As suggested by continuing and recent reviews (Baumann & Cangemi, 2020; Ladd & Arvaniti, 2023), defining prominence is not an entirely straightforward enterprise. For the purpose of the present study, prominence is considered in two regards.

Firstly, following commonly used terminology from Jun (2005, 2014), a language’s prosodic system can be described as having “head prominence” and/or “edge prominence.” In the former, the expression of prominence is linked to a prosodic head. Relevant to the present study, in American English this head is a metrically prominent syllable in a phrase. Metrically prominent syllables may be marked as phrasally prominent, and produced with a prominence-lending pitch movement (a pitch accent). This sort of prominence will henceforth be described as “phrasal prominence.” Of note, In languages which are described as lacking head prominence, the notion of a prosodic head is not relevant and intonational F0 events demarcate domain (phrase) edges; Ladd and Arvaniti (2023) raise the question if, in languages of this sort, the concept of prosodic prominence is a useful one at all.

Another definition of prominence can be made without reference to metrical structure, or the prosodic features of a particular language. This is a language-general notion of “standing out.” Two definitions are as follows:

The definition in (1) uses the concept of a word, though the same definition could also apply to sub-word units. Both this definition and the perceptual definition in (2) are evidently related to phrasal prominence: a phrasally prominent, pitch-accented syllable/word will be prominent in the sense of both (1) and (2).1 However the definitions are broader in that other properties besides phrasal prominence may also impact the (perceived) “strength” of a word in relation to surrounding material (including, e.g., word frequency as in Baumann & Winter, 2018). One relevant example of this is domain-initial strengthening (e.g., Cho & McQueen, 2005; Keating, 2006; Keating, Cho, Fougeron, & Hsu, 2004). Here the phonetic properties of segments are strengthened in phrase-initial positions, though not necessarily in analogous fashion to strengthening under phrasal prominence (Kim, Kim, & Cho, 2018). These strengthening effects can be seen as enhancing the acoustic/phonetic prominence of a given segment, if prominence is defined as in (1) and (2) above.

As will be described in Section 1.2, vowel-initial glottalization in American English can be related both to phrasal prominence, and to the more general definitions given in (1) and (2). On the one hand, it is probabilistically predicted by phrasal prominence: Phrasally prominent vowel-initial words are more likely to be preceded by glottalization, discussed below. On the other hand, vowel-initial glottalization is also predicted by phrasing, and can be seen as an instance of general acoustic/phonetic prominence strengthening for the following vowel. Both of these views of prominence are thus relevant when considering vowel-initial glottalization effects.

“Glottalization” is used here as a cover term to refer to the production of a sustained closure of the vocal folds, i.e., a glottal stop [ʔ], and localized voice quality changes that are associated with constriction of the vocal folds during voicing (Garellek, 2013; Huffman, 2005). The cover term is useful if we consider the latter of these to be an “incomplete” or lenited glottal stop realization, as is common in the literature (Dilley, Shattuck-Hufnagel, & Ostendorf, 1996; Pierrehumbert & Talkin, 1992).

Of the languages described in the UPSID database (Maddieson & Precoda, 1989), about half use glottalization contrastively (often represented as /ʔ/). However, in many languages that do not use glottalization contrastively, it is well documented that glottalization is nevertheless pervasive in speech, for example in English, Dutch, and Spanish (Dilley et al., 1996; Garellek, 2014; Jongenburger & van Heuven, 1991). An important task for speech research is thus accounting for the prevalence and distribution of glottalization in spoken language.

One clear predictor of glottalization in American English (among other languages) is prosodic organization, both related to prosodic boundaries and prosodic prominence, as noted above. Glottal stops are described as being “inserted” at the beginning of vowel-initial words in prosodically strong positions, where prosodically strong positions include the beginning of a prosodic phrase (Dilley et al., 1996; Pierrehumbert & Talkin, 1992), and in words which bear phrasal prominence (Dilley et al., 1996; Garellek, 2013). Dilley et al. (1996) in particular show that phrase-medial, word-initial vowels in pitch-accented (phrasally-prominent) syllables are glottalized at higher rates as compared to non-prominent equivalents. Notably however, not all pitch accented word-initial vowels are glottalized, and vowels in words which lack pitch-accent but do not have a reduced vowel are more likely to be glottalized than reduced vowels. Speakers also vary widely in their overall rate of glottalization and the extent to which prominence impacts their rate of glottalization. In this sense, glottalization in word-initial vowels is only probabilistically related to phrasal prominence marking, though with a clear tendency to co-occur with phrasal prominence. Redi and Shattuck-Hufnagel (2001) document similar patterns, and consistent inter-speaker variation, and state: “It is clear from these results and from earlier studies that phrase-level glottalization is not obligatory […] glottalization may serve as a marker of ‘degree of finality’ (when it occurs at phrase boundaries) or ‘degree of prominence’ (when it occurs at pitch-accented syllables). Perceptual experiments will be necessary to evaluate the hypothesis that glottalization unrelated to segmental allophony is interpreted by listeners as evidence for a boundary or a prominence, and to determine whether it is interpreted along a continuum or as a contrastive binary feature” (p. 427). The present study addresses both of these perceptual questions.

