study, in accordance to the principle of knowledge update via accumulation of evidence (Etz et al. 2018).
To summarise, the following research questions and respective hypotheses were formulated:
- Is the duration of the interval between two consecutive stop releases (the release-to-release interval) in monosyllabic and disyllabic words affected by the voicing of C2 in English?
- H1a: The duration of the release-to-release interval is not affected by C2 voicing in disyllabic words.
- H1b: The release-to-release interval is longer in monosyllabic words with a voiced C2 than in monosyllabic words with a voiceless C2.
- Is the duration of the release-to-release interval affected by (a) the number of syllables of the word, (b) the quality of V1, and (c) the place of C2?
- H2a: The release-to-release interval is longer in monosyllabic than in disyllabic words.
- H2b: The duration of the release-to-release interval decreases according to the hierarchy /ɑː/, /ɜː/, /iː/.
- H2c: The release-to-release interval is longer when C2 is labial.
- What is the estimated difference in the effect of voicing on vowel and stop closure duration between monosyllabic and disyllabic words?
- H3: The effect of voicing on vowel duration is greater in monosyllabic than in disyllabic words (no specific hypothesis for the effect of number of syllables on the effect of voicing on closure duration).
5.2 Methods
The following subsections describe the experimental and statistical methods of this study. The research design and data analyses were pre-registered on the Open Science Framework prior to data collection (https://osf.io/m4rzy/). Data (Coretta 2019a) and analysis scripts (https://osf.io/32fst/) are also available on the Open Science Framework. Choices on experimental design and analysis were made within the Bayesian framework of statistical inference (see 5.2.1 and 5.2.7 for details).
5.2.1 Sample size and stopping rule
Sample size and a stopping rule were decided prior to data collection with a Bayesian method of sample determination based on the Region Of Practical Equivalence (ROPE, Kruschke 2015; Vasishth, Mertzen, et al. 2018). A “no-effect” region of values around 0 is first identified. This null region (the ROPE) can be thought of as a Bayesian 95% credible interval of a distribution, the values within which can be interpreted as a negligible or null effect. For this study, a ROPE between −10 and +10 ms has been chosen. The width of 20 ms is based on the estimates of the just noticeable difference in Huggins (1972) and Nooteboom & Doodeman (1980). Differences in release-to-release durations below 10 ms (either positive or negative) will be interpreted as compatible with a null effect.
Once a ROPE width is set, the goal is to collect data during sequential testing until the width of the 95% credible interval (CI) of the tested effect is equal to or less than the ROPE width (in this study, 20 ms). In other words, the objective is to reach estimate precision, rather than significance (as in frequentist null hypothesis testing). Inference can then be made based on the credible interval of the sought effect. When the precision goal is reached (the CI width is equal or lower than the ROPE width), three possible scenarios can arise: (1) the CI of the effect completely overlaps with the ROPE around 0, in which case the data supports a practically equivalent null effect; (2) the CI of the effect completely lies outside the ROPE, which indicates that the data support the effect to be within that CI; (3) the CI partially overlaps with the ROPE, in which case no decision can be made on whether the data support one hypothesis over the other, although it still possible to infer the sign of the effect (if the CI partially overlaps with the right side of the ROPE without including 0, there is evidence for a positive effect, while if the CI overlaps with the left side of the ROPE without including 0, there is evidence for a negative effect).
An initial minimum of 20 participants was chosen for sequential testing. Due to resource and time constraints specific to this particular study, a second condition had to be included in the stopping rule such that data collection would be have to stop on 5 April 2019, independent of the ROPE condition.
5.2.2 Participants
The participants of this study were 15 native speakers of British English, who were born and raised in the Greater Manchester area. The speakers were all undergraduate students at the University of Manchester with no reported hearing or speaking disorders, and with normal or corrected to normal vision. The participants signed a written consent form and received £5 for participation.
5.2.3 Equipment
Audio recordings were obtained in a sound-attenuated room in the Phonetics Laboratory of the University of Manchester, with a Zoom H4n Pro recorder and a RØDE Lavalier microphone, at a sample rate of 44100 Hz (16-bit, downsampled to 22050 Hz for analysis). The Lavalier microphone was clipped on the participants clothes, about 20 cm from the mouth, displaced a few centimetres to one side.
5.2.4 Materials
| teep | teepus | teek | teekus |
| teeb | teebus | teeg | teegus |
| terp | terpus | terk | terkus |
| terb | terbus | terg | tergus |
| tarp | tarpus | tark | tarkus |
| tarb | tarbus | targ | targus |
The test words were CV́C(VC) words, where C = /t/, V = /iː, ɜː, ɑː/, C = /p, b, k, g/, and (VC) = /əs/. /əs/ was chosen for its lower parsability as a native suffix, in order to prevent morphological complexity in disyllabic words. This structure specification generates 24 test words, shown in Table 5.1. All of these are nonce words, with the exception of turk and tarp, and of teek via the homophone teak. Building stimuli from a structure template rather than from the lexicon ensures greater experimental and statistical control. Moreover, the use of nonce words removes or reduces confounds from some usage variables, like for example lexical frequency.22 Each word was embedded in the following frame sentences: I’ll say X this Thursday, You’ll say X this Monday, She’ll say X this Sunday, We’ll say X this Friday, They’ll say X this Tuesday. Each word + frame combination was included once in the stimuli list, so that each speaker read a total of 120 sentence stimuli (24 words × 5 frames). A total of 1800 observations were recorded (120 stimuli × 15 speakers).
5.2.5 Procedure
The experimental procedure was first explained to the participants prior to recording. The participants also familiarised themselves with the materials by reading them aloud. They were instructed not to insert pauses anywhere within the sentence stimuli and to keep a similar intonation contour for the total duration of the experiment. They were also given the chance to take any number of breaks at any point during recording. Misreadings or speech errors were corrected by asking the participant to repeat the stimulus. The reading task took around 6 to 10 minutes, while the total experiment session lasted about 25 minutes. Data collection started on 19 February 2019 and ended on 5 April 2019.
