1.1. Perceptual adjustments to phonemic categories

How do listeners adapt to and accommodate variable pronunciations? Listeners’ rapid adaptation abilities have been studied under the umbrellas of perceptual learning or phonemic recalibration—behaviours that allow for listeners’ adjustments to phonemic categories. In the lab, adaptive behaviours have been extensively tested using an experimental paradigm termed lexically-guided perceptual learning, which generally functions as follows: Listeners are exposed to novel or phonetically ambiguous pronunciations of a particular sound in lexical contexts that provide the linguistic scaffolding to guide the interpretation of the intended word. Norris, McQueen, and Cutler (2003) introduced the paradigm using Dutch words containing the fricatives /f/ and /s/. If an ambiguous sound—denoted [?sf]—was presented in the context of a word like witlof “chicory,” listeners expanded their /f/ category to include these pronunciations. Conversely, if the ambiguous [?sf] sound was heard in the context of an /s/ word like naaldbos “pine forest,” listeners adjusted their /s/ category. The adjusted category boundaries were demonstrated in a categorization task with items on a synthesized [s]-[f] continuum. Listeners who heard the ambiguous sound in /f/ contexts showed learning targeted in a particular direction, expanding their /f/ category to specifically accommodate the ambiguous fricative at the expense of /s/, while those who heard the ambiguous sound in the context of /s/ words expanded their /s/ category at the expense of /f/. Crucially, listeners did not make perceptual adjustments when the phonetically ambiguous sound was heard in the context of nonwords, as there was no lexical content listeners could use to make the connection between [?sf] and /s/ or /f/. These adjustments reflect actual retuning of linguistic categories and not exclusively decision criterion shifts (Clarke-Davidson, Luce, & Sawusch, 2008).

There seem to be some limitations to perceptual learning, and further identifying and articulating (some of) those limitations is a goal of the current study. Adaptation seems to be limited, for example, if the phonetic ambiguity of the crucial phoneme is extreme, as is the case of heavily non-native-accented speech, where listeners also do not show adaptation (Witteman, Weber, & McQueen, 2013). Likewise, if listeners do not accurately recognize the words with ambiguous pronunciations as words, they are less likely to learn the novel pronunciations (Scharenborg & Janse, 2013). Adaptation appears limited in its generalization to voices and across-languages, only applying when the voices or cross-linguistic patterns share sufficient acoustic-phonetic or perceptual similarity (Eisner & McQueen, 2005; Kraljic & Samuel, 2005; Reinisch & Holt, 2014; Mitterer & Reinisch, 2017, Reinisch, Weber, & Mitterer, 2013). Together, adaptation requires lexical or other top-down support, in addition to signal-based or bottom-up similarity.

Taking advantage of the natural variation in stop realization within and across languages, scholars have identified some additional limits to the generalization of what is learned. For example, native English listeners naïve to Dutch-accented English not only show adaptation to the devoiced final-stops in Dutch-accented English, but can generalize voiced stop devoicing to initial position as well, at least when not presented with counterevidence to initial devoicing patterns (Eisner, Melinger, & Weber, 2013). Counterevidence inhibited this learning: Listeners do not generalize to initial position when listeners are exposed to items with initial voiced stops, which do not undergo devoicing in Dutch-accented English. When presented with a native English speaker who only exhibited word-final stop devoicing (and not the more global and systematic effects of a non-native accent), listeners showed learning of the pattern, but did not generalize it to word-initial positions, as they had with the Dutch-accented English speaker. These results crucially suggest that the overall accent of a talker informs the generalization strategy. Using Mandarin-accented English, Xie, Theodore, and Myers (2017) demonstrated that adaptation to final /d/-devoicing results in both a recalibration of category boundaries and a reorganization of the category-internal structure. Adaptation was evident in a cross-modal priming task, where listeners showed higher levels of lexical activation for devoiced /d/-final words without negative effect on lexical activation for /t/-final words. These adjustments to Mandarin-accented final-stop devoicing (as observed in the priming data) are different from those results observed by McQueen et al. (2006), who found that listeners completely remapped categories (e.g., negative consequences for doof when training was words like doos).

Adaptation has also been found for vowels (Maye, Aslin, & Tanenhaus, 2008; Witteman et al., 2013; Weatherholtz, 2015; Babel, Senior, & Bishop, 2019). The novel pronunciations used in the work on vowel adaptation implement changes in vowel pronunciation that use or model natural accents, which often fully and categorically replace one vowel quality with that of another vowel category (e.g., tooth pronounced as [tʊθ], not the canonical [tuθ] for North American dialects of English). Some of these vowel studies have tackled the mechanisms of perceptual adaptation by assessing whether exposure to a novel accent adjusts criteria for word identification. For example, Maye et al. (2008) crucially controlled for whether learning was specifically tuned in the targeted direction of the novel pronunciations or whether exposure to a unique accent led to a more general relaxing of criteria for word identification. They accomplished this by testing word endorsement rates of items like [wiʧ] when listeners had been exposed to [wɛʧ] for what is typically pronounced as [wɪʧ] “witch” in North American English. Their results indicate that listeners’ adjustments to the novel pronunciations were targeted in the specific direction of listeners’ exposure, as opposed to a more general relaxation of criteria for what constitutes an acceptable realization of a vowel. Conversely, several experiments in Weatherholtz (2015) do provide evidence of relaxing criteria for vowel categories: Listeners exposed to a back vowel raising chain shift learned an exposure-specific pattern, whereas listeners exposed to a back vowel lowering chain shift showed evidence of category relaxation, accepting items that showed a lowering or raising pattern. While the mechanism behind this exposure-specific category relaxation is not understood, it is not necessarily a surprising outcome, as small distortions to a word’s consonant or vowel pronunciations do not impede successful recognition (Connine, Blasko, & Titone, 1993; Connine, Titone, Deelman, & Blasko 1997; Andruski, Blumstein, & Burton, 1994; Brouwer, Mitterer, & Huettig, 2012). Successful recognition of a word despite a deviant pronunciation, however, may or may not have an effect on the subsequent encoding of that item (Todd, Pierrehumbert, & Hay, 2019).