Garellek (2013, 2014) further suggests a functional motivation for vowel-initial glottalization in American English, using electroglottography (EGG) to examine voicing in vowel-initial words. Garellek (2014) found that phrase-initial vowels, particularly non-prominent vowels, were generally produced with less vocal fold contact during voicing, corresponding to breathy voicing (not glottalization). This suggests, for Garellek’s data at least, glottalization is not having a systematic effect on non-prominent vowels phrase-initially and is more related to prominence marking (cf. Dilley et al., 1996). This effect also became larger at higher-level phrasal domains. Breathier phrase-initial voicing can be attributed to phrase-initial pitch reset, where falling pitch (immediately after reset) results in relaxation of the cricothyroid and thyroarytenoid muscles, and vocal fold abduction (Mendelsohn & Zhang, 2011; Zhang, 2011). Breathier voicing generally leads to decreased intensity and weaker formant energy (Garellek & Keating, 2011; Gordon & Ladefoged, 2001), and Garellek (2014) accordingly proposes that phrase-initial glottalization, most evident in his data in prominent phrase-initial vowels, occurs as a countervailing influence which mitigates the effects of pitch-reset-induced breathiness on voice quality. Glottalization in prominent phrase-initial vowels “strengthens” these vowels, as described by Garellek, in the sense that it engenders more high frequency energy and overall intensity, and boosts frequency information that will be useful in vowel perception (Kreiman & Sidtis, 2011; cf. Garellek, 2013 who found a boost of harmonic energy between 1500–2500 Hz). Glottalization may also be functionally useful in prominence-marking in separating prominent vowel-initial words from surrounding material, and modulating the amplitude envelope in the vicinity of prominent vowels to make them stand out. Preceding silence from a glottal stop will likewise give a boost to listeners’ auditory system at the onset of the vowel (Delgutte, 1980; Delgutte & Kiang, 1984). This view of phrase-initial (and phrase-medial) glottalization implicates (acoustic/phonetic) prominence as a driving force behind it, in that vowels which are preceded by glottalization are enhanced (though this may be either at prosodic domain edges to mitigate phrase-initial breathiness, or at phrasally prominent prosodic heads). In this sense, glottalization in word-initial vowels in American English can be seen as an example of phonetic prominence strengthening, which is related probabilistically to phrasal prominence.

In addition to prosodic prominence, other factors have been shown to influence the rate and distribution of glottalization preceding a vowel in various languages. These include speech rate (Pompino-Marschall & Żygis, 2010; Umeda, 1978) and vowel height (Groves, Groves, Jacobs, et al., 1985; Michnowicz & Kagan, 2016; Pompino-Marschall & Żygis, 2010; Thompson, Thompson, & Efrat, 1974). As documented in German and Spanish, the relative openness of vowels in a vowel hiatus environment predicts the production of glottalization between them: Lower (more open) vowels are more likely to be preceded by a glottal stop (Mckinnon, 2018; Pompino-Marschall & Żygis, 2010). However, relevant to the present study, in American English this is not systematic. Umeda (1978) found no relationship between relative differences in vowel height and production of a glottal stop in a hiatus environment, and Garellek (2013) found that the rate of production of glottal stop in a vowel-initial word was not related to vowel height. Given this, it appears that vowel-initial glottalization is not well predicted by vowel height in American English as it is in e.g., German. This point will be returned to in Section 3.3 in light of the results.

Given the aforementioned patterns attested in the speech production literature, we can now consider some ways in which prosodic information impacts speech perception, and how these prior findings relate to the objectives of the current study.

In some studies, prosodic information (e.g., an intonational tune), has been shown to exert a predictive, or anticipatory, influence on speech processing. For example, Weber, Grice, and Crocker (2006) found that German intonational tunes are used by listeners to disambiguate temporarily ambiguous sentences as S(ubject) V(erb) O(bject) or OVS, prior to critical case information which disambiguated the constituent order. Similar anticipatory effects of pitch accent type were shown by Ito and Speer (2008), where by a prominent (L+H*) pitch accent was interpreted as conveying contrastive focus on one element in adjective-noun pairs, generating anticipatory looks to a referent (e.g., as participants decorated a Christmas tree: “Hang the blue ball, now hang the GREEN ….” generates anticipatory looks to a green ball). Results such as these in Weber et al. (2006) and Ito and Speer (2008) (among others, e.g., Dahan, Tanenhaus, & Chambers, 2002; Nakamura, Harris, & Jun, 2022; Snedeker & Trueswell 2003) indicate that prosodic cues, especially intonational tunes, can be used to anticipate upcoming speech in terms of syntactic, discourse, and information structure.

Complementing this research, the role of prosodic features such as prominence in the perception of speech segments (and relatedly in pre-lexical and lexical processing) has been a recent topic of interest in the literature (Kim, Mitterer, & Cho, 2018; McQueen & Dilley, 2020; Mitterer, Cho, & Kim, 2016; Mitterer, Kim, & Cho, 2019). In comparison to the results described in the preceding paragraph, data in this line of research offers a different view of the way in which listeners use prosodic information in their perception of fine-grained phonetic detail, and their integration of prosody in perception of cues to segmental contrasts. As alluded to above, it is well documented in the speech production literature that prosodic organization modulates cues that are relevant in the perception of segmental contrasts (see e.g., Keating, 2006 for an overview). For example, voice onset time (VOT) in aspirated stops, an important cue for voicing contrasts, varies systematically as a function of prosodic factors. VOT is longer at the beginning of prosodic domains and in phrasally prominent positions (Cole et al., 2007; Keating et al., 2004; Kim, Mitterer, & Cho, 2018). Another example of prosodically modulated cues to segmental contrasts, described in more detail in Section 2, is that of vowel formants. To the extent that phrasal prosody impacts segmental realization along these lines, the listener is hypothesized to benefit from integrating prosodic information in their perception of segmental cues (Kim & Cho, 2013; Mitterer et al., 2016).

A model which has framed this line of inquiry and received empirical support is that of Prosodic Analysis (Cho, McQueen, & Cox, 2007; McQueen & Dilley, 2020). The model’s architecture stipulates simultaneous parses of segmental information and prosodic information from the speech signal, though the role of each of these in processing is different. Adopting an activation-competition view of word recognition, the model postulates that segmental information activates entries in the lexicon, while phrasal prosodic information is used to select among possible lexical candidates. In the original formulation of the model this entails the reconciliation of prosodic boundaries and word boundaries to determine lexical selection (cf. Christophe, Peperkamp, Pallier, Block, & Mehler, 2004). Empirical support for the model comes from studies showing a delayed influence of prosodic boundary information in processing (Kim, Mitterer, & Cho, 2018; Mitterer et al., 2019), consistent with a post-lexical influence in word recognition. This framing of the role of prosody in processing departs somewhat from the anticipatory effects described above, and this follows from the fact that prosodic characteristics are good predictors of sentence and discourse structure as in Weber et al. (2006) and Ito and Speer (2008); however, they are not good predictors of particular lexical items themselves (i.e., generally speaking, a given word can be produced with a range of prosodic expressions, phrase-medially, phrase-initially, and so on). In this sense, the Prosodic Analysis model (and existing data) suggests that prosodic information is not used to anticipate a given word, but is instead integrated with bottom-up cues in lexical processing with a relative delay, consistent with modulation of activated lexical hypotheses. In other words, if the listener’s task is to identify a lexical item (in the absence of other good predictive information), prosodic cues may be integrated in this process but not used to anticipate what word will be said prior to acoustic information about that word is perceived. What the Prosodic Analysis model and available data show more generally is the importance of considering both prosodic and segmental factors as being processed in parallel in speech recognition, with many outstanding questions (see McQueen & Dilley 2020 for a recent overview).