5.2.6 Data processing and measurements
A forced-aligned transcription was obtained with the SPeech Phonetisation Alignment and Syllabification software (SPPAS, Bigi 2015). The automatic annotation was corrected by the author according to the principles of phonetic segmentation detailed in Machač & Skarnitzl (2009). A custom Praat script was written to automatically detect the burst onset of the consonants in the test words, using the algorithm in Ananthapadmanabha, Prathosh & Ramakrishnan (2014). The output was checked and manually corrected by the author when necessary.
The following measures were obtained via a custom Praat script:
- Duration of the release-to-release interval: from the release of C1 to the release of C2.
- V1 duration: from appearance to disappearance of higher formant structure in the spectrogram in correspondence of V1 (Machač & Skarnitzl 2009).
- C2 closure duration: from disappearance of higher formant structure in the V1C2 sequence to the release of C2 (Machač & Skarnitzl 2009).
- Speech rate: calculated as the number of syllables per second (number of syllables in the sentence divided by the sentence duration in seconds, Plug & Smith 2018).
5.2.7 Statistical analysis
The choice of Bayesian over frequentist statistics stems from a recent discussion of the problems associated with the reliance of p-values in statistical inference (Wagenmakers 2007; Munafò et al. 2017; Kirby & Sonderegger 2018; Roettger 2019). Bayesian statistics also offers a straightforward framework for investigating the absence of differences across conditions (a “null effect”) based on the ROPE (5.2.1), as it is in part the case in this study. Another favourable aspect of Bayesian methods is that more focus is given to the distributions of the enquired effects, rather than on point estimates (which are less informative when matters of statistical power are taken into consideration, see a discussion of Type S-M errors in Kirby & Sonderegger 2018) and an arbitrary significance cut-off point. Furthermore, Bayesian inference is centred around an incremental procedure of reallocation of credibility between natural states and on evidence based on observed data (Kruschke 2015), rather than on a series of hypothetical experimental replications (Wagenmakers 2007).23 For an introduction to Bayesian statistics in phonetics, see Vasishth, Beckman, et al. (2018), and Nicenboim, Roettger & Vasishth (2018), while for a general introduction see Etz et al. (2018), McElreath (2015), Kruschke (2015), and references therein. While a thorough discussion of Bayesian methods would be beyond the scope of this paper, it is relevant to provide the less familiar reader with the basic tools for interpreting analyses and results.
Particular weight will be given to the estimated distributions of the sought effects in presenting the results of this study. The estimated distribution of an effect (or parameter) is the posterior distribution of that effect (or parameter). The posterior distribution is an approximation of the parameter distribution, and it takes into account the specified prior for that parameter, i.e. the theoretical probability of the parameter as known or derived by the researcher. The inclusion of priors in the analysis is at the heart of Bayesian modelling, which relies on prior knowledge for the estimation of parameter values. For each relevant term in the models, the 95% credible intervals (CI) should be taken as a summary of the posterior distribution, and inference should be based on the posterior rather than on the point estimate (the posterior mean, represented here with \(\bar{\theta}\)). A 95% CI can be interpreted as the 95% probability that a parameter lies within that interval range. For example, if the 95% CI is between 10 and 30 ms, there is a 95% probability that the true parameter value is between 10 and 30 ms, with extreme values being less likely than values in the centre of the interval.
In each model, priors are specified for each of the parameters to be estimated. The priors are in the form of particular distributions, like the Gaussian (normal) or the Cauchy distribution. A prior defines the prior knowledge of where the parameter might lie within a range of values. For example, a prior as a normal distribution with mean 200 ms and standard deviation 50 indicates the researcher’s belief that the parameter lies between 100 and 300 ms with 95% probability (i.e., the mean minus twice the standard deviation, and the mean plus twice the standard deviation).
Statistical analysis was performed in R v3.5.3 (R Core Team 2019). Bayesian regression models were fit with brms (Bürkner 2017; Bürkner 2018). Each model was run with four MCMC chains and 2000 iterations per chain, of which 1000 for warm-up. A Gaussian (normal) distribution was used in all the models as the response distribution. All factors were coded using treatment contrasts (the first level in this list was set as the reference level): number of syllables (disyllabic, monosyllabic), vowel (/ɑː/, /ɜː/, /iː/), C2 voicing (voiceless, voiced), C2 place of articulation (velar, labial). Speech rate was centred when included in the models so that the intercept could be interpreted as the intercept at mean speech rate. A seed (1234) was set in all models to ensure reproducibility of the output. The priors used in the models reported here will be discussed along with the results in the following sections.
A concern could be raised that the priors might have greater influence on the posterior distributions than the observed data. A sensitivity analysis based on posterior z-scores and shrinkage (Betancourt 2018) indicates that the models discussed in this study are highly informed by the observed data and don’t heavily rely on prior specifications.
5.3 Results
This section reports the results of the Bayesian models, grouped by outcome variable (release-to-release, vowel duration, closure duration). A description of the model structure and priors is given for each model, followed by the presentation of the posterior distributions of the relevant terms. Each model is assigned a number (1 to 5), and the text refers to these.
Model convergence was reached in all the reported models (\(\hat{R}\) = 1) and no major divergences in the MCMC chains were observed. The posterior predictive check plots indicate that the observed distributions are slightly positively skewed so that a log-normal distribution would have been more appropriate. Previous work has shown that speech-units duration does follow, as a general trend, a log-normal distribution (Rosen 2005; Ratnikova 2017), and the practice of transforming duration data to the logarithmic scale is not uncommon (Gahl & Baayen 2019). However, the deviations from a Gaussian distribution here were minimal, and an informal comparison of one of the models fitted with a log-normal distribution led to virtually identical results.
