1.1. American English coda glottalization

American English voiceless stops in syllable codas are sometimes pronounced with audible glottal constriction (Bellavance, 2017; Cohn, 1993; Eddington & Channer, 2010; Eddington & Taylor, 2009; Huffman, 2005; Kahn, 1976; Kaźmierski, 2018, 2020; Kaźmierski, Wojtkowiak, & Baumann, 2016; Kilbourn-Ceron, 2017; Kilbourn-Ceron, Clayards, & Wagner, 2020; Levon, 2006; Pierrehumbert, 1994, 1995; Redi & Shattuck-Hufnagel, 2001; Roberts, 2006). For example, the word ‘bat’ may be pronounced as [bæt], [bæt], [bæʔt], or [bæʔ]. Voiceless stop glottalization is common across languages (Harris, 2001; Kohler, 1994; Michaud, 2004), though in English it is better documented outside North America, where it has more widespread socio-indexical meaning (Ashby & Przedlacka, 2014; Clark & Watson, 2016; Docherty & Foulkes, 1999a; Docherty, Hay, & Walker, 2006; Fabricius, 2002; Gordeeva & Scobbie, 2013; Henton & Bladon, 1988; Higginbottom, 1964; Holmes, 1995; Johnston, 2007; Kerswill, 2007; Mathisen, 1999; Mees & Collins, 1999; Milroy, Milroy, Hartley, & Walshaw, 1994; Newbrook, 1999; Penney, Cox, Miles, & Palethorpe, 2018; Penney, Cox, & Szakay, 2019; Ramisch, 2007; Roach, 1973, 1979; Stuart-Smith, 1999; Tollfree, 1999, 2001; Watt & Milroy, 1999; Williams & Kerswill, 1999).

The accompanying audio files (a–c) exemplify three acoustic types of American English coda glottalization. These are visualized as waveforms and spectrograms in Figure 1.¹ The first type (1a) shows aperiodicity in the waveform and visibly-irregular glottal pulses preceding a typical [t] closure. Voicing irregularity is characteristic of glottal constriction, although one does not necessarily entail the other. The second type (1b) shows irregular glottal striations that appear throughout the short syllable coda, and no silent closure. In this type, it is difficult to determine whether or not an alveolar constriction exists based on acoustic or auditory evidence. In the third type (1c), well-defined glottal irregularity precedes a brief silence. Here, the stable formants and unidentifiable stop release (though see Bellavance, 2017) suggest that the silence is a sustained glottal closure rather than an oral one, and the accompanying audio file lacks a perceptible alveolar consonant transition.

Figure 1

Waveform and spectrogram representations of three utterances of the phrase not very produced by the second author. Each panel illustrates a different type of coda glottalization associated with the coda of not, which is indicated by the box in each panel. Spectrograms were calculated with a 10 ms Gaussian window and 0.7 ms step length.

These various types are sometimes distinguished with terms such as preglottalization, glottal reinforcement, glottalization, glottaling, glottal replacement, and full glottalization (e.g., Esling, Fraser, & Harris, 2005; Higginbottom, 1964; Milroy et al., 1994). In the current study, we refer to all three types with the cover term coda glottalization. Glottalized codas all share the primary feature of audible glottal constriction, despite its variable alignment with the nucleus vowel, and with an oral stop closure that may be missing or incomplete. We refer to types (1b) and (1c) as glottal stops, using the broad symbol [ʔ] to transcribe both types. These are defined by glottal constriction in the absence of an audible oral closure. While type (1b) does not have a sustained closure that results in silence, it involves the same basic articulation as (1c), and glottal stop production is inherently noisy and often incomplete (Garellek, 2013). For transcription purposes, we include type (1a) under the broad symbol [t]: Although it exhibits coda glottalization, it also has a definite alveolar closure.

1.2. Phonological distribution of coda glottalization

Coda glottalization in American English is attested primarily for /t/ and sometimes /p/ (Bellavance, 2017; Cohn, 1993; Eddington & Channer, 2010; Eddington & Taylor, 2009; Huffman, 2005; Kahn, 1976; Kaźmierski, 2018, 2020; Kaźmierski et al., 2016; Kilbourn-Ceron, 2017; Kilbourn-Ceron et al., 2020; Levon, 2006; Pierrehumbert, 1994, 1995; Roberts, 2006). In mainstream American English, irregular voicing accompanying an oral closure (as in 1a) is attested for both sounds, but irregular voicing or glottal closure without an identifiable oral closure (as in 1b–c) is attested only for coda /t/. Across other English varieties, glottalization is most often reported with /t/ and less often with /p/ and /k/, but coda glottalization is attested for all three voiceless stops /p, t, k/ in some varieties (Docherty & Foulkes, 1999a; Docherty, Foulkes, Milroy, Milroy, & Walshaw, 1997; Johnston, 2007; Jones, Kalbfeld, Hancock, & Clark, 2019; Milroy et al., 1994; Penney et al., 2019; Roach, 1973, 1979; Stoddart, Upton, & Widdowson, 1999; Stuart-Smith, 1999; Tollfree, 1999, 2001; Trudgill, 1999).