With respect to glottalization specifically, recent perception and processing studies in Maltese, a language in which /ʔ/ is contrastive, suggest that listeners are sensitive to its prosodic patterning in the language (Mitterer et al., 2019; Mitterer, Kim, & Cho, 2021a, 2021b). In addition to marking a phonemic contrast in Maltese, vowel-initial words can be glottalized when they are at the beginning of a prosodic phrase as a form of phrase-initial strengthening. Glottalization thus serves a sort of dual function: It is phonemic and conveys contrast, and also patterns based on prosodic organization. Mitterer et al. (2019) show that listeners are aware of this dual patterning. When a word is phrase-initial, the listener is more likely to attribute the presence of glottalization as being driven by prosody, thus inferring a phonemically vowel-initial word. In contrast, when glottalization precedes a vowel phrase-medially, the listener is more likely to infer that the word is phonemically/contrastively glottalized. Consistent with the prosodic analysis model, these effects were seen to be delayed in time, as assessed in a visual world eyetracking study, and supporting that prediction from the prosodic analysis model. Mitterer et al. (2021a) show that glottal stops differ from other stops (e.g., /t/) in that they do not strongly constrain lexical access, suggesting that listeners’ interpretation of glottalization is intimately linked to prosodic features in a way that differs from other stops. Mitterer et al. (2021b) further show that glottalization is clearly interpreted as a prosodic feature in that it impacts syntactic parsing decisions in the resolution of attachment ambiguity: The presence of word-initial glottalization leads listeners to posit a preceding prosodic boundary, and thus the presence of a syntactic boundary. These results together thus suggest that vowel-initial glottalization can be treated as prosodic cue in perception by listeners, even when glottalization is contrastive.

Steffman (2021a) offers another relevant comparison for the present study. Steffman examined the influence of prosodic prominence on listeners’ perception of vowel contrasts, as cued by the intonational tune and durational patterns of a phrase. Vowels are strengthened phonetically by formant modulations described in Section 2. Steffman thus tested how phrase-level prominence impacted the perception of vowel formants, and further examined the timecourse of its influence. As noted above, in American English, the expression of prominence is related to the placement of pitch accents in a phrase, which are linked to metrically prominent syllables and (in the autosegmental-metrical model of American English intonation, e.g., Pierrehumbert, 1980) are implemented as F0 targets in an intonation contour. Steffman manipulated F0, duration and intensity in a phrase to shift perceived pitch accentuation, and the perceived prominence of a target word, in the stimuli. In one condition, the target word (which was categorized by listeners) was relatively prominent, interpretable as having an (H*) pitch accent in the phrase “I’ll say [TARGET] now” (where [TARGET] indicates the target word; this could be uttered in a broad focus context). In the other condition, the target word was preceded by focus on the verb “say”: “I’ll SAY [target] now,” where “say” bore a prominent L+H* pitch accent (this could be uttered in a contrastive focus context, e.g., A: “Will you write [target] now?”, B: “I’ll SAY [target] now”). In this condition the target is post-focus and non-prominent (more details on the stimuli in Steffman, 2021a are given in Section 4.5.2, which compares that data to the results of this study). This prominence manipulation is one of phrasal/global prominence cues, and was found to impact listeners’ perception of the target vowel in line with the patterns which will be described in Section 2.

Using eyetracking data, Steffman additionally found that, in contrast to the strictly delayed influence of prosodic boundaries documented in previous studies (Kim, Kim, & Cho, 2018; Mitterer et al., 2019), phrasal prominence showed subtle earlier influences in vowel perception, though these effects were quite small, and strengthened over time to be more robust later in processing. The presence of the earlier effect was discussed in Steffman (2020, 2021a) as reflecting prominence processing at multiple stages, described in terms of the Multistage Assessment of Prominence in Processing (MAPP) model. This model proposes that prosodic information needn’t show a strictly delayed (post-lexical) influence in processing as in the Prosodic Analysis model. Instead, early effects reflect “phonetic prominence”: The relative acoustic/phonetic salience of a word (signaled by whatever cues lend prominence in this sense). The fact that the effect was strongest later in time was interpreted as the result of a more abstract/phonological prominence percept (e.g., the presence or absence of pitch-accentuation), which is reconciled with lexical candidates, under the hypothesis that the lexicon contains information about prosodically conditioned pronunciation variants along the lines of Brand and Ernestus (2018); Mitterer et al. (2021a); Pitt (2009). Notably, this multi-stage effect was generated from stimuli that varied both in terms of phonological prominence (pitch accent structure within the phrase), and necessarily, the relative phonetic prominence of the target word. One prediction from the MAPP model is thus that cues which convey only “phonetic prominence,” i.e., vowel initial glottalization, without varying a more global prominence in terms of pitch accent structure etc., should show a clear early effect, and a different online processing pattern than the effect in Steffman (2021a). The present data thus address this prediction from the model directly as a test of how different cues to prominence may be processed differently.

In order to help explain the creation of the stimuli, let us first consider several empirical predictions. If a vowel preceded by glottalization is perceived as prominent, a more prominent acoustic realization of that vowel may be expected by listeners. In this case, it would mean a lower and backer realization of the vowel (with higher F1 and lower F2), with a prominent /ɛ/ essentially becoming acoustically more like /æ/. The corresponding perceptual response would thus be a shift in categorization of the F1/F2 continuum, with more sonorant (lower, backer) F1/F2 values categorized as /ɛ/ in a prominent context (when preceded by glottalization), as compared to a non-prominent one. Empirically, this predicts increased /ɛ/ responses under prominence. Such an effect would constitute perceptual re-calibration for a prominent vowel realization. It is worth noting here that Steffman (2021a) found this effect with the same contrast, when prominence was cued by global/phrasal context as described above.