5.3.1 Release-to-release duration
A Bayesian regression was fit to model the duration of the release-to-release interval (model 1). The following terms were included as fixed effects: C2 voicing (voiceless, voiced), number of syllables (disyllabic, monosyllabic), centred speech rate, an interaction between C2 voicing and number of syllables. A by-speaker and by-word random intercept, and a by-speaker random coefficient for C2 voicing were entered as random effects. The following priors were used. Two weakly informative priors based on the results from Coretta (2019d) were chosen for the intercept and the effect of C2 voicing. The former prior is a normal distribution with mean 200 ms and SD = 50, while the latter a normal distribution with mean 0 ms and SD = 25. A weakly informative prior as a normal distribution with mean 50 ms and SD = 25 was specified for the effect of number of syllables. The prior is based on differences in vowel duration between mono- vs disyllabic words, which range between 30 and 100 ms (Sharf 1962; Klatt 1973). The same prior was used for the interaction between C2 voicing and number of syllables, based on the reported differences in voicing effect in mono- vs disyllabic words (Sharf 1962; Klatt 1973). The prior for the effect of centred speech rate is a normal distribution with mean -25 ms and SD = 10, and is based on results from Coretta (2019d). For the random effects, a half Cauchy distribution (location = 0, scale = 25) was used for the standard deviation and the residual standard deviation, and a LKJ(2) distribution for the correlation among the random terms.
| Predictor | Mean | SD | Q2.5 | Q97.5 | CI width |
|---|---|---|---|---|---|
| Intercept | 264.29 | 9.56 | 245.20 | 283.01 | 37.81 |
| Voicing = voiced | -4.76 | 9.87 | -24.40 | 13.53 | 37.93 |
| Num. syll. = monosyllabic | 16.40 | 9.48 | -2.49 | 35.13 | 37.62 |
| Speech rate (cntr.) | -36.13 | 2.06 | -40.11 | -32.10 | 8.01 |
| voiced × monosyll. | 17.01 | 12.59 | -6.66 | 42.05 | 48.71 |
Table 5.2 gives the posterior mean, posterior standard deviation, 2.5 and 97.5 quantiles (lower and upper bounds of the 95% credible interval), and the credible interval’s width of the fixed effects of model 1. According to the hypotheses H1a-b set out in 5.1.3, the effect of the C2 voicing predictor (with reference to disyllabic words) should be 0 (i.e. the posterior distribution of the effect should be entirely within the ROPE; in other words, no effect of voicing in disyllabic words), and the interaction between C2 voicing (= voiced) and number of syllables (= monosyllabic) should be positive (the release-to-release interval should be longer when the consonant is voiced in monosyllabic words). Based on H2a, the effect of the number of syllables (= monosyllabic) should be positive (the interval is longer when the word is monosyllabic). However, note that the precision goal (CI width ≤ 20 ms, based on the ROPE) was reached only for centred speech rate (CI width = 8.14 ms), so that these results come with a high degree of uncertainty. The posterior distribution of the estimated effect of C2 voicing on the release-to-release duration in monosyllabic words has a 95% credible interval (95% CI) between -23.86 and 15.45 ms (the mean is -4.43 ms, SD = 10.03). The 95% CI of the estimated interaction between C2 voicing (= voiced) and number of syllables (= monosyllabic) tends towards positive values, between -8.41 and 41.41 ms (\(\bar{\theta}\) = 16.53 ms, SD = 12.72). The difference in duration of the release-to-release interval between monosyllabic and disyllabic words is more clearly positive, between -1.58 and 36.53 ms (95% CI, \(\bar{\theta}\) = 17.34, SD = 9.76). Speech rate has a strong negative effect on the release-to-release duration with 95% CI = [-40.14, -32.13].
A second Bayesian regression (model 2) was fitted with the release-to-release duration as the outcome variable to test the effects of vowel and C2 place of articulation, which were entered as terms in the model without interactions. Centred speech rate was also included. The random effects structure was the same as with the first model. The relevant priors from the first model were kept. For the effects of vowel (/ɜː/, /iː/) and place of articulation (labial), the very weakly informative prior used is a normal distribution with mean = 0 ms and SD = 30. This prior was based on duration differences depending on vowel height (Heffner 1937; House & Fairbanks 1953; Hertrich & Ackermann 1997) and labial place of articulation (Sharf 1962), which both range between 10 and 30 ms.
| Predictor | Mean | SD | Q2.5 | Q97.5 | CI width |
|---|---|---|---|---|---|
| Intercept | 289.29 | 8.61 | 271.73 | 305.49 | 33.76 |
| Vowel = /ɜː/ | -8.55 | 6.84 | -21.58 | 5.84 | 27.41 |
| Vowel = /iː/ | -37.07 | 6.95 | -50.75 | -22.58 | 28.17 |
| C2 place = labial | 2.27 | 5.70 | -8.80 | 14.25 | 23.04 |
| Speech rate (cntr.) | -37.53 | 2.06 | -41.55 | -33.63 | 7.92 |
The summary of the fixed effects of model 2 are given in Table 5.3. As with model 1, only the CI width of speech rate reached the intended precision. Based on the hypotheses H2b-c in 5.1.3, the effects of vowel = /ɜː/ and vowel = /iː/ should be negative, and the latter effect should be more negative than the former. The posterior distribution of the effect of the vowel /ɜː/ shows that this vowel tends to a negative effect, with a 95% CI between −21.90 and 4.87 ms (\(\bar{\theta}\) = −8.58 ms, SD = 6.9). The vowel /iː/ has a more robust negative effect on release-to-release duration, with a 95% CI between −50.10 and −22.26 (\(\bar{\theta}\) = −36.94 ms, SD = 6.96). Hypothesis H2c states that the effect of C2 place of articulation when the consonant is labial is positive (i.e. the interval is longer when C2 is labial). The robustness of the effect of C2 place of articulation (velar vs labial stop) is less compared to the other effects: The mean of the posterior is 2.46 ms (SD = 5.68), and the 95% CI is [−9.15, 13.28].