The specificity of the category—for example whether what is learned is tied to a phoneme-level, an allophonic-level, or an even more surface-level of representation—appears to be sensitive to the exposure conditions (Eisner et al., 2013; Xie et al., 2017; Mitterer & Reinisch, 2017). What remains unclear about the processes involved is whether a more general category relaxation may be a distinct mechanism that also supports perceptual learning, one that complements a more directionally targeted phonemic recalibration mechanism. Zheng and Samuel (2019) present evidence that phonemic recalibration, accent adaptation, and selective adaptation are all distinct mechanisms. Zheng and Samuel’s approach to accent adaptation leverages a non-native accent that has specific shifts in pronunciation that lead to what have been termed ‘bad maps’ (Sumner, 2011). These ‘bad maps’ are categorical mismatches in pronunciation patterns, as in the case of Mandarin-accented English /θ/ pronounced as [s]. In such cases, there is deviation from a listener’s local accent throughout the signal, though crucially, the acoustics-to-phoneme mapping is more disrupted for particular sound categories. Focusing on such categories, Zheng and Samuel found that listeners exposed to Mandarin-accented English sentences were subsequently more likely to call nonwords words (e.g., calling a nonword like [sæŋkfʊɫ] a word), which indicates general criteria relaxation. Support for a general category relaxation mechanism can also be found in the literature on accent accommodation, both in development (Schmale, Cristia, & Seidl, 2012; Schmale, Seidl, & Cristia, 2015; White & Aslin, 2011), and with adult listeners (Baese-Berk, Bradlow, & Wright, 2013). Inconclusive support for expansion mechanisms also comes from computational modelling. Hitzcenko and Feldman (2016) implement Kleinschmidt and Jaeger’s (2015) ideal adaptor framework in an effort to computationally model the Maye et al. (2008) data and assess whether the mechanism(s) at play in the listeners in Maye and colleagues’ study is one of (i) expanding, (ii) shifting, (iii) shifting and expanding, or (iv) remapping. The relevant parameters are the category means and covariance and the model’s confidence in those values (e.g., high or low). The four implementations hit-and-miss the behavioural data to different degrees, but overall, the expand, shift, and expand and shift models all provide decent fits with models that include expansion perhaps aligning somewhat better to Maye and colleagues’ data.

It is important to broach the question of how degree of deviation might limit or constrain the ease with which listeners adapt—or even whether they adapt at all. There are limitations on the acoustic similarity of what kind of substitution is acceptable while still allowing for perceptual adaptation (e.g., Witteman et al., 2013; Babel, McAuliffe, Norton, Senior, & Vaughn, 2019), as is the case with other behaviours that showcase perceptual flexibility (e.g., the phoneme restoration effect; Samuel, 1981). Though note, it is possible that simply increasing the amount of exposure may provide a boost in learning for highly divergent pronunciations. There also seem to be other constraints on lexically-guided perceptual learning with such constraints potentially stemming from a number of underlying causes. For example, Kraljic, Brennan, and Samuel (2008a) illustrate that listeners do not adapt to more [ʃ]-like /s/ sounds in the context of str- clusters. In another example, Kraljic, Samuel, and Brennan (2008b) find that listeners do not shift phonetic categories if ambiguous productions are accompanied by video clips suggesting the speaker had a pen in her mouth during production. They reason that if phonetic variation can be attributed to an entity beyond the speaker (i.e., phonologically conditioned dialect patterns or pen-in-mouth disruptions), listeners will fail to learn the pattern (for a similar account, see Cutler, 2012). Alternatively, if there is not sufficient lexical activation to support the mapping of the novel pronunciation to a phonemic category, boundary adjustments do not occur (Jesse & McQueen, 2011; McAuliffe & Babel, 2016). Thus, the amount or nature of phonetic variation, quantity of exposed items, talker-related attribution (i.e., causal inference), and lack of top-down support may all constrain listeners’ perceptual adjustments.

Lastly, we can also ask whether all speech sounds are updated equivalently in perceptual learning paradigms. The answer to this question appears to be no. Zhang and Samuel (2014) report on a lexically-guided perceptual learning study where listeners were exposed to an ambiguous /s/-/f/ fricative in the context of /s/ or /f/ words in highly predictable sentences. In addition to showing perceptual learning in clear speech, conversational speech, and under two cognitive load manipulations, Zhang and Samuel found an asymmetry across the /s/ and /f/ conditions. Compared to a control condition, listeners exposed to the ambiguous fricative in the context of /s/ words learned the novel pronunciation, generalizing their updated category distribution to the categorization test stimuli. Listeners exposed to the same ambiguous fricative in the context of /f/, however, did not. Zhang and Samuel’s explanation for this hinges on the less robust spectral cues for /f/ compared to /s/, suggesting that listeners may simply be less sensitive to phonetic variation in non-sibilant fricatives. Note, however, that several other studies have found evidence for the perceptual retuning of /f/ in the face of /s/ (e.g., Schuhmann, 2014; Bruggeman & Cutler, 2020; Reinisch & Holt, 2014; and for retuning of /f/ compared to a control group, see Chan, Johnson, & Babel, 2020). Drozdova, Van Hout, and Scharenborg (2016) also uncover an asymmetry in the perceptual retuning of /l/ and /ɹ/ in British English with British English and Dutch (L2 English) listeners. Dutch listeners showed more recalibration for English /ɹ/ than the British English listeners, which Drozdova and colleagues link to the wide range of rhotic variation in (L1) Dutch, reasoning that this larger phonetic range allows listeners to tap into pre-existing variation they have been exposed to, and thus adapt with ease.