The materials used in all experiments reported here were created by re-synthesizing the speech of a male American English speaker. The speech used in making the stimuli was recorded in a sound-attenuated booth in the UCLA Phonetics Lab, using an SM10A Shure^TM microphone and headset. Recordings were digitized at 32 bit with a 44.1 kHz sampling rate.

The analysis of categorization data in all experiments reported here was carried out using a Bayesian logistic mixed-effects regression model, implemented with the R package brms (Bürkner, 2017). The models were run using R version 4.1.2 (R Core Team, 2021) in the RStudio environment (RStudio Team, 2021). Weakly informative normally distributed priors were employed for both the intercept and fixed effects (mean = 0, standard deviation = 1.5 in log-odds space).5,6

In reporting effects two measures are given, both characterizing the estimated posterior distribution for a given fixed effect. First we report the estimate and 95% credible intervals (CrI) for an estimate. This gives the effect size (in log-odds), and characterizes the distribution/certainly around the estimate. When 95% credible intervals exclude 0, this suggests a consistently estimated directionality, and accordingly a robust influence. In comparison, 95% credible intervals which include 0 would indicate substantial variability in the estimated direction of an effect, and therefore a non-reliable impact on categorization. An additional metric is reported: The “probability of direction,” henceforth pd, computed with bayestestR package (Makowski, Ben-Shachar, & Lüdecke, 2019). This metric is conceptually similar to reporting CrI, but is useful in that it corresponds more intuitively to a frequentist model’s p-value: The pd indexes the percentage of a posterior distribution which shows a given sign, with values ranging between 50 and 100 percent. A posterior centered precisely on zero (i.e., no evidence for an effect), will have a pd of 50, while a posterior with a skewed negative or positive distribution will have pd that approaches 100. Convincing evidence for an effect would come from pd values that are greater than 97.5 (the pd value that corresponds to 95% CrI excluding zero; a pd value of 100 would indicate all of the distribution for an estimate excludes the value of zero, this would be very strong evidence for an effect). Tables showing all fixed effects estimates for each model are included in the Appendix.

Models were coded to predict categorization responses, with an /ɛ/ response mapped to 1, and an /æ/ response mapped to 0. The formant continuum was coded as a continuous variable, and scaled and centered. In Experiment 1, glottalization was contrast coded with the presence of a glottal stop mapped to 0.5, and the absence of a glottal stop mapped to –0.5. Categorization responses were predicted as a function of continuum step, glottalization, and the interaction of these two fixed effects. In Experiment 2, the glottalization continuum was treated as a continuous variable, and was scaled and centered. Categorization responses were predicted as a function of glottalization continuum, formant continuum, and their interaction. As a control variable, stimulus repetition was also included as a fixed effect, referring to the repetition of a given unique stimulus over the course of the experiment, to account for the possibility of listener bias in categorizing repeated stimuli. This variable was centered and scaled, and ranged from 1–10 in Experiment 1, 1–7 in Experiment 2 and 1–8 in Experiment 3 (due to the different number of repetitions of each unique stimulus). Random effects in the each model included random intercepts for participant and random slopes for all fixed effects and the interaction between glottalization and continuum step.

The results of Experiments 1 and 2 are shown together in Figure 2. In both Experiment 1 and Experiment 2 changing formant values along the continuum shifted categorization in the expected way; increasing (scaled) step values along the continuum decreased the log-odds of an /ɛ/ response (Experiment 1: β = –3.42, 95%CrI = [–3.76, –3.10]; pd = 100; Experiment 2: β = –3.06, 95%CrI = [–3.41, –2.73]; pd = 100). Neither experiment showed a credible effect of the stimulus repetition variable (pd = 72 in Experiment 1, pd = 83 in Experiment 2), indicating that there was not a categorization bias introduced by repetitions of unique stimuli.

In Experiment 1, the glottal stop condition showed a credible effect in shifting categorization (β = 1.74, 95%CrI = [1.30,2.17]; pd = 100). As shown in Figure 1A, a preceding glottal stop increased /ɛ/ responses. This result lines up with the predictions outlined in Section 2.1, suggesting that listeners do indeed adjust their perception of the contrast in line with sonority expansion: A vowel preceded by a glottal stop is expected to be realized as a more prominent variant, i.e., lower and backer in the vowel space.

In Experiment 2, the glottalization continuum additionally showed a credible effect in shifting categorization responses (β = 0.40, 95%CrI = [0.30,0.50]; pd = 100). This is evident in Figure 2B as increasing rightward shifts along the glottalization continuum, with the strongest glottalization cues (step 4) showing the largest difference from step 1 (no glottalization). The results are further shown in Figure 3B, which collapses across all steps of the formant continuum, showing a graded increase in /ɛ/ responses as glottalization cues increase in strength. The effect size (in log odds) is smaller than in Experiment 1, though direct comparisons are not straightforward because of the way that the variables were coded. In Experiment 2, the estimate is for a one-unit change in the scaled value of glottalization continuum step. Relating the scaled and centered values to actual continuum values and comparing the difference between step 1 and step 4 (weakest to strongest glottalization cues) yields an estimated log-odds change of approximately 1.05, suggesting a slightly smaller effect than the full stop in Experiment 1. This may be expected because glottalization cues, even at their strongest in Experiment 2, are in a sense “weaker” than the full stop in Experiment 1. This effect size estimate is in agreement with an alternative parameterization of the model in which glottalization continuum was treated as a four level categorical variable, included in the open access repository.7