The credible intervals of the effects in the models reported above have widths which are greater than the chosen ROPE width of 20 ms. The wide credible intervals indicate that the estimated posterior distributions of the effects have a somewhat high degree of uncertainty with them. This uncertainty is potentially due to not controlling for vowel and number of syllables in the first and second model respectively. An exploratory model (model 3) was thus fitted to the data, in which all the terms from the two models above were included. The same priors of the two separate models were used in the combined model.
Including all the relevant terms in the model (C2 voicing and place, vowel, number of syllables in interaction with C2 voicing) reduces the width of the credible intervals substantially. Figure 5.2 shows the posterior distributions of the model terms with a variety of credible intervals. Hypothesis H1a states that the effect of C2 voicing (= voiced) on the release-to-release duration of disyllabic words should be null (i.e., the posterior should be contained within the ROPE). The posterior distribution of the C2 voicing effect on release-to-release duration in this aggregated model (95% CI = [−10.45, 5.65]) is tighter than that of model 1 (95% CI = [-23.86, 15.45]) while the mean (−2.43 ms, SD = 4.06) is virtually unchanged (−4.43 ms, only a 2 ms difference). According to hypothesis H2a, the effect of number of syllables (= monosyllabic) on the release-to-release should be positive. The estimated effect of syllable number is robustly positive (95% CI = [9.17, 22.48]), with a mean (16.03 ms, SD = 3.32) similar to that in model 1. Based on hypothesis H1b, the interaction between C2 voicing (= voiced) and number of syllables (= monosyllabic) should be positive. The posterior distribution of the interaction between number of syllables and C2 voicing (95% CI = [2.65, 20.98]) suggests a positive and medium-sized interaction effect (\(\bar{\theta}\) = 11.67 ms, SD = 4.71). This result indicates that the duration of the release-to-release is greater in monosyllabic words with voiced C2 than in monosyllabic words with voiceless C2, compatibly with H1b. The effects of vowel and place of articulation have similar means as in model 2, but the credible intervals are smaller. According to hypotheses H2b-c, the effects of vowel = /ɜː/ and vowel = /iː/ should be negative, the later effect should be more negative than the former, and the effect of C2 place of articulation when the consonant is labial should be positive. The release-to-release is on average 10.05 ms (SD = 2.95, 95% CI = [−15.92, −4.24]) shorter if the vowel is /ɜː/ and 39.3 ms (SD = 2.99, 95% CI = [−45.03, −32.76]) shorter if the vowel is /iː/. C2 place of articulation (labial) has a negligible positive mean effect (2.6 ms, SD = 2.39, 95% CI = [−2.29, 7.28]).
| Predictor | Mean | SD | Q2.5 | Q97.5 | CI width |
|---|---|---|---|---|---|
| Intercept | 280.55 | 7.00 | 266.09 | 294.27 | 28.18 |
| Voicing = voiced | -2.53 | 4.14 | -10.81 | 5.55 | 16.36 |
| Num. syll. = monosyllabic | 16.04 | 3.33 | 9.42 | 22.85 | 13.42 |
| Vowel = /ɜː/ | -10.01 | 2.94 | -15.89 | -4.08 | 11.81 |
| Vowel = /iː/ | -38.91 | 2.98 | -44.87 | -32.95 | 11.92 |
| C2 place = labial | 2.45 | 2.38 | -2.11 | 7.12 | 9.24 |
| Speech rate (cntr.) | -36.14 | 2.01 | -39.90 | -32.24 | 7.66 |
| voiced × monosyll. | 11.68 | 4.69 | 2.46 | 20.87 | 18.41 |
## Warning: 'geom_halfeyeh' is deprecated.
## Use 'stat_halfeye' instead.
## See help("Deprecated") and help("tidybayes-deprecated").## Warning: 'stat_intervalh' is deprecated.
## Use 'stat_interval' instead.
## See help("Deprecated") and help("tidybayes-deprecated").
Figure 5.2: Posterior distributions and Bayesian credible intervals of the effects on release-to-release duration (model 3). For each effect, the thick blue-coloured bars indicate (from darker to lighter) the 50%, 80%, and 95% CI. The black point with bars are the posterior median (the point), the 98% (thin bar) and 66% (thicker bar) CI. The shaded grey area around 0 is the ROPE.
5.3.2 Vowel duration
A Bayesian regression model was fitted to test vowel duration (model 4). The following terms were entered: C2 voicing (voiceless, voiced), vowel (/ɑː/, /ɜː/, /iː/), number of syllables (disyllabic, monosyllabic), centred speech rate, all possible interactions between C2 voicing, vowel, and number of syllables. The same random structure as in the previous models was used (a by-speaker and by-word random intercept, and a by-speaker random coefficient for C2 voicing).
For the prior of the intercept of vowel duration, a normal distribution with mean 145 ms and standard deviation 30 was used (Heffner 1937; House & Fairbanks 1953; Peterson & Lehiste 1960; Sharf 1962; Chen 1970; Klatt 1973; Davis & Summers 1989; Laeufer 1992; Ko 2018). A normal distribution with mean 50 ms and standard deviation 20 was used as the prior for the effect of voicing on vowel duration (based on the above studies). A normal prior with mean 50 and standard deviation 25 was chosen instead for the effect of number of syllables and the interaction C2 voicing/number of syllables. For the effects of vowel, vowel/number of syllables interaction, and the three-way interaction vowel/number of syllables/C2 voicing, the prior was a normal distribution with mean 0 and standard deviation 30, based on differences reported in the studies above. A slightly more informative prior was used for the interaction between C2 voicing and vowel (mean = 0, SD = 20). The same priors as in the previous models were included for the random effects.