Coda glottalization occurs both word-internally (as in litmus pronounced [liʔ.məs]) and word-finally (as in eight people pronounced [e^Iʔ.p^hi.pl]). It is especially common at phrasal junctures (Huffman, 2005). While it is reportedly variable, coda glottalization is more likely when the following onset consonant is a sonorant (Cohn, 1993; Davidson, Orosco, & Wang, under review; Huffman, 2005; Pierrehumbert, 1994, 1995; Roberts, 2006), though Huffman (2005) argues that this is true only phrase-medially.

1.3. The glottis and American English voicing contrasts

It has been proposed that coda glottalization is related to the stop voicing contrast in American English (Huffman, 2005; Keyser & Stevens, 2006; Pierrehumbert, 1994, 1995). The vocal folds normally vibrate when air flows through the glottis, producing voicing. To produce a voiceless pulmonic sound, the glottis can be configured to inhibit vocal fold vibration, such as through either glottal spreading or constriction (Garellek, 2019). If glottal spreading is used, a possible side-effect is breathy voice or aspiration during the gestural transition (Löfqvist & McGowan, 1992; Seyfarth & Garellek, 2018). If glottal constriction is used, the side-effect would instead be audible glottalization or creaky voice (Ashby & Przedlacka, 2014; Garellek, 2012).

1.4. The current proposal

Our proposal is based on the assumption that voicelessness for American English coda stops is typically produced through glottal constriction (Fujimura & Sawashima, 1971; Kahn, 1976; Westbury & Niimi, 1979), as opposed to the glottal spreading used for voiceless onsets. If the oral constriction for a coda stop is not audible, not produced, or not aligned with this glottal constriction gesture, the result is perceptible coda glottalization (Selkirk, 1972, p. 194, Browman & Goldstein, 1992; Cohn, 1993; Manuel & Vatikiotis-Bateson, 1988; Parrell & Narayanan, 2018). In the present study, we explore and attempt to explain the variability of perceptible coda glottalization. Our data are drawn from audio recordings, and are thus not suited to evaluate the presence or absence of a physical glottal constriction gesture. Whether this physical gesture is indeed always present (Kahn, 1976), is an optional variant of voiceless codas (Huffman, 2005), or is itself conditioned, is a question for future research (see discussion Section 5.6.1).

We propose that perceptible coda glottalization is variable for two reasons (following Browman & Goldstein, 1990). First, the alignment and completeness of the oral and glottal constriction gestures vary in connected speech. Glottal constriction that begins earlier than the oral constriction may appear as pre-glottalization (similar to 1a), and a complete early glottal closure (as in 1c) can silence an oral one. If the oral constriction is incomplete or not produced (as in 1b–c), then audible glottalization may be the only sign of the expected coda stop.

Oral /p, t, k/ constrictions are reduced at different rates and in different ways, which accounts for the variability in perceptible coda glottalization among these sounds. In mainstream American English, /t/ lacks an oral constriction much more often than /p/ and /k/, which is likely due to the low information content of /t/ in American English (Cohen Priva, 2008; 2012; 2015; 2017). A labial /p/ constriction is usually not reduced in English (Jun, 1996), and an incomplete /k/ more typically results in spirantization (Riebold, 2011), which is both highly audible and likely to conceal the acoustic and visual signs of glottal constriction. When it is not omitted, a coronal /t/ constriction is shorter and has less alveolar contact when it is followed immediately by a non-coronal consonantal constriction, especially at faster speech rates (Barry, 1991; Browman & Goldstein, 1995; Byrd & Tan, 1996; Kühnert & Hoole, 2004; Sung & Kochetov, 2018).³ Given these patterns, glottalization should be attested most often for /t/, especially when followed by a labial or velar onset consonant, and especially at faster speech rates.

The second source of variability in coda glottalization is because some phonological contexts facilitate perceptible irregular voicing more than others (Huffman, 2005). At phrase junctures, irregular voicing is common (Redi & Shattuck-Hufnagel, 2001), where it may be used independently to signal a boundary (Kilbourn-Ceron, 2017; Kreiman, 1982; Slifka, 2007; Umeda, 1978). When a voiceless stop precedes a boundary, irregular voicing at the boundary may reinforce the perceptibility of the stop’s glottal constriction.

In this context, we also test the hypothesis in Huffman (2005) that coda glottalization is more audible before /n, m, l/ due to an anticipatory increase in supraglottal volume, which facilitates transglottal airflow and thus makes glottalization more audible (Cohn, 1993; Moll & Daniloff, 1971, see Section 1.3). To preview our results, we find that coda glottalization before sonorants is near-categorical and not limited to /n, m, l/, which is inconsistent with this final hypothesis. Elsewhere, however, it is consistent with the other physical causes proposed above. For this reason, we ultimately argue that perceptible coda /t/ glottalization is planned before sonorants, whereas elsewhere it is primarily the consequence of unplanned physical and mechanical variation.

2.1. Overview

To support this proposal, we examine the distribution of perceptible glottal stops in a conversational corpus of American English speech. The focus of the analyses is the phonetic and phonological context surrounding the variable coda glottalization: the following onset consonant, the position of the coda in a phrase, stress before and after each coda, and the talker’s rate of speech.