This overview of the theoretical and empirical literature on adaptation leads to the objectives for the current study, which are multifaceted. We seek to (i) directly test whether asymmetries in sound patterns affect perceptual adjustments, and (ii) assess whether the adaptation mechanism is targeted in the direction of the exposed pronunciation or reflects a general relaxing of criteria or both. We investigate a potential asymmetry in perceptual learning by testing the learnability of changes in the voiced and voiceless English coronal fricatives: /z/ and /s/. Specifically, we assess whether listeners perform differently when exposed to naturally-produced devoiced /z/ (perceived as [s]) and voiced /s/ (perceived as [z]). This question has roots in questions about learnability in phonology (e.g., the presence of channel biases in phonology; e.g., Moreton & Pater, 2012, Martin & Peperkamp, 2020), and in whether there are boundary conditions on retuning for novel pronunciations in lexically-guided perceptual learning.

We use a lexical decision task as a test of perceptual learning, and quantify differences in word endorsement rates across experimental and control groups for items with a voicing change (i.e., /z/-devoicing or /s/-voicing). This paradigm allows us to quantify both learning specificity and the generalization of the learned (or not) voicing change in novel words not presented during the exposure phase. As opposed to the two-alternative forced choice categorization task often used in phoneme recalibration studies, our use of a lexical decision paradigm to quantify learning has the distinct advantage of allowing us to test whether phonemic adjustments are directional, showing increased word endorsement only in the direction of the exposed change, or whether the adjustments involve general relaxation of category boundaries. To assess whether exposure to devoiced /z/ or voiced /s/ results in directional adaptation or general relaxation, we tested word endorsement rates of both [s] and [ʒ] pronunciations in canonical /z/ words. Endorsement of [ʒ] pronunciations in addition to [s] pronunciations would be evidence in support of a general category relaxation mechanism. Likewise, [ʃ] pronunciations in canonical /s/ words are used to test category relaxation in response to exposure to /s/-voicing. Both the voicing change shifts and the place of articulation shifts are pronunciation changes of one feature edit distance from the canonical pronunciation, and as a result are close to the canonical productions. Following our corpus study in Section 2, we provide much more finely honed predictions of listener behaviour.

1.2. Coronal fricatives in English

English coronal fricatives offer a nice test case because they are clearly asymmetrical in their cross-linguistic distribution and within-English voicing patterns. English coronal fricatives also have attested alveopalatalized variants in casual speech, though these are rare (at least in the Buckeye Corpus [Pitt et al., 2007]; see corpus study in Section 2). Aerodynamic constraints make voiced fricatives less common cross-linguistically compared to voiceless fricatives (Ohala, 1983; Moran & McCloy, 2019). Devoicing patterns are logically assumed to result from the difficulty in maintaining the pressure gradient between the oral and subglottal cavities required to maintain voiced turbulent airstreams. The difficulty of maintaining voicing during fricative production explains why voiceless fricatives are typically longer in duration than their voiced counterparts. These aerodynamic constraints result in devoicing patterns being described as more phonetically natural, particularly for obstruents in utterance-final position (Ladefoged & Johnson, 2011). Also, phonological rules of English fricative final-devoicing are more common than the reverse—within and across varieties of English—though there is the small set of voiceless fricative final words which can voice when coupled with the plural morpheme (e.g., houses can be pronounced as [haʊsəz] or [haʊzəz]; MacKenzie, 2018). In a study of such words, MacKenzie (2018) finds that /s/-voicing rates are dropping precipitously in apparent time, further reiterating the asymmetry between /s/-voicing and /z/-devoicing. Overall, both cross-linguistic and English-specific patterns favour the learnability of /z/-devoicing and the reduced learnability of /s/-voicing, offering a clear test for whether asymmetries in phonemic adjustments are biased by experience.

While there is at least one dialect of English that engages in word-initial fricative voicing (West Country English; Wells, 1982), voicing of phonologically voiceless fricatives is rare both within and across English dialects. Conversely, word-final fricative devoicing is widely documented in varieties of American and British English (e.g., Docherty, 1992; Veatch, 1989; Stevens, Blumstein, Glicksman, Burton, & Kurowski, 1992). Similarly, many languages have phonological patterns where obstruent voicing contrasts are described as neutralizing in word-final positions. Phonetic studies illustrate that these neutralizations are often incomplete, with subtle acoustic-phonetic cues differentiating the underlying obstruent categories (e.g., Warner, Good, Jongman, & Sereno, 2006; Warner, Jongman, Sereno, & Kemps, 2004; and references therein). Kleber, John, and Harrington (2010) demonstrate that naïve listeners are perceptually sensitive to incomplete neutralizations, and use subtle acoustic distinctions in recognition (although far from categorically). In an artificial language learning study, Myers and Padgett (2014) demonstrated that English-speaking listeners more readily learned an utterance-final fricative devoicing pattern compared to a voicing pattern, though participants learned and generalized both patterns to word-final utterance-medial environments. To maintain our focus on phoneme-specific adjustments, we focus on adaptation to fricative voicing changes in word-medial position where the pattern is not tied to morphological alternations, but rather to pronunciation variation attributable to a speaker or dialect.