| Predictor | Mean | SD | Q2.5 | Q97.5 | CI width |
|---|---|---|---|---|---|
| Intercept | 125.08 | 5.97 | 113.03 | 136.45 | 23.42 |
| Voicing = voiced | 13.67 | 5.26 | 3.57 | 24.43 | 20.86 |
| Vowel = /ɜː/ | -9.11 | 5.08 | -18.83 | 1.41 | 20.24 |
| Vowel = /iː/ | -36.82 | 5.10 | -46.39 | -25.95 | 20.44 |
| Num. syll. = monosyllabic | 14.98 | 4.98 | 5.10 | 25.15 | 20.05 |
| Speech rate (cntr.) | -18.03 | 1.53 | -20.98 | -15.09 | 5.90 |
| voiced × /ɜː/ | 0.36 | 6.80 | -13.66 | 13.26 | 26.92 |
| voiced × /iː/ | 6.71 | 7.05 | -7.54 | 20.00 | 27.54 |
| voiced × monosyll. | 3.87 | 6.64 | -9.46 | 16.47 | 25.93 |
| /ɜː/ × monosyll. | 0.46 | 6.95 | -13.18 | 14.15 | 27.33 |
| /iː/ × monosyll. | -16.10 | 6.99 | -29.95 | -2.39 | 27.57 |
| voiced × /ɜː/ × monosyll. | -2.74 | 9.43 | -21.42 | 16.17 | 37.59 |
| voiced × /iː/ × monosyll. | 14.68 | 9.36 | -4.06 | 32.76 | 36.81 |
Table 5.5 reports the summary of model 4, while Figure 5.3 shows the posterior distributions and credible intervals. The precision target was reached in the non-interacting predictors (permitting a few milliseconds above 20), with the exception of the intercept. All the interactions terms have CI widths above 25 ms. The 95% CI of the posterior distribution of the duration of /ɑː/ is included in the range 112.94–136.77 ms (\(\bar{\theta}\) = 124.91 ms, SD = 5.96). The vowel /ɜː/ is 9.03 ms shorter (SD = 5.16) with CI = [−19.08, 1.63], while /iː/ is 36.77 ms shorter (SD = 5, 95% CI = [−46.42, −26.67]). C2 voicing has a small but robust positive effect on vowel duration in disyllabic words. The posterior distribution of the effect of voicing on /ɑː/ has mean 13.65 ms (SD = 5.16) and 95% CI = [3.73, 24.09]. The posterior of the interaction of voicing with vowel when the vowel is /ɜː/ is quite spread out around 0, with the 95% CI between −13.70 and 13.94 ms. This indicates that /ɑː/ and /ɜː/ are similar in their behaviour of voicing-driven durational differences. On the other hand, the effect of voicing is on average 6.73 ms greater (SD = 6.59, 95% CI = [−6.54, 19.26]) when the vowel is /iː/.
According to hypothesis H3, the interaction effect between C2 voicing (= voiced) and number of syllables (= monosyllabic) should be positive (i.e. the effect should be greater in monosyllabic than in disyllabic words). The magnitude of the voicing effect in disyllabic vs monosyllabic words is modulated by the identity of the vowel. The posterior distribution for the interaction C2 voicing/number of syllables when the vowel is /ɑː/ has mean 4.03 ms (SD = 6.7) and 95% CI [−8.98, 17.69]. This distribution indicates the possibility for a very small increase of the effect from disyllabic to monosyllabic words with /ɑː/. The three-way interaction C2 voicing/vowel/number of syllables suggests that the effect of voicing in monosyllabic words with /ɜː/ is very similar to that of monosyllabic /ɑː/-words (\(\bar{\theta}\) = −2.94, SD = 9.46, 95% CI = [−21.37, 15.77]). On the other hand, the effect increases by 14.46 ms (SD = 9.18, CI = [−3.59, 31.99]) in monosyllabic words with /iː/ relative to disyllabic /iː/-words. Note that the credible intervals of these interaction effects are quite large, so that a wide range of values are probable at 95% confidence.
## Warning: 'geom_halfeyeh' is deprecated.
## Use 'stat_halfeye' instead.
## See help("Deprecated") and help("tidybayes-deprecated").## Warning: 'stat_intervalh' is deprecated.
## Use 'stat_interval' instead.
## See help("Deprecated") and help("tidybayes-deprecated").
Figure 5.3: Posterior distributions and Bayesian credible intervals of the effects on vowel duration (model 4). For each effect, the thick blue-coloured bars indicate (from darker to lighter) the 50%, 80%, and 95% CI. The black point with bars are the posterior median (the point), the 98% (thin bar) and 66% (thicker bar) CI. The shaded grey area around 0 is the ROPE.
5.3.3 Consonant closure duration
To test various effects on C2 closure duration, model 5 was fit with closure duration as the outcome variable and the following predictors: C2 voicing (voiceless, voiced), C2 place of articulation (velar, labial), number of syllables (disyllabic, monosyllabic), all interactions between these predictor terms, and centred speech rate. The random effects were again a by-speaker and a by-word random intercept, and a by-speaker random coefficient for C2 voicing.