There was additionally a credible interaction between continuum and glottal stop condition in both Experiments (Experiment 1: β = –0.77, 95%CrI = [–1.15, –0.41]; pd = 100; Experiment 2 β = –0.26, 95%CrI = [–0.42, –0.12]; pd = 100). The interaction was inspected using the estimate slopes function from the modelbased package (Makowski, Ben-Shachar, Patil, & Lüdecke, 2020), which estimated the marginal effect of the formant continuum across glottalization conditions; see the online repository for code implementing this assessment of the interactions. In Experiment 1, this assessment showed a larger effect of formant continuum step for the glottal stop condition (β = –3.80, 95%CrI = [–4.23, –3.41]) as compared to the no glottal stop condition (β = –3.03, 95%CrI = [–3.39, –2.70]). The same trend was observed for an increase along the glottalization continuum in Experiment 2, where the most glottalized endpoint showed the largest effect of formant continuum step (β = –3.40, 95%CrI = [–3.86, –3.00]) as compared to the not glottalized endpoint (β = –2.71, 95%CrI = [–3.07, –2.38]), with an increase in the effect along the glottalization continuum. A larger effect of formant continuum step is analogous to a steeper categorization slope, and in this sense the presence of the interaction can be taken to suggest that glottalization leads to sharper categorization of differences in vowel formants.8 This makes sense if we consider glottalization as rendering the target vowel more perceptually prominent, though glottalization also simply acoustically sets the target apart from preceding context, which enhances auditory processing as noted above (Delgutte, 1980; Delgutte & Kiang, 1984).

We can now consider the results of Experiment 1 and 2 in relation to the aforementioned relation between vowel height and vowel-initial glottalization, whereby a general cross-linguistic pattern is that lower vowels favor glottalization (e.g., Brunner & Zygis, 2011). On the one hand, this relationship could be treated as a statistical pattern by listeners: Glottalization could lead to the expectation of a lower vowel phoneme (in the present study, /æ/). The results indicate that this is clearly not the case, as glottalization favors perception of /ɛ/. The fact that a lower vowel percept is not favored by preceding glottalization comports with the findings that there is not a predictive relationship between phonological/categorical vowel height and the production of vowel initial glottalization (Garellek, 2013; Umeda, 1978), such that listeners do not use preceding glottalization to identify the vowel as being the lower vowel category /æ/. What the results indicate instead is that vowel-initial glottalization leads listeners to re-calibrate such that the acoustic space which is mapped to a given vowel category is lower and backer (in F1/F2), in line with sonority expansion. This relation to (acoustic) vowel height is a restatement of the predicted prominence effect, though future work will benefit from looking at other vowels, including those which are not realized as acoustically lower/backer under prominence (e.g., American English /i/, Cho, 2005).

The data from Experiments 1 and 2 thus supports the prediction that vowel-initial glottalization serves a prominence-marking function for listeners. Notably, we can see that different realizations of glottalization engender similar perceptual effects, with a clear relationship between strength of glottalization and the magnitude of the perceptual shifts evidenced by listeners. The effect seems to change fairly continuously as a function of the glottalization continuum, addressing “whether [glottalization] is interpreted along a continuum or as a contrastive binary feature” (Redi & Shattuck-Hufnagel, 2001, p. 427).

Experiment 3 made use of the same materials as Experiment 1, though it used a subset of the 10 step continuum. The method by which the Experiment 3 stimuli were selected was the same as that used in Mitterer and Reinisch (2013). The overall interpolated categorization function for Experiment 1 was inspected. The point at which the interpolated function crossed 50% (i.e., the most ambiguous region in the continuum) was identified. The three steps on each side of this crossover point were used in Experiment 3. This led to the selection of steps 4–9 from Experiment 1. There were accordingly 12 unique stimuli used (6 continuum steps × 2 prominence conditions).

Forty participants, none of whom had taken part in Experiment 1 or 2, were recruited from the same population as previous experiments to participate in Experiment 3. Testing was carried out in a sound-attenuated room in the UCLA Phonetics Lab.

Participants were seated in front of an arm-mounted SR Eyelink 1000 (SR Research, Mississauga, Canada) set to track the left eye9 using pupil tracking and corneal reflection at a sampling rate of 500 Hz, and set to record remotely (i.e., without a head mount) at a distance of approximately 550 mm. At the start of the experiment, participants’ gaze was calibrated with a 5-point calibration procedure.

Stimuli were presented binaurally via a PELTOR^TM 3M^TM listen-only headset. The visual display was presented on a 1920 × 1080 ASUS HDMI monitor. In each trial, participants were presented with a black fixation cross (60px by 60px) in the center of monitor. The target words themselves were displayed in 60pt black Arial font, with one word centered in the left half of the monitor, and the other in the right half of the monitor. The side of the screen on which the words appeared was counterbalanced across participants, though for a given participant the same word always appeared on the same side of the screen as in Kingston, Levy, Rysling, and Staub (2016); Reinisch and Sjerps (2013). Two interest areas (300px by 150px) were defined around the target words. These were slightly larger than the printed words, to ensure that looks in the vicinity of the target words were also recorded, following Chong and Garellek (2018) and Kingston et al. (2016).

The onset of the audio stimulus was look-contingent, such that stimuli did not begin to play until a look to the fixation cross had been registered. This was done to ensure that participants were not already looking at a target word at the onset of the stimulus. As soon as a look to the fixation cross was registered, the audio stimulus began, and the target words appeared simultaneously with the onset of the audio. The trial ended after participant provided a click response. The next trial began automatically after a click response was registered. At the start of each new trial, the cursor position was re-centered on the computer screen, following Kingston et al. (2016). Trials were separated by an interval of one second. Eye movements were recorded from the first appearance of the fixation cross until the participants provided a click response and the next trial began.

There were four practice trials, with each continuum endpoint being presented in each prominence condition once. Following this, there were a total of 96 test trials; each of 12 unique stimuli was presented a total of eight times, with stimulus presentation completely randomized. The experiment, including calibration, took approximately 20 minutes to complete.

Given the variables under consideration and the previous accounts of prosody and prominence in processing described in Section 1.3, we can operationalize some predictions for Experiment 3, which will motivate the analyses described below. First, a general expectation is that vowel-internal formant cues should exhibit a rapid influence in online processing as shown, for example, by Reinisch and Sjerps (2013). It takes approximately 200 milliseconds to program a saccadic eye movement (e.g., Matin, Shao, & Boff, 1993), meaning that we expect (at least) a 200 ms lag between the time that a given stimulus dimension is presented to listeners and the time it influences their looking behavior. Given this, we can predict to see an influence of vowel acoustics (modeled with the continuum variable) in online processing as early as 200 ms from the onset of the target vowel, a rapid effect.