Smith (1997) examined the propensity to devoice /z/ in English in depth for a small set of speakers (n = 4) in utterances where the target fricatives occurred in word-initial, word-medial, word-final, and utterance-final position. As a testament to the multidimensional nature of voicing contrasts, Smith measured vocal fold vibration with an electroglottograph (EGG), duration of the fricatives and the preceding vowels, and oral airflow. Smith found that her four talkers varied considerably in terms of the degree of /z/-devoicing—one speaker typically produced vocal fold vibration throughout /z/, one speaker produced most of her /z/ tokens with no vocal fold vibration, and two speakers typically vibrated their vocal folds for some portion of the fricative. Three of the four speakers produced a clear duration difference between /z/ and /s/, regardless of whether the /z/ was realized as voiced, partially devoiced, or fully devoiced. Speakers produced longer vowels before phonologically voiced /z/, although three of the four speakers did not produce longer vowels before phonologically voiced /z/ when produced with full voicing. Generally, across all types of /z/ realizations, speakers produced /z/ with lower mean and maximum airflow. Together, these results suggest that the difference between /s/ and /z/ extends well beyond the categorical distinction of whether or not vocal folds are vibrating. Smith summarizes her results as providing evidence that American English speakers can “devoice /z/ in almost any environment” (Smith, 1997: 498).

Thus, not only is there typological evidence that voiced fricatives are more likely to devoice than the reverse, but at least in lab speech, English /z/ is more likely to devoice than /s/ is to voice. There is no literature, to our knowledge, documenting the probability or magnitude of /s/-voicing. This lack of evidence, however, does not mean /s/-voicing does not exist. Thus, to assess /z/-devoicing and /s/-voicing on a categorical level and to quantify listeners’ previous experience with sibilant fricatives and voicing patterns, we consider spontaneous speech, which exhibits large amounts of reduction and pronunciation variation (Johnson, 2004; Pluymaekers, Ernestus, & Baayen, 2005).

2.2. Results

The counts for labels that were automatically and manually applied are shown for /s/ and /z/ words in Tables 1 and 2. There are many instances of applied categories that have small counts and are likely erroneous labels due to mistakes in the human decision process or overly simple assumptions for our automatic labeling (e.g., /s/ as [ɑ]). What is clear from these numbers is that /s/ is much less likely to surface as [z] than /z/ is to surface as [s]. Over 96% of /s/ words have an [s] as their transcribed category, and only 1% of /s/ words were transcribed as [z]. On the other hand, 77% of /z/ words are labeled with [z], and a total of 18% of words were transcribed with an [s] in place of the citation /z/. These data from spontaneous speech confirm prior laboratory findings: /z/ is much more likely to surface as [s] than /s/ is to surface as [z].

The voicing-matched alveopalatal English fricatives are equally likely to surface: [ʃ] and [ʒ] both occur 1.7% of the time for /s/ and /z/, respectively. This equivalence is useful in our adjudication between directional adaptation and general category relaxation, as both substitutions are attested, but equivalently unlikely.

2.3. Discussion and predictions

The results of this corpus study confirm that, at least for this dialect of North American English, listeners are more likely to be exposed to /z/-devoicing than /s/-voicing in spontaneous speech.² The probability of listeners hearing substitutions of [ʒ] for /z/ and [ʃ] for /s/ is low but roughly equivalent. These data allow us to clarify our predictions for how asymmetries in sound patterns will impact adaptation and acceptability of category variation.

We predicted that English listeners would be willing to identify words with /z/-devoicing as words a priori, but that exposure to devoiced /z/ in lexical contexts would cause listeners to be even more likely to identify these items as words. We expected that listeners exposed to devoiced /z/ would update their expectations about the talker in the experiment, and generalize to novel devoiced /z/ words, as this talker would be labelled a ‘devoicer.’ Given that the devoiced /z/ taps into previous experiences with devoiced /z/, we expected that listeners would not generally relax category boundaries for /z/. That is, after being exposed to devoiced /z/, we did not expect listeners to change their threshold for /z/, such that [ʒ] pronunciations would become acceptable.

Conversely, we expected listeners to exhibit a different response for /s/, as fully voiced /s/ is incredibly rare. While we anticipated that exposure to voiced /s/ pronunciations would increase the identification of these items as words, we expected this pattern to be qualitatively different from the word endorsement rates for the more frequent pattern of /z/-devoicing, as this reflects listeners’ experiences with the more probable /z/-devoicing than /s/-voicing; the baseline word endorsement rates for /z/-devoicing are predicted to be much higher than those for the /s/-voicing. Because /s/-voicing is such a rare change in listeners’ experiences, the talker may essentially be labelled as atypical and the adaptation mechanism may be different, reflecting the listener’s lower degree of certainty that the produced [z] indeed maps onto the /s/ category. Under this general relaxation response scenario, listeners may show increased acceptability of [ʃ] pronunciations as well as to novel words with a [z] pronunciation. Given the rarity of [ʒ] for /z/ and [ʃ] for /s/ in spontaneous speech, we anticipated these items as having lower word endorsement rates than items with the voicing change for listeners in the control groups, who were given no reason to adjust their phoneme boundaries or criteria in the exposure phase. The hypotheses described above are summarized in Table 3.