As priors, a normal distribution with mean 90 ms (SD = 20) was used for the intercept, based on Sharf (1962) and Luce & Charles-Luce (1985). The means reported in these studies also indicate that the closure of the stop in monosyllabic words is 10-30 ms shorter when the stop is voiced. A normal distribution with mean -20 ms (SD = 10) was chosen as the prior of the effect of C2 voicing on closure duration. The same studies indicate that labial stops have a closure which is 10-20 ms longer than the closure of velar stops. For the effect of C2 place, a normal distribution with mean 15 ms (SD = 10) was used.
| Predictor | Mean | SD | Q2.5 | Q97.5 | CI width |
|---|---|---|---|---|---|
| Intercept | 74.55 | 2.91 | 68.81 | 80.33 | 11.52 |
| Voicing = voiced | -20.79 | 3.01 | -26.54 | -14.68 | 11.86 |
| C2 place = labial | 5.19 | 2.77 | 0.01 | 10.89 | 10.88 |
| Num. syll. = monosyllabic | 3.07 | 2.80 | -2.20 | 9.03 | 11.23 |
| Speech rate (cntr.) | -9.17 | 1.24 | -11.55 | -6.67 | 4.88 |
| voiced × labial | 1.40 | 3.94 | -6.47 | 8.88 | 15.35 |
| voiced × monosyll. | 1.83 | 3.94 | -6.01 | 9.47 | 15.48 |
| labial × monosyll. | -0.82 | 3.85 | -8.56 | 6.48 | 15.05 |
| voiced × labial × monosyll. | 6.41 | 5.58 | -4.75 | 17.34 | 22.09 |
The summary of model 5 is shown in Table 5.6. See Figure 5.4 for the posteriors and credible intervals of the effects. The 96% CI width of all the terms, with the exception of the three-way interaction (voicing/place/number of syllables), is below 20 ms (the precision goal has been reached). As explained in 5.1.3, no specific hypothesis addressing closure duration was formulated. The posterior distribution of the intercept for closure duration (corresponding to the duration of voiceless velar stops in disyllabic words) has mean 74.75 ms (SD = 2.86) and 95% CI = [69.07, 80.59]. The effect of C2 voicing on closure duration is certainly negative, between −26.77 and −14.74 ms (95% CI). The posterior mean of this effect is −20.79 ms (SD = 3.06). A very small positive effect of place of articulation (labial) is suggested by the 95% CI from −0.03 to 10.76 ms (\(\bar{\theta}\) = 5.19 ms, SD = 2.77). A possibly even smaller effect of number of syllables or no effect at all can be inferred from the posterior distribution which has mean 2.98 ms and SD 2.9 (95% CI = [−2.8, 8.77]). Note that the 95% CIs of the posterior distributions of all the effects, with the exception for the effect of voicing, are within the ROPE around 0.
## Warning: 'geom_halfeyeh' is deprecated.
## Use 'stat_halfeye' instead.
## See help("Deprecated") and help("tidybayes-deprecated").## Warning: 'stat_intervalh' is deprecated.
## Use 'stat_interval' instead.
## See help("Deprecated") and help("tidybayes-deprecated").
Figure 5.4: Posterior distributions and Bayesian credible intervals of the effects on closure duration (model 5). For each effect, the thick blue-coloured bars indicate (from darker to lighter) the 50%, 80%, and 95% CI. The black point with bars are the posterior median (the point), the 98% (thin bar) and 66% (thicker bar) CI. The shaded grey area around 0 is the ROPE.
5.4 Discussion
This study set out to build on the results discussed in Coretta (2019d) by investigating durational properties of the release-to-release interval in English monosyllabic and disyllabic words. It was expected that the release-to-release interval would not be affected by C2 voicing in disyllabic words but it would in monosyllabic words. Moreover, a conceptual replication of studies on the effect of consonant voicing on vowel and closure durations was sought, with a focus on comparing the effect in mono- vs disyllabic words. This section discusses in turn the results in relation to the release-to-release interval duration (5.4.1) and to vowel and closure durations (5.4.2) by comparing them with the hypotheses of this study. 5.4.3 synthesises and links these findings back to the articulatory grounding of the temporal properties of the release-to-release interval in mono- and disyllabic words (5.1.1). Limitations and future work are also discussed.
5.4.1 Release-to-release interval
The first question (see 5.1.3) asked whether the voicing of C2 in disyllabic and monosyllabic words in English influences the duration of the release-to-release interval. Coretta (2019d) showed that the release-to-release interval duration is not affected by C2 voicing in disyllabic words of Italian and Polish. The hypotheses were that, in English, the interval is not affected in disyllabic words, like in Italian and Polish, but that it is in monosyllabic words. In sum, the results of this study indicate that the release-to-release duration of disyllabic words in English is relatively stable independent of whether C2 is voiceless (like in /tɑːpəs/) or voiced C2 (/tɑːbəs/). On the other hand, the release-to-release in monosyllabic words is longer if C2 is voiced (like in /tɑːb/ vs /tɑːp/).
Two pre-registered Bayesian regression models were fitted to the release-to-release duration (model 1-2). The established ROPE target has not been achieved (see 5.2.1). An exploratory model (model 3) including all predictors from model 1 and 2 resulted in higher estimate precision (CI widths below 20 ms). The results of model 3 suggest a negligible effect of C2 voicing on the interval duration in disyllabic words (hypothesis 1a), with a 95% probability that the true effect is between −10 and +5 ms. At lower levels of probability, the posterior distribution indicates an effect between −6 and 1 ms (60% probability). If the voicing of C2 is conditioning the duration of the release-to-release interval, this effect is very small.
The possible small effect of C2 voicing in disyllabic words could be related to an annotation bias which affects the identification of stop releases. English voiceless stops are generally followed by aspiration, and the glottal friction that makes up aspiration could mask the burst of the release. If the release of the post-vocalic voiceless stops is annotated later than the actual release (by mistaking peaks in the aspiration noise for the release burst), this could lead to longer release-to-release durations when C2 is voiceless compared to when it is voiced. Such annotation bias could explain the quite small negative effect of voicing on the interval duration, and why it is in the opposite direction of the one predicted for monosyllabic words (i.e. longer release-to-release when C2 is voiced).