Taking this timing as a baseline for what constitutes a rapid effect, consider the timecourse predictions for vowel-initial glottalization from both the Prosodic Analysis model and MAPP model.

Importantly, as described in Section 2.2 the glottalization manipulation only preceded the target vowel in time in the stimuli, and the target itself is acoustically identical across glottalization conditions.

We now turn to the results, first examining the categorization responses and a preliminary look at the eye movement data. Following this, more details on the eyetracking analysis and the eyetracking results are presented in Section 4.5.

As shown in Figure 3 panel A, categorization results from Experiment 3 essentially replicated Experiment 1. Formant cues from the continuum exerted a reliable influence in categorization (β = –2.64, 95%CrI = [–3.01, –2.28]; pd = 100), and we can see the categorization function is overall fairly well-anchored. The glottal stop effect from Experiment 1 was also replicated, with the presence of a preceding glottal stop increasing listeners’ /ɛ/ responses (β = 2.51, 95%CrI = [2.00, 3.04]; pd = 100). An overall bias towards /ɛ/ is also evident in the tendency of listeners to categorize the target as /ɛ/, especially when it is preceded by a glottal stop. As with the previous Experiments, the control variable for stimulus repetition did not show a credible effect (pd = 88). There was additionally a credible interaction between continuum and glottal stop condition, mirroring what was seen in Experiment 1 (β = –0.43, 95%CrI = [–0.78, –0.11]; pd = 99). Comparison of the marginal effect of formant continuum across glottalization conditions also aligns with what was seen in Experiment 1 in showing a larger effect in the glottal stop condition (β = –2.85, 95%CrI = [–3.29, –2.46]) as compared to the no glottal stop condition (β = –2.42, 95%CrI = [–2.81, –2.05]).

Figure 3 panel B shows the eye movement data from the experiment, plotting eye movement trajectories as a function of continuum step and glottalization condition. The measure plotted on the y axis is listeners’ preference to fixate on /ɛ/, computed as the proportion of looks to /ɛ/ minus looks to /æ/ in each 20 ms time bin. Here a value of zero indicates no preference, a positive value indicates a preference to fixate on /ɛ/, and a negative value indicates a preference to fixate on /æ/. Note that the time which is marked as zero on the x axis is the precise point in the stimulus (in either glottalization condition) where there begins to be any difference based on vowel continuum acoustics, corresponding to the positioning of the dashed line in Figure 1. In other words, the stimuli up until this time will be different based on the glottalization manipulation preceding the target vowel, but there are not yet any formant cues to vowel identity at this point. We can see the effect of continuum step in the separation of lines based on coloration, with more /ɛ/-like continuum acoustics leading to a preference to fixate on /ɛ/. This separation, or fanning out, of trajectories appears to occur at roughly 200 ms from the onset of the vowel. The effect of vowel-initial glottalization is also evident in the separation we see based on line type: In line with the categorization data, a preceding glottal stop (dashed lines) facilitates looks to /ɛ/, an online effect corresponding to the categorization results we have seen thus far. We can also note that there is an /ɛ/-bias in eye movements, as also suggested by the categorization data, with steps 1–4 showing a strong /ɛ/ preference. Qualitatively, it thus appears that both vowel-internal acoustic cues, and preceding glottalization, are both shaping listeners’ perception of the target word.

Two complementary analyses of the eyetracking data are presented here. The dependent measure in each analysis was a “preference measure,” which offers a normalized measure of listeners’ propensity to fixate on a target (cf. Reinisch & Sjerps, 2013). This measure is computed as log-transformed looks to “ebb” minus log-transformed looks to “ab,” using the empirical logit (Elog) transformation given in Barr (2008).10 This measure was computed within a given time bin in a trial, the size of which was different in the two different analyses, described below. The analysis window of 0–1200 ms from the onset of the target vowel in the stimulus is used.

In the first eyetracking analysis, eye movement data from Experiment 3 was analyzed by a Generalized Additive Mixed Model (GAMM) using the R packages mgcv (Wood, 2006) and itsadug (van Rij, Wieling, Baayen, & van Rijn, 2016). GAMMs have recently been suggested to offer an appealing alternative to moving window analyses in that they allow for an encoding of the temporal contingency across time bins, and further allow for modeling non-linearity in the data (see Zahner, Kutscheid, & Braun, 2019 for a discussion of the advantages of GAMMs for eyetracking data). The data was sampled at 20 ms time bins for the GAMM analysis (as in Steffman, 2021a; Zahner et al., 2019). The GAMM was an AR1 error model, fit using the technique described in e.g., Sóskuthy 2017, to reduce residual autocorrelation. The rho parameter was specified in the model based on a previous run of the same model without the AR1 component (see the open access repository for code implementing this). The number of knots in some terms were increased (to k = 20) following inspection with the gam.check function, after which the number of knots was adequate as determined by that function.

The model was fit with parametric terms for continuum step (scaled and centered), glottalization condition, and the interaction between these fixed effects. The control variable of stimulus repetition was additionally included. Parametric terms in the GAMM model are analogous to fixed effects in mixed effects models and capture if listeners’ fixation preference in the analysis window as a whole varies as a function of the predictors. Smooth terms in GAMMs are additionally fit to model changes over time, and (potentially) non-linear patterns in the data. The model was fit to capture the interaction between continuum acoustics and time using a non-linear tensor-product interaction term, which allows us to examine how, over time, vowel acoustics mediate listeners’ preference to fixate on a given target. Crucially, this term was interacted with glottalization condition as a “by” term in the tensor-product term, modeling the potential interaction between glottalization, and the influence of continuum acoustics over time. As a control variable, an additional tensor-product term was fit for (scaled) stimulus repetition over time, modeling how the dependent variable changed over time as a function of repetition. This term showed no systematic effect of repetition on looking behavior (in line with previous categorization analyses), so it will not be discussed further, though a plot of the predictions from the model for the influence of stimulus repetition is included on the open access repository. Random effects in the model were specified using the reference-difference smooth method described in Soskuthy (2021), with factor smooths for participant, and for participant by glottal stop condition (coded as an ordered factor). In both factor smooth terms, the m parameter was set to 1, following Baayen, van Rij, de Cat, and Wood (2018) and Soskuthy (2021). The numerical GAMM model output is included in the appendix, though the terms in the model as it was coded are generally not useful for intrepreting timecourse questions of interest here (Nixon, van Rij, Mok, Baayen, & Chen, 2016; Zahner et al., 2019).