The two analyses in this study are exploratory. In the first analysis, we model the rate of hand-annotated coda glottalization in word-final position before each individual onset consonant. This model estimates and adjusts for other factors, such as speech rate, but these other factors are treated as independent predictors. The second analysis uses a model tree to automatically search for significant contrasts and interactions within the data. In both cases, we provide an account for the phonological distribution indicated by the models, and argue that this distribution is most consistent with our proposal that (except before sonorants) variation in perceptible American English coda glottalization derives from phonetic reduction of the oral constriction and reinforcement of the glottal one. Our annotated dataset is available at https://doi.org/10.5281/zenodo.3332888.

2.2. The Buckeye Corpus

The data from this study come from the Buckeye Corpus of Conversational Speech (Pitt et al., 2007). The Buckeye Corpus contains interviews with forty white middle-class native English speakers who had grown up in central Ohio. Twenty speakers were under thirty years old and twenty were over forty; twenty were female and twenty were male. Although central Ohio straddles several dialect regions for white American English speakers, the demographic limitations of the corpus should be noted. The corpus comprises about 300,000 words of speech by the interviewees, and each interview is up to an hour long. Interviewees were asked about their background, and were asked to discuss their opinions about everyday topics.

The audio files in the Buckeye Corpus were recorded at a 16 kHz sampling rate with 16-bit depth. The recordings were transcribed orthographically, then automatically force-aligned based on a phonetic dictionary. The dictionary transcriptions and segment time-stamps were then hand-corrected by trained annotators to create a close phonetic transcription for each word token. Among the other close transcription conventions (described in the corpus manual), the close transcriptions include [ʔ] in place of some dictionary /t/ segments. Longer portions of irregular voicing were noted by annotators in a separate transcription file.

2.3. Data used in analysis

For this study, we identified a subset of the recorded words containing a singleton /t, p/ coda. We annotated both /t/ and /p/ codas in the corpus with the intent of analyzing both segments in this study. However, we found that /p/ codas meeting the criteria were too rare (947 included tokens) and acoustically heterogeneous in the natural speech of the Buckeye Corpus to draw useful conclusions. The data that we used in our present analyses therefore include only /t/ codas. Our complete annotated dataset, including /p/ codas with acoustic measurements, is made available for other researchers (see the accompanying Data Accessibility statement and Seyfarth & Garellek, 2015 for a preliminary analysis).

The tokens included in this dataset are all those in the corpus that met these criteria:

1. The target segment was in a singleton syllable coda in the dictionary transcription of the word. Syllabification uses the procedure from Gorman (2013, Appendix B), with two modifications. First, /pj/ and /tw/ sequences were considered to be permissible word-medial onsets (as in popular, between). Second, postvocalic /ɹ/ was never included in a syllable nucleus.

2. If the target segment was /t/ in the dictionary transcription of the word, annotators had given it a close transcription of [t, ʔ, d, ɾ, t͡ʃ], or [s]. If the target segment was dictionary /p/, annotators had also transcribed it as [p]. Over 95% of codas were word-final, which made it straightforward to check the correspondence between the dictionary and close transcriptions. For the remaining word-medial codas, we identified a possible matching segment in the close transcription that was preceded by any vowel and followed by any consonant, and then hand-checked the set of matches.⁴ This criterion excluded 1,693 /t/ codas and 81 /p/ codas, which were primarily auditorily-deleted codas.

3. If the target /t, p/ coda was word-final, it was not followed by a vowel in either the dictionary or close transcription of the following word. Although /t/ glottalization is also attested word-finally before vowels (Eddington & Channer, 2010; Eddington & Taylor, 2009; Kaźmierski, 2018; Kaźmierski et al., 2016; Kilbourn-Ceron, 2017; Roberts, 2006; Umeda, 1978), an analysis of coda glottalization in this environment is more complex because it is confounded by resyllabification and glottal stop insertion in onsetless syllables (Garellek, 2013). We discuss this environment further in Section 5.4.

4. The vowel preceding the target segment was at least 50 milliseconds (ms) long.

5. The word containing the target segment did not contain a speech error, disfluency, or other interruption; it did not overlap with the interviewer’s speech or non-speech recording noise; and its close transcription matched the segment timestamps provided in the corpus.

2.4. Annotation procedure

2.5. Data summary

In total, the data include 12,451 singleton coda /t, p/ segments (11,504 /t/ and 947 /p/) including 477 unique words produced by 40 speakers. Nearly all (12,170 tokens; 98%) were word-final codas, and slightly over half (6,931 tokens; 56%) were phrase-medial. The four most common word types (that, it, but, not) comprise 50% of the data, and the fifteen most common comprise 85%, with a long tail. An earlier version of this dataset, including some acoustic measures and incomplete phrase annotations, was described and analyzed in Seyfarth and Garellek (2015).

The codas were followed by one of 21 different consonants or by a pause. Pre-pausal tokens were always utterance-final, and not closely followed by another speech sound. The most common environment was pre-pausal (4,056 tokens; 33%); codas were also common before /ð/ (1,556 tokens; 12.5%) or /w/ (1,037 tokens; 8.3%). The data included between 128–651 tokens before every other segment, except for uncommon /v/ (17 tokens; 0.1%) and /t͡ʃ/ (41 tokens; 0.3%).