3.1. Word and sentence selection

Target lexical items with non-initial /s/ or /z/ were identified. These items contained only a single sibilant fricative. To reduce ambiguity in the lexical frame, these items were confirmed to not be able to form words if the target fricative was replaced by another fricative (see the list of stimuli in the Appendix). Further, all targets had two to four syllables, and were embedded within moderately predictable sentences containing no sibilant fricatives other than the critical /s/ or /z/. Filler sentences (n = 100) were composed with no sibilant fricatives.

3.2. Audio recordings

An adult female monolingual English speaker produced the sentence and single word materials. These auditory stimuli were recorded using a head-mounted microphone with a SoundDevices USB PreAMP in a sound-attenuated cubicle at a 44.1kHz sampling rate with 16 bit depth. Recorded materials were trimmed of extraneous silence and RMS-amplitude normalized to 70 dB, ensuring no clipped samples. Three versions of each critical /s/ and /z/ sentence were recorded, corresponding to the canonical pronunciation (control), (de)voicing, and alveopalatalized item types. To address the potential for unintended mispronunciations and inconsistencies, the speaker produced several instances of each critical sentence and word. The final recordings were selected based on perceived clarity and consistency by the third author.

4.2. Analysis and results

Participant responses were removed if the reaction time was below 200 ms, or more than three standard deviations above the grand mean—this resulted in the removal of 0.2% of the data.

The experimental data were analyzed using Bayesian multilevel logistic regression models implemented with the brms R package (Bürkner, 2017). The brms package provides a simple interface to the popular and widely-used Stan probabilistic programming language (Stan Development Team, 2021), and adopts the familiar formula specification of models. Bayesian analysis methods are desirable from both conceptual and practical perspectives, as outlined by Vasishth and colleagues in a 2018 tutorial paper on the topic (Vasishth, Nicenboim, Beckman, Li, & Kong, 2018). While similar to frequentist mixed effects models in accounting for variability in the population (‘fixed effects’) and between groups within the population (‘random effects’), Bayesian models allow for graded statements about the strength of evidence for all parameters. A practical benefit is that Bayesian models can be fit with maximal random effects and not run into convergence problems.

Inference in Bayesian models is based on the posterior distributions of parameters, which represent a range of possible parameter values accompanied by probabilities, rather than the point estimates provided in frequentist models. The posterior distribution is a combination of prior expectations (discussed below) and the likelihood of observing the data under the model. The posterior distribution gives information about the size of the effect and the confidence with which the effect size can be interpreted. As the models described in this paper use logistic regression, the posterior distributions represent the possible parameter values in the log odds space. For example, a posterior distribution with most of its probability mass in the [2.3, 4.6] range indicates that the parameter leads to a 10-fold to 100-fold increase in odds for the response variable—that is, the parameter indicates a large effect size. Bayesian models also allow us to express a degree of confidence in an effect’s direction; this is shown in the reporting of the proportion of the posterior distribution with the same direction as the posterior mean. Prose descriptions of confidence are necessarily subjective, and we use the following conventions: Probabilities equal to 1 represent strong or consistent evidence, probabilities between 0.9 and 1 represent moderate evidence, probabilities between 0.8 and 0.9 are weak, and those below 0.8 offer little to no evidence.

In all models, the dependent variable was Word Endorsement, where 1 corresponds to participants’ response of ‘word,’ and 0 to ‘not a word.’ Each of the models had population-level (fixed) effects for Item Type, Condition, and their interaction. Both were weighted effect coded categorical variables, in which levels are compared against the weighted mean. For example, the effect for a specific level of Item Type would indicate that the level differs from the weighted mean across all levels. Weighted effect coding accounts for unbalanced data (i.e., more filler items than critical items) and facilitates the interpretation of both main effects and interactions (Nieuwenhuis et al., 2017). When levels are balanced, weighted effect coding is identical to effect coding.⁵ Levels and weights are reported for each model below. Additionally, in all models, there were by-Participant random slopes for Item Type. This is summarized in the brms formula used for each model: Word Endorsement ~ Item Type × Condition + (1 + Item Type | Participant) + (1 | Word).

Models also shared general specifications. Priors for the intercept and all population-level effects were regularizing, weakly informative Student’s t distributions (𝜈 = 3, μ = 0, σ = 2.5), following the recommendations from Gelman, Simpson, and Betancourt (2017) and Vasishth et al. (2018). This prior largely serves to keep parameter estimates within a sensible range for the log-odds space—that is, approximately 93% of this t distribution falls between log odds of –6.9 and 6.9, which corresponds to probabilities of 0.001 and 0.999, respectively. While extreme parameter values are possible with this prior, the prior is regularizing because it requires more evidence for extreme parameter values to be reflected in the posterior. The default weakly informative brms priors for the group-level and correlations of group-level parameters were used—Student’s t (𝜈 = 3, μ = 0, σ = 2.5), half Student’s t (𝜈 = 3, μ = 0, σ = 2.5), and LKJ correlation (η = 1) distributions, respectively. All models were run with four chains. Each chain had 5,000 iterations (including 2,500 warm-up iterations). This resulted in a total of 10,000 post-warmup samples in each model. Across all models, there were no divergent transitions, and Rˆ values were very close to 1 (and all < 1.05), which indicates that chains were well-mixed and that all models converged. Likewise, the effective sample size for all population-level effects was sufficiently large. For discussion regarding model specifications, see Vasishth at al. (2018), and McElreath (2020: p. 287). In the following sections, results are reported as posterior mean estimates of model parameters (β), along with 95% highest density credible intervals (henceforth, CrI)—the narrowest interval that contains 95% of the posterior distribution. As noted above, the posterior distribution of a parameter indicates a range of possible parameter values and their associated probabilities. While the interpretation of the posterior mean is analogous to the point estimates provided in frequentist mixed-effects models, the posterior distribution as a whole offers a richer picture. As the credible intervals for some parameters span zero, the probability of the effect’s direction is also reported (i.e., if β is negative, the probability of β < 0).