On the other hand, the release-to-release interval in monosyllabic words is longer when C2 is voiced (for example, /tɑːb/) vs when it is voiceless (/tɑːp/). The interaction term between number of syllables in the word and C2 voicing is positive, between +2.5 and +21 ms (at 95% probability), which means that the effect of C2 voicing increases by 2.5 to 21 ms in monosyllabic words relative to the effect in disyllabic words. This result is compatible with hypothesis 1b that the release-to-release interval is longer in monosyllabic words with a voiced C2 than in monosyllabic words with a voiceless C2. As discussed in 5.1, the absence of release-to-release isochrony in monosyllabic words is possibly due to the absence of a second vowel which would constitute the left articulatory anchor for vowel isochrony, which in turn is argued to be the necessary element for the release-to-release temporal stability.
The second question posed at the beginning of the paper was about other effects on the release-to-release duration. As expected by hypothesis 2a, the release-to-release is longer in monosyllabic than in disyllabic words. At 95% probability, the effect of number of syllables (from di- to monosyllabic) is between 9 and 22.5 ms. As for hypothesis 2b, the results are more robust for /iː/ than for /ɜː/. When the vowel is /iː/, the release-to-release interval is 33 to 45 ms shorter compared with an interval with /ɑː/. The posterior distribution of the effect when the vowel is /ɜː/ substantially overlaps with the ROPE, although it tends towards the negative side. If there is an effect with this vowel compared to /ɑː/, it is negative and possibly around −10 ms. Finally, the hypothesis that the release-to-release interval is longer when the second consonant is labial (H2c) is not strongly corroborated. The posterior distribution of the effect of C2 place of articulation (labial) has very high precision (9.5 ms) and it is between 0 and 5 ms (at somewhat less than 80% probability). However, the distribution lies within the ROPE, indicating a practically null effect.
5.4.2 Vowel and closure duration
Question 3 addressed the effect of voicing on vowel and closure duration, and the possible differences between disyllabic and monosyllabic words. The effect of voicing on vowel duration found in this study was estimated to lie between 4 and 25 ms. This range of values is very similar to that reported in Coretta (2019d) for Italian and Polish disyllabic words (the 95% confidence interval for the effect in these languages is [8, 25]), monosyllabic words were not tested). When compared to the values in previous studies that investigated disyllabic words Klatt (1973), the effect size found in this study tends towards smaller values. However, note that the posterior distribution of the effect in the current study is entirely contained in the meta-analytical posterior distribution of the effect in the other studies, which roughly ranges between −15 and +65 ms (see B). Thus, we can assume that the deviation of this study from previous ones is not substantial. As for the effect of number of syllables on vowel duration, a similar effect to that of voicing was found, whereby vowel durations increase by 5 to 25 ms in monosyllabic words compared to disyllabic words. This relation corresponds to what has been previously reported in the literature. Finally, given that the 95% CIs of the effects of voicing and number of syllables overlap with the right side of the ROPE without including 0, the data supports positive effects, but inference on their magnitude should be carefully weighted.
It was expected that the voicing effect on vowels would be stronger in monosyllabic than in disyllabic words (hypothesis 3). The credible intervals of the posterior distributions from model 4, which are larger than the ROPE, make interpretation less straightforward. At 80% probability, the difference in voicing effect between mono- and disyllabic words is between −5 and +12.5 ms. The distribution is skewed towards the positive side, and this is compatible with results from previous studies, although the CI includes 0. The magnitude, however, is considerably lower than what previously reported. More data is needed to reach a sensible estimate precision and reduce uncertainty.
The three-way interaction between C2 voicing, vowel, and number of syllables reveals that the effect in monosyllabic words with the vowel /ɜː/ is similar to that of words with /ɑː/. On the other hand, the effect is larger if the vowel is /iː/. Model 4 estimates an effect increase of about 14.5 ms ([−4.27, 33.41]). Note that the credible interval is very wide (38 ms) and it spans over both negative and positive values, although tends more towards the latter. Moreover, the vowel /iː/ followed by a voiceless stop has, according to the model, the same duration in monosyllabic and disyllabic words. While it is not clear why the vowel should have the same duration in these contexts, this pattern suggest a possible process of /iː/ shortening in monosyllabic words. More research is warranted in relation to the observed patterns.
Turning now to consonants, there was no specific hypothesis concerning differences in the effect of voicing on closure durations between mono- and disyllabic words. C2 voicing has a robust negative effect on closure duration in disyllabic words, so that voiced closures are 14.6-26.8 ms shorter than voiceless closures. The posterior of the interaction between C2 voicing and number of syllables lies entirely within the ROPE, indicating that the effect of voicing on closure duration is similar in mono- and disyllabic words. The effects of number of syllables and place of articulation have credible intervals that are narrower than 20 ms (the ROPE width) but they lie entirely within the ROPE around 0. If these variables do have an effect on closure duration, the present analysis suggests that the means of these effects are between 0 and 5 ms. For C2 place, these values are smaller than those in Sharf (1962), which indicate a difference of 15 ms between velar and labial closure durations.
As a general trend, the differences in vowel and closure duration found in this study are smaller than those known from the literature, and considerably so in the case of vowels. A possible reason for this discrepancy could be found in problems arising from Type M errors (as briefly discussed in 5.1), and in differences of speech rate, as evidenced by comparing average segment durations. While the model’s intercept of vowel duration in this study is approximately 125 ms (SD = 5.89), the mean vowel duration in the studies surveyed in the meta-analysis (B) is 150 ms (SD = 36). These longer durations may be indicative of lower speech rates in older studies so that the effect of voicing on the absolute scale may have been greater there than at the higher speech rates of the current study, assuming a linear increase of the effect. The relative increase in vowel duration when the following stop is voiced is of 1/3 in this study compared to 1/2 in previous work, which would be suggestive of a true difference in the effect. Nonetheless, Ko’s findings (2018) support the idea that the voicing effect is not stable across speaking rates, with the consequence that the effect in absolute terms is increased at decreasing speaking rates, compatibly with the discrepancies reported here. More studies like Ko (2018) which directly test the effect of speaking rates on the voicing effect are needed to settle this issue.