The model described above will be compared to one which did not include a non-linear interaction term for glottalization condition with continuum step and time. In this model, glottalization condition was not included as a “by” term in the tensor-product interaction, but instead in a separate smooth modeling the effects of glottalization over time. This latter model thus captures an independent effect of glottalization, but crucially, not an interaction with continuum step. The fit of the two models will be compared in light of the predictions described in Section 4.3, testing prediction 2 from each model. The code for both models and model comparison is contained in full on the open-access repository.

The second analysis presented here is a traditional moving window analysis, which assesses how vowel-internal formant cues influence eye movements in relation to the glottal stop manipulation. The moving window analysis serves the purpose of comparing across Expeirment 3 and Steffman (2021a) with a focus on the relative timing of the influence of continuum step (formants) and the prominence manipulation. Notably, a GAMM model can be used for this purpose too, however in the GAMM numerical time estimates for the influence of continuum step needs to be computed with difference smooths on a pairwise basis, i.e., the timing of an effect between step 1 and step 2, step 1 and step 3, and so on. The moving window analysis thus offers a more global picture of the timing of continuum step. This additional analysis is accordingly carried out to provide converging evidence for the effects in Experiment 3, and to further offer a compact comparison with data from Steffman (2021a). Models were fit for each experiment separately, due to the fact that the continuum acoustics, and the nature of the prominence effects were different across them. Despite these substantial differences, the relative timing of the continuum effect and prominence effect can be considered comparable, given the predictions in Section 4.3. In other words, because the prosodic analysis model predicts a two-stage influence of formants and then prominence we can test this prediction in both Experiment 3 and the data from Steffman (2021a), and evaluate the relative timing of the effects across experiments (prediction 1 from each model in Section 4.3), even though the formant acoustics and prominence cues are not directly comparable to one another. More details from Steffman (2021a) are given in Section 4.5.2, following the presentation of the GAMM results of Experiment 3.

Time bins of 100 ms were used in the moving window analysis, with the preference measure computed at 100 ms intervals across a trial. A 100 ms window was selected because it provides a fairly fine-grained temporal assessment, while also granting a reasonable amount of independence from bin to bin (Barr, 2008; Mitterer & Reinisch, 2013), a known issue in moving window analyses. The dependent measure was predicted as a function of (scaled) formant continuum step, and glottalization context (coded as in the categorization models), and the interaction of these two fixed effects in each time bin. Stimulus repetition was again included as a fixed effect. Random effects were random intercepts for participant and random slopes that were the same as the fixed effects and interaction term. These models were run in brms as with models of the categorization data. The assessment of the models will be in terms of when, over binned time, each has a robust effect on listeners’ fixations, with a focus on the relative timing of continuum step and prominence.

Let us first consider these results as they relate to the hypothesized prominence-marking function of word-initial glottalization in American English in the speech production literature. The presence of glottalization preceding a vowel led to listeners’ expectation of a more prominent (in this case, sonorous) variant of that vowel being produced. Such an expectation led listeners to map acoustically identical formant values to /ɛ/ more often under prominence (versus /æ/). This data thus supports the proposal that glottalization cues prominence to listeners, in line with its implementation as a prominence marker in production. This interpretation more generally accords with Mitterer et al. (2021a; 2021b) in that glottalization is an important prosody-related cue which is recruited in perception.

It is worth noting here that across all conditions in the present experiments the target word was pitch accented, such that the prominence effects seen here suggest different levels of perceptual prominence within pitch accented words, and fine-grained variation in prominence perception as shown in Experiment 2. If we consider “pitch-accented” to be a phonological specification of prominence category, these results speak to the importance of considering within-category variation in perceived prominence as meaningfully impacting the perception of segmental material, in line too with Dilley et al. (1996) showing that pitch accented vowel-initial words are often glottalized, but not always (i.e., there is a probabilistic relationship between pitch accentuation and vowel-initial glottalization). This further raises the question of listeners’ behavior when prominence cues conflict, for example when glottalization precedes an unaccented phrase-medial vowel (possible, but less common as shown in Dilley et al. 1996). This study and Steffman (2021a) showed an effect for two different prominence cues, and one prediction is that these cues are additive when combined, allowing for the possibility of a sort of “perceptual garden path” effect when they conflict. We could thus predict an overall delay in recognition and (potentially) revised fixation behavior in eyetracking as cues unfold, for example if glottalization information precedes the relevant pitch accent information in time. On the other hand, if listeners instead wait until both cues have been received it could be taken to suggest that they are integrating them into a more holistic and abstract prominence percept. Pitting cues (e.g., glottalization and pitch accentuation) against one another in this sense will also allow for testing precedence and possible interactions (e.g., perhaps glottalization is an important cue only when words are pitch accented). Tests of this sort will help us to better understand the ways multiple cues are used in combination by listeners, and hopefully, what sort of representation of prominence is implicated.

More broadly, the results suggest that future research will benefit from considering other patterns of prominence strengthening as relevant in segmental perception. For example, consider the lengthening of VOT in voiceless stops which is observed in prominent syllables (Cole et al., 2007; Kim, Kim, & Cho, 2018). Given the present results we can predict that prominence-signaling lengthening of VOT may impact perception of the following vowel. If found, this would further indicate the importance of fine-grained prominence-strengthening cues in segmental perception. A key takeaway from these results is accordingly the view that prosody should be considered not only in terms of suprasegmental parameters, nor strictly abstract structural terms (phrase boundaries, pitch accents) but should be viewed holistically and as encoded in fine-grained detail and the modulation of cues such as VOT and formant structure.