The transcription of the consonant following the target codas was based on the dictionary transcription of the following word. We chose to use dictionary transcriptions of the following consonant, rather than close transcriptions, because we expected that the pronunciation of the following consonant might be equally influenced by the target coda (e.g., see Kaźmierski et al., 2016, on /j/ onsets).

2.5.1. Variable realizations of coda /t/

Figure 2 shows the empirical distribution of coda /t/ realizations before each type of following consonant. This figure reflects our corrected transcriptions, except for the deleted codas, which were not manually reviewed and are not included in our subsequent analyses.⁹ The overall average rate of coda /t/ deletion is lower than reported in previous work (Bybee, 2002; Guy, 1991), which is most likely because our dataset includes only singleton codas produced by white speakers.

Figure 2

Observed distributions of coda /t/ realizations before each type of following consonant. Deleted tokens were excluded from the analysis, but all other tokens were included.

Other than [t, ʔ] realizations, it is relatively common for coda /t/ to be voiced before /h/. Under the assumption that coda /t/ normally involves glottal constriction to inhibit voicing (Section 1.4), this makes sense: The glottal constriction during a voiceless /t/ will be reduced when it is immediately followed by a voiceless /h/, which is defined by a glottal spreading gesture. This leaves the glottis in a neutral position that is conducive to coarticulatory voicing between voiceless coda /t/ and voiceless onset /h/. On the other hand, this result would not make sense if coda /t/ (like onset /t/) normally involved glottal spreading to inhibit voicing: A subsequent /h/ would at most enhance a hypothetical /t/ glottal spreading gesture. This would continue to inhibit voicing, and thus coarticulatory voicing at a /t.h/ juncture should be rare, but this is not what we find.

It is also relatively common for coda /t/ to be voiced before other voiced coronals, affricated before /j/ (see also Kaźmierski et al., 2016), and spirantized before voiceless fricatives, especially /s, ʃ/. All of these realizations are phonetically natural in their respective environments.

3.1. Model procedure

The probability of glottalization for the word-final /t/ codas was modeled with a multilevel logistic regression using the brms package for R (Bürkner, 2018; R Core Team, 2018). The dependent variable was whether each token was annotated as glottal stop ([ʔ]) or not ([t, d, ɾ, t͡s, t͡ʃ, s]). The model included an overall intercept and parameters for the following onset segment (21 possible consonants, or a pause) and all interactions between following segment and phrase position (medial or final). Including interaction parameters between following segment and phrase position allows us to model how the effects of the following segment might change when a phrase juncture intervenes between a /t/ coda and that segment.

Also included were parameters for the preceding nucleus vowel (13 possible vowels), the presence of stress on the target and following syllable, the presence of phrasal creak, the talker’s speech rate in syllables per second, and the speaker’s age (old or young) and gender (female or male), and an interaction between age and gender. All categorical predictors were sum-coded with values of –0.5 or 0.5 for each parameter.

In order to facilitate generalizations across speakers and word types, the model included group-level intercepts for each speaker, each word, and each following word. For each speaker, the model included group-level slopes for the following segment and all interactions with phrase position, and for speech rate. For each word type, the model included group-level slopes for phrase position. There were no group-level slopes for the following word type.

All parameters were estimated with the default priors provided by brms, except the population-level slopes, which were estimated with Gaussian priors with μ = 0 and σ = 2. The model was fit via Markov chain Monte Carlo with Stan using the default sampler (Carpenter et al., 2017; Stan Development Team, 2018) with four chains and 2,000 samples per chain, discarding the first 1,000 samples per chain as warm-up. Convergence and model fit were assessed via the potential scale reduction statistic Rˆ (all <1.004) and visual inspection of the posterior predictive density (Gabry, Simpson, Vehtari, Betancourt, & Gelman, 2019).

3.2. Effects of the following onset consonant

Figure 3 shows the rate of perceptible glottal stops in place of /t/ before each onset consonant type (or pause), as estimated from the model posterior. The upper panel shows the estimated rates in phrase-medial position, and the lower panel shows the estimated rates in phrase-final position. The onset consonants are ordered from left to right by the rate of glottal stop production, with the highest rates of glottalization on the left, so the x-axis differs between the two panels.

Figure 3

Model estimates for the probability of a coda /t/ being realized as a glottal stop (y-axis) when a coda appears before each type of onset consonant (x-axis), ordered by rate of glottal stop production. Top panel shows glottal stop probabilities in phrase-medial position; bottom panel shows when a phrase juncture intervenes between the coda and the following onset. Note that the x-axis order differs between panels. Bar heights are median posterior estimates, and error bars are one standard error above and below the median. Model estimates for onset consonants with zero observations in the data are not included in the figure. In this and all subsequent plots, estimates are given with other predictors held at their mean value (for continuous predictors) or averaged over all levels of other predictors (for categorical predictors).

By referring to the model estimates rather than the raw (empirical) frequencies in the corpus, we are able to adjust for idiosyncratic effects associated with individual speakers and lexical contexts, as well as other variables that may affect glottalization rates but are not evenly distributed across contexts. The Bayesian model estimates also allow us to take into account imbalances in the frequencies of different contexts: The estimated rates for contexts that are rarely observed are drawn towards the conditional means (e.g., the estimated rate before rare /v/ is drawn toward the average rate before all other segments).