5.4.3 General discussion
Coretta (2019d) proposes that the voicing-related adjustments in the relative timing of the closure onset within an isochronous speech interval (acoustically identified as the release-to-release interval) is the diachronic precursor of the cross-linguistically widespread effect of voicing on vowel duration.24 Given that the duration of the release-to-release interval in Italian, Polish, and English disyllabic words is not affected by the voicing of the post-vocalic consonant, the relative durations of vowel and closure are thought to depend on the timing of the VC boundary within that interval. A later VC boundary implies a longer vowel and a shorter closure, while, vice versa, an earlier boundary produces a shorter vowel and a longer closure. Behind the differential timing of the VC boundary within the release-to-release interval, several other accounts can be envisaged, like accounts relating to laryngeal and supraglottal adjustments (Halle & Stevens 1967; Beguš 2017; Coretta 2019e).
As discussed in 5.1.1, a prerequisite of the articulatory account proposed here for the emergence of the voicing effect is the temporal stability of the acoustic release-to-release interval and of the related articulatory gestures. However, it was expected that English mono-syllabic words do not show such temporal stability, even though the voicing effect is present in this context and allegedly even grater than in disyllabic words (although cf. B).
As mentioned at the end of 5.1.1, the absence of temporal stability and presence of the voicing effect in monosyllabic words can be reconciled by drawing from known mechanisms of perceptual enhancement. Perceptual biases, like the ones proposed by perceptual accounts of the voicing effect (Javkin 1976; Kluender, Diehl & Wright 1988; Sanker 2019), can contribute to the increase of the effect of voicing, for example as a means to enhance the perceptual difference of voiceless vs voiced stops (Lisker 1974; Lisker 1986; Stevens & Keyser 1989). In particular, vowels can be further lengthened when followed by voiced stops and/or further shortened when followed by voiceless stops, so as to produce a greater and more perceptible difference.
In the case of disyllabic words, movements of the VC boundary within the isochronous interval will logically affect both vowel duration and closure duration. On the other hand, the absence of a second vowel acting as temporal articulatory anchor (as per vowel-to-vowel isochrony) in monosyllabic words would allow articulatory stretching or compression to operate independently on the vocalic and the consonantal gestures. In the monosyllabic context, the gestural duration of vowels and following consonants can be modified in such a way that could result in a change in timing of the onset of the consonant closing gesture and in the disruption of the release-to-release isochrony.
The presence of a voicing effect with absence of release-to-release temporal stability could be obtained, for example, by keeping the vocalic gesture when the following stop is voiced active for a longer time relative to when the following stop is voiceless. Although more research is needed in this area, the articulatory studies in Raphael (1972) and Jong (1991) do suggest that the vocalic gesture in monosyllabic words is executed for a prolonged time when the following consonant is voiced. While differences in magnitude of the voicing effect should be further investigated, the potentially greater effect of voicing in monosyllabic words (albeit by a small fraction) could be ascribed to the fact that, while vowel-to-vowel isochrony constraints how vowels and consonants can be produced in disyllabic words, mechanisms affecting the VC boundary (articulatory and/or perceptual) in monosyllabic words are less constrained due to the non-application of vowel-to-vowel isochrony.
5.5 Conclusion
This paper set out to investigate temporal properties of the so-called “voicing effect,” by which vowels are shorter when followed by voiceless stops and longer when followed by voiced stops. Coretta (2019d) proposes that the voicing effect emerges via a mechanism of relative timing of the VC boundary within a temporally stable interval. Such interval was argued to be the interval between two consecutive releases, as evidenced by acoustic data from Italian and Polish disyllabic words. The temporal stability of the release-to-release in relation to consonantal voicing is thought to derive from two properties of gestural phasing, namely the isochrony of the distance between the vowels in a VCV sequence, and in-phase alignment of onset consonants and the following vowel. On the other hand, the lack of an articulatory anchor (a second vowel) in monosyllabic words would allow the release-to-release duration to be affected by C2 voicing and differ in the monosyllabic context.
This study adds to the current status of knowledge on temporal aspects of the voicing effect by showing that the release-to-release interval is not affected by C2 voicing in English disyllabic words, as in Italian and Polish, and that, instead, it is longer in monosyllabic words when C2 is voiced. While the timing of the VC boundary within the release-to-release in disyllabic words affects both vowel and closure durations in a logically dependent way, vowel and closure durations can be modulated more independently in monosyllabic words. The less constrained operation of production and perceptual mechanisms affecting the timing of the VC boundary was argued to be the reason for the seemingly greater effect of voicing reported for monosyllabic words. The data in this study, and the cumulative evidence from previous studies as evinced by a Bayesian meta-analysis, however, do not equivocally provide support for a difference in the effect between mono- and disyllabic words, and future work is necessary to shed light on the matter.
To conclude, the results of this study suggest some directions of research. Future studies should further investigate the articulatory temporal patterns of vocalic and consonantal gestures in disyllabic words. In particular, a complete assessment of the isochrony (or lack thereof) of consecutive vocalic gestures should include a variety of oppositions, involving voicing, place of articulation, number of consonants, syllabic affiliation, and prosodic contexts. Moreover, work is needed to shed light on the timing of the consonant closing gesture relative to the articulatory gesture of the preceding vowel in voiceless vs voiced stops. Finally, the scenario of emergence of the voicing effect offered here should be examined in relation to other consonantal effects on vowel duration, like other laryngeal effects and effects of manner of articulation.
The three real words in the materials have low lexical frequency (Zipf 1-7 log-frequency: tarp 2.23, teak 2.76, and turk 2.91) according to the SUBTLEX-UK corpus (Van Heuven et al. 2014).↩︎
I am not advocating here against p-values in absolute terms. On the contrary, p-values are still useful in that they provide us with a practical solution in situations that involve, for example, decision-making.↩︎
Note that isochrony here is intended as pertaining the context of voiceless vs voices stops only.↩︎