4.2.1. Model: Learning /z/-devoicing

This model assesses whether or not participants learned /z/-devoicing, and excludes critical test items that were not previously heard during exposure. In the model, Item Type was weighted effect coded with four levels (Item Type Filler: Filler Word = 1, [s] = 0, [ʒ] = 0, Nonword = –1.69; Item Type Heard [s]: Filler = 0, [s] = 1, [ʒ] = 0, Nonword = –0.17; Item Type Heard [ʒ]: Filler = 0, [s] = 0, [ʒ] = 1, Nonword = –0.16). Condition was weighted effect coded with two levels (Control = 1, Experimental = –0.76). Participants’ mean word endorsement rates are summarized by Item Type and Condition in Figure 1.

Figure 1

Mean word endorsement rates for filler words, nonwords, and all previously heard critical items across the two conditions used in the analysis of learning /z/-devoicing.

The results of the Bayesian multilevel regression model of word endorsement for the Control and Experimental listener groups’ adaptation to /z/-devoicing are described here and depicted in Figure 2.⁶ To limit the complexity of the interactions, the first model for Experiment 2 focused on word endorsement behaviour for filler words, filler nonwords, and words presented in the exposure phase sentences, which were presented at test with either [s] or [ʒ] pronunciations. Generalization to novel items is addressed later.

Figure 2

Population-level parameters for the learning /z/-devoicing model. Thin lines represent 95% CrI and thick lines represent 50% CrI. The posterior mean estimate for each parameter is indicated by the vertical tick mark.

The model intercept indicates that there is a slight and consistent bias to endorse items as words (β = 2.09, CrI = [1.69, 2.53], Pr(β > 0) = 1). There is evidence that participants in the Control condition have a slightly higher word endorsement rate overall (β = 0.30, CrI = [0.02, 0.59], Pr(β > 0) = 0.98). Filler word items were very likely to be endorsed as words (β = 3.96, CrI = [3.59, 4.37], Pr(β > 0) = 1), and the Condition × Item Type Filler interaction indicates that Control participants were slightly more likely to endorse filler words as words (β = 0.24, CrI = [0.04, 0.48], Pr(β > 0) = 0.99), providing evidence that Experimental condition listeners may have globally adjusted their criteria for word endorsement, becoming more conservative and calling fewer filler items words—we offer possible explanations for this in the interim discussion. Overall, the population-level parameter for Item Type [s] provides evidence that previously heard [s] items are more likely to be endorsed as words (β = 1.39, CrI = [0.36, 2.40], Pr(β > 0) = 0.99). The interaction of Condition and Item Type [s] provides strong evidence that Control listeners were somewhat less likely to identify [s] pronunciations as words (β = –0.76, CrI = [–1.27, –0.28], Pr(β < 0) = 1). This is clear evidence in support of an adjustment to [s] pronunciations in /z/ words for the Experimental participants. While critical items presented as [ʒ] were less likely to be endorsed as words for both experimental and control listener groups (β = –2.68, CrI = [–3.90, –1.55], Pr(β < 0) = 1), the evidence that conditions differed in how they responded to [ʒ] items was weak (β = –0.44, CrI = [–1.34, 0.48], Pr(β < 0) = 0.84). Together with the low estimate for Item Type [ʒ], this indicates that, overall, listeners in both conditions were very unlikely to endorse these items as words, though note the wide range of variability in Figure 1.

4.2.2. Model: Generalizing /z/-devoicing

A second model addressed the question of whether participants generalized their learning of /z/-devoicing to novel words, that is, words not heard during the exposure phase. For this model, Filler Words and Nonwords were excluded, and all critical items included (both Heard and Novel). All aspects of the model structure were identical to that of the previous section, with the following exceptions. Item Type was weighted effect coded with four levels (Item Type Novel [s]: Heard [s] = –0.97, Novel [s] = 1, Heard [ʒ] = 0, Novel [ʒ] = 0; Item Type Heard [ʒ]: Heard [s] = –0.94, Novel [s] = 0, Heard [ʒ] = 1, Novel [ʒ] = 0; Item Type Novel[ʒ]: Heard [s] = –0.94, Novel [s] = 0, Heard [ʒ] = 0, Novel [ʒ] = 1). Condition was weighted effect coded with two levels (Control = 1, Experimental = –0.75). Participants’ mean word endorsement rates are summarized by Item Type and Condition in Figure 3.

Figure 3

Mean word endorsement rates for previously Heard and Novel critical Item Types (both [s] and [ʒ] pronunciations) across the two conditions used in the analysis of generalizing /z/-devoicing.