3.3. Phrase-final position

Besides the effect of the following onset consonant, the rate of perceptible coda glottalization increases overall in phrase-final position compared to phrase-medial position. Phrase-final position is associated with 0.39 greater log-odds of glottalization compared to phrase-medial position (95% posterior density interval: 0.01 to 0.79). This is consistent with our proposal that phrase-final irregular voicing increases the audibility (and perhaps the extent) of the glottal constriction associated with /t/, making it more likely to completely obscure or replace the oral /t/ constriction.

The 95% interval for the overall effect of phrase position is large. This is likely because the effect of phrase-final position is not uniform with respect to the different (phrase-initial) onset consonants. As shown in Figure 3, coda glottalization rates before obstruents are generally higher when a phrase juncture intervenes (lower panel) than when it does not (upper panel). However, an intervening phrase juncture instead very slightly lowers coda glottalization rates before sonorants (see also Seyfarth & Garellek, 2015).

If the sonorant effect is considered to be an allophonic alternation, this decrease is consistent with the hypothesis that phonological alternations that depend on upcoming sounds are sensitive to the availability of those sounds during speech production (Côté, 2013; Kilbourn-Ceron, 2017; Kilbourn-Ceron et al., 2020; Kilbourn-Ceron & Sonderegger, 2018; Kilbourn-Ceron, Wagner, & Clayards, 2016; Tanner, Sonderegger, & Wagner, 2017; Wagner, 2012). Because a sonorant is arguably less accessible during planning when it occurs in a new phrase than when it occurs during the existing one, a phonological (planned) sonorant effect should apply slightly less often across phrase junctures than within the same phrase.

3.4. Other effects of phonetic context

3.4.1. Speech rate

Marginal predictions across increasing speech rate are visualized in Figure 4. A one-syllable-per-second increase in speech rate is associated with a median increase of 0.11 log-odds of glottalization (95% posterior density interval: 0.05 to 0.17). Deletion of oral /t/ is more likely at faster speech rates (Kul, 2015; Tanner et al., 2017). Thus, this result is consistent with our proposal that coda glottalization may be more perceptible when the oral constriction is incomplete or not produced, which is more likely at faster speech rates (Barry, 1991; Byrd & Tan, 1996; Kühnert & Hoole, 2004; Kul, 2015; Parrell & Narayanan, 2018; Sung & Kochetov, 2018).

Figure 4

Model estimates for the effect of speech rate (x-axis) on the probability of coda glottalization (y-axis). The line shows the median posterior estimate for the probability of coda glottalization at different values of speech rate. The line shading shows one standard error above and below the median. The histogram shows the distribution of speech rate in the data.

3.4.2. Stress

If the following syllable is stressed, the log-odds of coda glottalization increase by 0.49 (median estimate; 95% posterior interval: 0.27 to 0.72). Eddington and Channer (2010) also report that prevocalic word-final /t/ glottalization is more likely when the following syllable is stressed. Stress on the syllable containing the /t/ coda itself did not have a reliable effect on whether the coda was glottalized or not (median estimate = 0.10 decrease in log-odds of glottalization; 95% posterior density interval: –0.47 decrease to 0.28 increase). Marginal median predictions from the model are visualized in Figure 5.

Figure 5

Model estimates for the probability of coda glottalization (y-axis), depending on stress in the syllable containing the coda and in the following syllable (x-axis). Bar heights are median posterior estimates, and error bars are one standard error above and below the median.

We are unsure whether this effect can be described by our proposal. It is possible that stress on the following onset may be associated with reduction of the preceding consonantal coda gesture. Alternatively, phrasal stress (pitch and phrase accents) is also associated with laryngeal constriction (Bird & Garellek, 2019; Campbell & Beckman, 1997; Dilley, Shattuck-Hufnagel, & Ostendorf, 1996), perhaps as a means of achieving a prominent pitch peak. If the constriction is anticipated in the previous syllable, this would enhance the glottal constriction associated with the preceding coda and lead to more audible glottalization. However, this explanation would also seem to predict that accented (or stressed) syllables themselves have more coda glottalization, which we did not find to be the case.

On the other hand, the apparent effect of stress in the following syllable may also be related to the unbalanced distribution of words in the corpus. For example, we observed impressionistically that our data contained many two-word constructions in which the second word had initial stress (e.g., right nów, not réally), but few in which the second word was unstressed (under our criteria for stress; Section 2.5). That is, most such unstressed bigrams were sequences like that they, at the, etc., which are not meaningful constructions. When two syllables comprise a more holistic construction (see Bybee, 2001), a coda /t/ within a cluster between the two syllables may be more reduced (Hay, 2003), permitting more audible glottalization. Thus, the apparent effect of stress may actually result from the fact that there are more two-word constructions that have stress on the second syllable.

3.4.4. Vowel quality

Figure 6 shows the rate of coda glottal stops in syllables with different nucleus vowels, as estimated from the model. There is a general trend for less coda glottalization in syllables with high vowels, though the differences between vowels are very small and unreliable compared to the effects of the following onset consonant (cf. Figure 3).

Figure 6

Model estimates for the probability of coda glottalization (y-axis) when a coda appears after each type of nucleus vowel (x-axis). Bar heights are median posterior estimates, and error bars are one standard error above and below the median.