The eyetracking data further enrich our understanding of the interplay between prosodic and segmental/lexical processing. As noted in Section 1.3, previous examinations of prosodic influences in segmental processing support a delayed influence of prosodic structure, overall consistent with a post-lexical model of prosodic effects (as in the Prosodic Analysis model). Such an account of the present data predicts an asynchronous influence of segment-internal cues to a contrast and prosodic context, with segmental cues preceding prosodic context in the timecourse of their influence. The data in Experiment 3 are not consistent with this account, with simultaneous effects of formants and a preceding glottal stop in online processing.

These data thus present an extension from the Prosodic Analysis model in showing a richer set of prominence effects in segmental/lexical processing than a strictly post-lexical influence. In that sense, they are consistent with the predictions from the MAPP model. The comparison to Steffman (2021a) shows that different prominence cues have different relative timing in processing when compared to vowel-internal spectral information. This is taken as evidence that prominence processing may vary depending on the prominence cue.

Nevertheless, existing data on phrasal prosodic boundaries in processing show clear support for only a later influence of prosodic boundary information in the perception of segmental material (Kim, Mitterer, & Cho, 2018; Mitterer et al., 2019). In this sense, the present data suggest the field will benefit from considering that prominence information and prosodic boundary information may enter differently into processing. One possible view of the asymmetrical role of these prosodic dimensions is that prosodic boundary information is necessarily structural: The listener must determine the presence of a boundary based on phonetic cues, broader phonological context, word boundary information, and syntactic information. Inferences about these levels of representation can be presumed to take place in parallel, and with the consideration of multiple hypotheses, framed recently through the lens of Bayesian inference by McQueen and Dilley (2020).

Phrasal prominence, as defined in Section 1.1, could also be described as structural in the sense that in American English (among other languages) it is determined based on metrical structure and phrasing (e.g., the most prominent pitch accent, the nuclear accent, is the last one in an intonational phrase). However, prominence should also clearly be viewed at a more fine-grained level: The present study shows the importance of considering phonetic prominence, signaled by language-specific cues such as vowel-initial glottalization. In this sense, the determination of a given unit’s prominence therefore needn’t be determined by only a global or phrasal prosodic parse, but instead may be computed by the listener on a syllable-by-syllable basis. Phonetic prominence is thus useful for the listener to determine if a segment has undergone prominence strengthening effects, reconciling the extent to which a segment is perceptually prominent with its acoustics to determine how it should map to a phonemic category. This view implicates perceptual prominence at both sub-lexical and higher levels, in multiple stages of processing.

The MAPP model, as a two-stage model, predicts that structural/phonological versus phonetic prominence effects should be differentiable, and the present data confirm this prediction: Glottalization as a prominence cue is processed early, and differently from more global (and perhaps phonological) prominence distinctions, as shown in Section 4.5.3.

Additional tests of the model and of the nature of prominence in this domain will also benefit from considering how local or distributed cues are in time. Delayed influences in prosodic boundary processing studies (Kim, Mitterer, & Cho, 2018; Mitterer et al., 2019) have been observed with only localized manipulations (e.g., lengthening of just one syllable in Mitterer et al., 2019), such that it is clear that locality does not translate directly into rapid cue use, at least where boundary processing is concerned. Future work addressing questions of cue locality and cue functionality (prominence versus boundary marking) might approach the issue by attempting to cross these parameters and compare local to global prominence cues, as well as comparing local to global boundary cues, within the same experiment.

Additional tests for this sort of distinction between localized/phonetic and global/structural prominence cues could take the form of examining the extent to which each can be modulated by task factors. Certain early effects in processing are assumed to be relatively immune to task effects and cognitive load as shown by, e.g., Bosker, Reinisch, and Sjerps (2017). More global prosodic factors have recently been shown to be influenced by task and stimulus presentation factors (Steffman, 2019, 2021b). For example, Steffman (2021b) found that rhythmic effects in the perception of segmental cues are disrupted when stimuli vary in speech rate, while speech rate effects (typically assumed to result from low-level auditory processing) are robust to rhythmic variation and occur consistently. To the extent that the effects of vowel-initial glottalization seen here reflect early sub-lexical processing, we might expect them to be robust to these sorts of task effects, whereas global prominence effects may be more fragile.

In this vein, one outstanding question is the extent to which localized prominence strengthening effects are related to more general auditory processing. Though glottalization as prominence strengthening is certainly implemented in a language-specific fashion by speakers, it has the effect of making the following vowel acoustically prominent in a more general way (i.e., a vowel preceded by glottalization is rendered louder than, and perceptually more separated from, preceding material) which boosts auditory processing (Delgutte, 1980; Delgutte & Kiang, 1984). Pulling apart the role of language-specific phonetic knowledge and language-general prominence perception may be difficult as phonetic strengthening patterns tend to serve the function of making the strengthened segment more prominent perceptually (though Steffman, 2020 shows that the effects of prominence on vowel perception are specific to the vowel contrast in question). Some indirect evidence for a language-specific interpretation of glottalization cues comes from comparing the early time course of the effect seen here to the delayed influence documented in Mitterer et al. (2019), where a delayed effect is consistent with post-lexical prosodic analysis. This suggests that the processing of glottalization for American English listeners is different from its processing in Maltese. One account for this asymmetry has to do with the function of the glottalization cues in this study as compared to Mitterer et al. (2019). Importantly, in that study listeners’ task was to determine if a word was phonemically /ʔ/-initial. In that sense glottalization was a contrastive cue, the perception of which was modulated by phrasing due to it’s additional phrase-initial boundary marking function. The hypothesis then is that even though vowel-initial glottalization in Maltese may make the following material more phonetically prominent, when the lexical decision depends critically on prosodic phrasing (not prominence), this leads to a relative delay in processing. Carefully controlled cross-linguistic experiments may be useful as a further test of language-general versus language-specific effects going forwards, particularly across languages (and within a language) in which glottalization can have different functions.

In sum, relating the present results to other phonetic strengthening patterns and other languages will help build our understanding of the detailed interplay between segmental and prosodic processing in speech comprehension, and the development of models of this process.