The model intercept indicates that there is an overall bias towards endorsing items as words (β = 1.04, CrI = [0.47, 1.63], Pr(β > 0) = 1). There is strong evidence that novel words pronounced with [s] are endorsed as words (β = 1.16, CrI = [0.59, 1.76], Pr(β > 0) = 1), but little to no evidence that this interacts with Condition (β = 0.13, CrI = [–0.28, 0.56], Pr(β < 0) = 0.73). That is, listeners in the experimental condition did not generalize their learning of /z/-devoicing to novel /z/ words pronounced with [s]. Overall, listeners were less likely to endorse Heard (β = –1.74, CrI = [–2.35, –1.14], Pr(β < 0) = 1) and Novel (β = –2.23, CrI = [–2.82, –1.68], Pr(β < 0) = 1) items where /z/ was pronounced as [ʒ] as words. The evidence that this interacted with Condition is weak for Novel [ʒ] items (β = 0.20, CrI = [–0.18, 0.59], Pr(β < 0) = 0.85) and non-existent for Heard [ʒ] items (β = 0.06, CrI = [–0.42, 0.56], Pr(β > 0) = 0.61). These results are depicted in Figure 4.

Figure 4

Population-level parameters for the generalizing /z/-devoicing model. Thin lines represent 95% CrI and thick lines represent 50% CrI. The posterior mean estimate for each parameter is indicated by the vertical tick mark.

4.3. Discussion

Regardless of condition assignment, listeners were very likely to identify /z/ words with devoiced [s] pronunciations as words. Listeners in the experimental condition who were exposed to devoiced /z/ in training were more likely to endorse heard /z/ words with devoicing as words, indicating that they had adjusted their thresholds for acceptable or identifiable realizations of /z/ in a directional manner for the items they were exposed to. However, there was no evidence for generalization for listeners in the experimental group: While listeners in the experimental condition were more likely to identify previously heard devoiced /z/ words as words, there was no evidence that novel /z/ words pronounced with [s] were more likely to be identified as words. This outcome highlights the potentially word-specific nature of perceptual adaptation.

In the generalization model, there was weak evidence that control listeners were more likely to endorse novel /z/ words with [ʒ] pronunciations as words than experimental listeners, suggesting not only that experimental condition listeners used a directional adaptation mechanism to adjust their /z/ category, but also that control listeners exposed to multiple novel pronunciations at test may have relaxed their criteria for /z/ in novel words across the course of the test phase, as they were, for the first time, exposed to this talker producing /z/ as [s] and then [ʒ]. This suggests that while listeners did not make any general category relaxation for /z/ in the context of lexical items they heard as [z] in the exposure phase, there is weak evidence that /z/ category boundaries were slightly relaxed to include [ʒ]-like pronunciations for novel lexical items. When listeners had not previously been presented with counter-evidence for specific lexical items, control listeners show a tendency towards category relaxation in the test phase. This pair of outcomes indicates that the same underlying category can be subjected to different mechanisms depending on the nature of the variation in the input. Simply, the nature of the input triggers the adaptation mechanism.

Note that there was moderate evidence that listeners exposed to devoiced-/z/ pronunciations in the exposure phase were less accurate on the categorization of filler words. This is a small (well below a doubling of odds) but consistent effect, and may be due to an apparent difference in the word-nonword balance across conditions. Experimental participants have learned to perceive target items (which are part of the nonword count in the stimuli distribution) as words; they identify fewer filler words as words. An alternative possibility is that by virtue of learning that a talker makes certain pronunciation changes, listeners may anticipate additional pronunciation changes which render licit words as nonwords. Both of these interpretations are speculative and warrant future consideration, and stem from a small effect.

In summary, as expected, both listener groups were likely to call devoiced /z/ items words, as they are all exposed to such pronunciations in their natural input. Listeners who were exclusively presented with devoiced /z/ in the exposure phase were even more likely to identify these items as words. Their adaptations were limited to the specific items they were exposed to and in the direction of variation to which they were exposed. That is, /z/-devoicing was not extended to novel /z/ items that were not presented in the exposure sentences and alveopalatalized [ʒ] pronunciations of /z/ were not more likely to be identified as words. This lack of generalization was unexpected, and we discuss this further in the general discussion.

5.2. Analysis and results

The analysis for Experiment 3 was nearly identical to that of Experiment 2. Less than 0.2% of the data was removed due to reaction time filtering. The models for Experiment 3 have the same formula, weak priors, and specifications as described in Section 4.2. For reference, the formula was: Word Endorsement ~ Item Type × Condition + (1 + Item Type | Participant) + (1 | Word).

5.2.1. Model: Learning /s/-voicing

This model assesses whether or not participants learned /s/-voicing, and excludes critical test items that were not heard during exposure. In the model, Item Type was weighted effect coded with four levels (Item Type Filler: Filler Word = 1, [z] = 0, [ʃ] = 0, Nonword = –1.69; Item Type Heard [z]: Filler = 0, [z] = 1, [ʃ] = 0, Nonword = –0.16; Item Type Heard [ʃ]: Filler = 0, [z] = 0, [ʃ] = 1, Nonword = –0.16). Condition was weighted effect coded with two levels (Control = 1, Experimental = –1.12). Mean word endorsement rates are summarized by Item Type and Condition in Figure 5.

Figure 5

Mean word endorsement rates for filler words, nonwords, and all previously heard critical items across the two conditions used in the analysis of learning /s/-voicing.