Laryngeal and epilaryngeal constriction naturally co-occur with tongue lowering and retraction, and glottal constriction is a consequence of laryngeal constriction (Moisik, Czaykowska-Higgins, & Esling, 2019). Coda glottalization may therefore be more audible after low and retracted vowels due to this enhancement of the glottal constriction gesture. Reduction of the oral constriction may also be involved: Lowering the jaw to produce a low vowel increases the distance that oral articulators need to move in order to produce a constriction, and so an oral constriction is more likely to be incomplete after low vowels (see Brown, 2004; Brown & Raymond, 2012; Raymond & Brown, 2012, on the phonetic conditioning of historical Spanish /f/ > /h/). We emphasize, however, that the differences in glottalization rates among most of the vowel types are very small.

3.5. Age and gender of the speaker

Figure 7 shows marginal median predictions for coda /t/ glottalization by age and gender. Younger speakers had median 0.70 increased log-odds of coda glottalization compared to older speakers overall (95% interval: 0.18 to 1.21 log-odds). The overall difference between female and male speakers was smaller and less reliable (0.41 greater log odds for female speakers; 95% interval: –0.09 to 0.90), though the effect of gender may be larger within younger speakers (0.39 further increase in log-odds for female speakers who are younger; 95% interval: –0.59 to 1.39).

Figure 7

Model estimates for the probability of coda glottalization (y-axis), depending on the age and gender of the speaker (x-axis). Bar heights are median posterior estimates, and error bars are one standard error above and below the median.

In several varieties of American English, coda glottalization is reportedly more common for younger speakers (Eddington & Channer, 2010; Eddington & Taylor, 2009; Roberts, 2006), and Kaźmierski (2020) finds the same qualitative pattern for prevocalic word-final /t/ glottalization in the Buckeye Corpus. In some English varieties outside the United States, younger speakers also have higher rates of glottal stop production (Holmes, 1995; Mathisen, 1999 Penney et al., 2019; Smith & Holmes-Elliott, 2018; Stoddart et al., 1999; Watt & Milroy, 1999).

3.6. Word-medial codas

There were 166 tokens of word-medial coda /t/ in our data. The majority (103 tokens; 62%) were transcribed as glottal stops. As in word-final position, glottal stops occurred at the highest rates before approximants (85%) and nasals (70%), and at lower rates before obstruents (38%).

4.1. Model procedure

We fit a multilevel logistic model tree to all of the dictionary word-final and word-medial /t/ codas in our dataset using the glmertree, partykit, and lme4 packages for R (Bates, Mächler, Bolker, & Walker, 2015; Fokkema, Smits, Zeileis, Hothorn, & Kelderman, 2018; Hothorn & Zeileis, 2015; Zeileis, Hothorn, & Hornik, 2008). The procedure uses the model-based recursive partitioning strategy described in Fokkema et al. (2018) to separate the data into subgroups.

In our analysis, the dependent measure is whether a /t/ coda is realized as [ʔ] or not, as defined in Section 2.4. The model-fitting procedure partitions the data into subgroups that have different rates of glottalization. For example, the procedure might choose to partition the data on the basis of phrase position, which might happen if phrase-final and phrase-medial /t/ codas are systematically associated with different rates of coda glottalization. At each step of the procedure, the best two-way partition is selected based on a parameter instability test (Zeileis et al., 2008). Each partition creates two subgroups, and additional partitions within each subgroup are recursively created until there are no more possible divisions that reflect systematic differences in the data, until the subgroups reach a minimum number of observations, or until the tree reaches a maximum depth.

In our model, we included the following phonological features as candidates that could be used to partition the data: the voicing, place, and manner of the following onset consonant (all annotated as ‘none’ for pre-pausal codas); the height and backness of the nucleus vowel; whether the coda was phrase-medial or phrase-final; whether the coda was word-medial or word-final; stress in the target syllable and in the following syllable; as well as articulatory speech rate and phrasal creak. Parameter instability tests for each variable were conducted with Bonferroni-corrected α = 0.05 and post-pruning with the Bayesian Information Criterion. Thus, all partitions represent statistically significant contrasts in the data.

A final logistic regression is then fit to the data within each subgroup, which predicts the rate of glottalization for tokens within that subgroup. For these final regression models, we used only the age and gender of the speaker to predict glottalization rates, since we found impressionistically that these did not interact with the partitioning variables.¹⁰ A multilevel (mixed-effects) model tree (Fokkema et al., 2018) also takes into account group-specific effects (i.e., random effects), such as speaker-specific idiosyncrasies, which are estimated across the entire dataset rather than separately within each subgroup. In our model, we included group-level intercepts for the speaker, word type, and the following consonant type.

4.2. Results and discussion

The partitioning procedure found seven partitions, which are shown in Figure 8. In the tree, the root variable (manner of the following onset consonant) is the most important division within the data, while lower-level divisions are less important. The terminal nodes show the estimated probability of coda glottalization within the final subgroups. For example, the estimated probability of coda glottalization is 63% (third subgroup) when the following consonant is a plosive, is bilabial, and is voiced (following the tree branching from the top).¹¹

Figure 8

Partitions and final subgroups for the logistic model tree. Branches show two-way divisions within a predictor variable that involve significantly different rates of coda glottalization. Terminal nodes show the estimated probability of coda glottalization in each of the seven subgroups, including the confidence interval (CI) and number of observations (N) for each subgroup, averaging over the levels of age and gender. Bar plots below each terminal node show the model estimates for the probability of coda glottalization (y-axis), depending on the age and gender of the speaker (x-axis).

Sonority of the following onset consonant. In predicting the rate of perceptible coda /t/ glottalization, the most important two-way division was whether the following consonant was a sonorant (approximant or nasal) or an obstruent (all other manners). Coda /t/ before sonorants is associated with perceptible glottalization at the highest rate, relative to all other contexts.

Within the sonorants, codas that are followed by palatal /j/ have a lower rate of perceptible glottalization than codas that are followed by any of the other sonorants. This is likely because a /t.j/ sequence is often also pronounced [t͡ʃ], even across a word boundary (Kaźmierski et al., 2016, and see Figure 2). For this reason, the absolute proportion of glottalization before /j/ onsets is lower than before other sonorant onsets. Additionally, young and old speakers seem to have larger differences in this environment than in any other environment, based on the model estimates using age and gender of the speaker (shown in the bar plots below each terminal node).

Place of the following onset obstruent. Within the coda /t/s that were followed by an obstruent, glottalization is perceptible less often before coronal obstruents (or glottal /h/) than before the other obstruents. We argue that this is because the oral constriction of /t/ is reduced in magnitude when it is immediately followed by a second consonant, except when the second consonant is also coronal (Browman & Goldstein, 1990; Sung & Kochetov, 2018) or is /h/ (Kühnert & Hoole, 2004). Thus, due to this reduction of the /t/ oral constriction, the simultaneous glottal constriction is more perceptible before non-coronal obstruents. Additionally, if the following consonant is /h/, the glottal constriction gesture itself may be reduced due to blending with the glottal spreading that is necessary to produce the adjacent /h/ (see also Section 2.5.1). This makes glottal constriction less perceptible before /h/.

Phrase position. Within the coda /t/s that were followed by a coronal consonant (or /h/), glottalization was identified more often phrase-finally than phrase-medially. We argue that frequent irregular voicing in phrase-final syllables (Redi & Shattuck-Hufnagel, 2001) as well as in pre-pausal position (Slifka, 2006, 2007) would enhance the irregular voicing associated with glottal constriction, leading to an increase in the observed rate of coda glottalization. In the model tree, the pre-pausal codas (fourth subgroup) are separated from the phrase-final (pre-coronal) codas (fifth subgroup). This is likely an artifact of the partitioning procedure, because the estimated rate of coda glottalization before long pauses (52%) is nearly the same as the pre-coronal phrase-final codas (48%). We claim that the crucial conditioning context for coda glottalization here is simply phrase-final position, regardless of a pause.

If phrase-final position is important, why does phrase position only significantly divide the data when the following consonant is coronal (rightmost three subgroups in Figure 8)? We believe that phrase-final position also conditions glottalization when the following consonant is labial or velar, but the increase is not apparent in this environment because of a second phrase-related effect. As we argued above, a labial or velar consonant should reduce a preceding oral /t/ constriction and make glottalization more perceptible. However, the oral gesture should be reduced only when the labial or velar consonant is in the same phrase. The net result is that coda glottalization should be more perceptible before labials and velars regardless of a phrase boundary, because (i) phrase-finally, prosodically-conditioned irregular voicing enhances coda glottalization, and (ii) phrase-medially, a labial or velar onset consonant reduces the oral constriction of the coda. Thus, it appears that glottalization is not affected by phrase position in this branch of the model tree.

Voicing of the following consonant. Additionally, glottalization is estimated to be slightly more common before voiced labial and velar obstruents (63%) compared to voiceless ones (52%). Voiced obstruents typically involve supraglottal maneuvers to facilitate phonation, such as an increase in supraglottal volume, even when the phonation itself is absent (Ahn, 2018; Netsell, 1969; Westbury, 1983). If these maneuvers begin early, the consequent increase in transglottal airflow would facilitate irregular voicing when the glottis is constricted (Huffman, 2005).

Moreover, English voiceless obstruents in onset position typically involve glottal spreading. As with /h/ onsets, anticipation of glottal spreading in a voiceless onset should reduce the magnitude of glottal constriction in a preceding voiceless coda stop. This would further reduce the rate of coda glottalization before voiceless compared to voiced labial and velar onsets.

Stress after the coda. As in Section 3.2 and Eddington and Channer (2010), our analysis found that coda glottalization is more likely when the following syllable is stressed.

In all other environments, our account still allows for coda glottalization to be present at some low base rate due to occasional miaslignment or omission of the oral /t/ closure in conversational speech.

5.1. Summary of empirical findings

We explored the distribution of perceptible coda /t/ glottalization in a conversational speech corpus of American English. We found that singleton coda /t/ is pronounced as a glottal stop almost categorically before all sonorant onsets when not deleted, although somewhat less often before /j/, where coda /t/ can alternately be resyllabified into a [t͡ʃ] onset (Kaźmierski et al., 2016). Coda glottalization is common at phrase junctures, though it is even more likely when a sonorant onset follows the juncture. Additionally, glottalization often occurs phrase-medially before labial and velar obstruents, and slightly more often before voiced ones. Glottalization is more likely at higher speech rates across phonological contexts, as well as when the following syllable is stressed.