In the model of learning /s/-voicing, the intercept indicates that there is an overall bias towards endorsing items as words (β = 1.79, CrI = [1.41, 2.20], Pr(β > 0) = 1), though there was little to no evidence of a meaningful difference across conditions (β = 0.09, CrI = [–0.12, 0.31], Pr(β > 0) = 0.80). There is strong evidence that filler words were consistently endorsed as words (β = 4.13, CrI = [3.77, 4.52], Pr(β > 0) = 1), and that this interacted with Condition. Control participants were slightly more likely to endorse filler words as words (β = 0.32, CrI = [0.16, 0.49], Pr(β > 0) = 1), suggesting that listeners in the Experimental group may have globally adjusted their criteria for word endorsement, becoming more conservative with respect to what constitutes a word, as in Experiment 2. There is weak evidence that Heard [z] items were less likely to be endorsed as words (β = –0.64, CrI = [–1.76, 0.47], Pr(β < 0) = 0.87), and furthermore, the Condition × Item Type [z] parameter indicates that Control participants were less likely to endorse Heard [z] items as words (β = –0.57, CrI = [–1.27, 0.10], Pr(β < 0) = 0.95). This provides some evidence that listeners in the Experimental group adjusted their /s/ criteria to accommodate [z] pronunciations. There is strong evidence that [ʃ] items were less likely to be endorsed as words overall (β = –3.38, CrI = [–4.76, –2.06], Pr(β < 0) = 1). Additionally, Control participants were less likely to identify [ʃ] items as words (β = –0.51, CrI = [–1.43, 0.40], Pr(β < 0) = 0.87), though given the probability of the effect’s direction, this should be interpreted as weak evidence that Experimental participants generally relaxed their word endorsement thresholds. The population-level parameters for the model of learning /s/-voicing are depicted in Figure 6.

Figure 6

Population-level parameters for the learning /s/-voicing model. Thin lines represent 95% CrI and thick lines represent 50% CrI. The posterior mean estimate for each parameter is indicated by the vertical tick mark.

5.2.2. Model: Generalizing /s/-voicing

This model addresses the question of whether participants generalized their learning of /s/-voicing to novel words—those not heard during exposure. As in Section 4.2.2, Filler Words and Nonwords were excluded, while all Heard and Novel critical items were retained. The same model structure was used. Item Type was weighted effect coded with four levels (Item Type Novel [z]: Heard [z] = –0.99, Novel [z] = 1, Heard [ʃ] = 0, Novel [ʃ] = 0; Item Type Heard [ʃ]: Heard [z] = –0.99, Novel [z] = 0, Heard [ʃ] =1, Novel [ʃ] = 0; Item Type Novel [ʃ]: Heard [z] = –1.01, Novel [z] = 0, Heard [ʃ] = 0, Novel [ʃ] = 1). Condition was weighted effect coded with two levels (Control = 1, Experimental = –1.14). Participants’ mean word endorsement rates are summarized by Item Type and Condition in Figure 7.

Figure 7

Mean word endorsement rates for previously Heard and Novel Item Types (both [z] and [ʃ] pronunciations) across the two conditions used in the analysis of generalizing /s/-voicing.

In the model of generalizing /s/-voicing, there is a slight overall bias to call items nonwords (β = –0.87, CrI = [–1.75, –0.04], Pr(β < 0) = 0.98), unlike each of the previous three models. Overall, there was some moderate evidence that novel /s/ items pronounced with [z] are more likely to be endorsed as words (β = 0.32, CrI = [–0.19, 0.84], Pr(β > 0) = 0.90), and strong evidence that [ʃ] pronunciations for /s/ items are less likely to be endorsed as words, whether Heard (β = –0.95, CrI = [–1.55, –0.41], Pr(β < 0) = 1) or Novel (β = –1.16, CrI = [–1.80, –0.56], Pr(β < 0) = 1). There is moderate evidence that Control participants are less likely to endorse items as words across the board (β = –0.57, CrI = [–1.30, 0.17], Pr(β < 0) = 0.94), suggesting that Experimental participants may have generally relaxed their criteria for /s/ which manifests here as greater acceptance of [ʃ] pronunciations. There was no evidence that Condition interacted with Item Type, as each interaction term substantially overlaps with zero, possibly because the pattern is well-captured by the main effect of Condition. The population-level parameters described here are depicted in Figure 8.

Figure 8

Population-level parameters for the generalizing /s/-voicing model. Thin lines represent 95% CrI and thick lines represent 50% CrI. The posterior mean estimate for each parameter is indicated by the vertical tick mark.

5.4. Discussion

Listeners were assigned to an experimental condition where /s/ words were produced with [z]—a voiced fricative at the same place of articulation—or to a control condition where a typical [s] was heard in the same sentence set. At test, all listeners were presented with /s/ items from exposure containing /s/-voicing, novel /s/ words with voicing, heard /s/ words pronounced with [ʃ], and novel /s/ words with [ʃ]. Bayesian multilevel logistic regression models present very weak evidence that listeners exposed to voiced /s/ adapted their /s/ category to specifically accommodate these pronunciations. That is, there is little to no evidence of a directional adjustment mechanism in play in response to exposure to /s/-voicing. However, there is weak-to-moderate evidence that, at test, listeners exposed to the voiced /s/ in exposure were more likely to call any non-canonical pronunciation of an /s/ word a word than listeners in the control condition, suggesting that any change in /s/ category structure was the result of a more general relaxation of /s/ criteria, as opposed to any directional adjustments towards voiced /s/.