5.2. Physical and phonological causes of coda glottalization

Our account is based on the assumption that American English voiceless stops in coda position include a glottal constriction gesture that is used to inhibit voicing (Fujimura & Sawashima, 1971; Huffman, 2005; Kahn, 1976; Westbury & Niimi, 1979). We claim that much of the distribution of perceptible coda glottalization can be explained by physical variation in the production of simultaneous oral and glottal constriction gestures. In phonetic environments which favor a reduced oral constriction, or those which favor irregular voicing, the inherent glottal constriction gesture is more likely to cause perceptible coda glottalization. However, we do not believe that coda /t/ glottalization before sonorants is caused by a favorable phonetic environment (contra Huffman, 2005, Section 1.4). In this environment, we claim that reduction or deletion of the alveolar closure to produce a glottal stop is phonologically planned.

5.3. Dialectal variation in coda glottalization

Coda glottalization varies across English dialects in both acoustics and phonological distribution. While the proposal here focuses on coarticulatory mechanics to account for coda glottalization in mainstream American English, a mechanics-based account does not entail that the distribution of coda glottalization should be identical across dialects and speakers. In particular, glottalization of some stops in some phonological environments is commonly associated with social meaning in other dialects (see references in Section 1.1), which largely has not been reported in white mainstream American English (but see Roberts, 2006 on rural Vermont English; Levon, 2006 on Reform American Jewish English; and Farrington, 2018; Fasold, 1981 on African American English). Inasmuch as particular varieties of coda glottalization carry socio-indexical meaning, we expect their production to be increasingly less predictable from mechanical factors. For example, Tyneside English has much higher rates of voiceless stop glottalization than American English (especially for /p/) which almost certainly results from the complex relationship between glottalization and social groups in that variety (Milroy et al., 1994).

Moreover, coarticulation itself is variable, and it depends on other articulatory and phonological patterns in a particular language (e.g., Cohn, 1993). Other varieties of English may reduce oral stop closures at different rates due to different usage patterns (Cohen Priva, 2017; Hay & Foulkes, 2016; Sóskuthy & Hay, 2017), or may specify a different alignment of the oral and glottal constriction gestures (Docherty & Foulkes, 1999a). Both of these will lead to different patterns of perceptible coda glottalization. Some speakers or dialects may also use alternate strategies to inhibit stop voicing. For instance, stop voicelessness can be achieved through glottal spreading rather than constriction (Section 1.3). If a language variety uses glottal spreading with coda /t/, the account discussed here would predict that a misaligned, incomplete, or omitted oral closure could produce audible coda aspiration and pre-aspiration, rather than glottalization (Parrell & Narayanan, 2018). Such coda aspiration patterns are attested in Liverpool English (Clark & Watson, 2016; Watson, 2002) and in Scottish and Welsh English (Gordeeva & Scobbie, 2013; Morris & Hejná, 2019, respectively). Other English varieties glottalize voiceless stops (and voiced stops: Farrington, 2018) in other phonological environments, which may interact with or supersede coda glottalization.

5.4. Word-final prevocalic glottalization

American English coda glottalization is also attested in word-final position before vowels, though at much lower rates than before consonants (Eddington & Channer, 2010; Eddington & Taylor, 2009; Kaźmierski, 2018, 2020; Kaźmierski et al., 2016; Kilbourn-Ceron, 2017; Kilbourn-Ceron et al., 2020; Roberts, 2006; Umeda, 1978). While prevocalic position does not necessarily favor /t/ reduction, glottal stops are often produced before vowel-initial words, especially at phrase junctures and in stressed syllables (Dilley et al., 1996; Garellek, 2013; Pierrehumbert & Talkin, 1991; Umeda, 1978). This glottalization pattern could reinforce coda /t/ glottalization, leading to increased rates of coda glottalization in word-final prevocalic position. Prevocalic coda glottalization should thus be more perceptible at phrase junctures and before stressed syllables.

This prediction is supported by the findings in Kilbourn-Ceron et al. (2020): Prevocalic coda /t/ glottalization is more likely when the word is lengthened (cf. speech rate in Kaźmierski, 2020), when it is followed by a short pause, or when the following word is relatively unpredictable. All of these are strong correlates of phrase junctures (e.g., Turk, 2010, Section 2.4.1), and Eddington and Channer (2010) further find that prevocalic coda /t/ glottalization is more likely before stressed syllables.

Eddington and Channer (2010) and Kaźmierski (2020) also propose that coda glottalization may be phonologically generalized from pre-consonantal environments (e.g., pre-sonorant) to prevocalic ones, as word-final /t/ occurs most often before consonants.

5.5. Other accounts for coda glottalization

5.6. Conclusions and future work

We hand-annotated and analyzed over ten thousand pre-consonantal singleton /t/ codas in a corpus of conversational English speech. Based on the distribution of glottal stops in the corpus, we argued that perceptible coda glottalization in this variety of American English can be understood primarily as a consequence of conditioned variability in the alignment and magnitude of simultaneous oral and glottal constriction gestures. In addition, coda glottalization before sonorant onsets may be a conditioned allophonic variant, as we found no plausible co-articulatory or mechanical motivation for the high rate of glottalization in this environment.