Work on lexically guided perceptual learning has demonstrated that listeners can readily adapt to idiosyncratic phonetic shifts produced by individual talkers (Samuel & Kraljic, 2009). Participants in a typical adaptation experiment complete an exposure phase (often a lexical decision task) followed by a test phase (usually a categorization task). When listeners are placed in a talker-specific condition (i.e., the same speaker is presented at both exposure and test), test phase categorization is altered based on the lexical bias in exposure. Consistently replacing /d/ with an ambiguous segment between /d/ and /t/ in exposure (e.g., ‘croco?ile’) results in more /ada/ response along an /ada/-/ata/ continuum at test, while replacing /t/ with the same ambiguous segment (e.g., ‘fron?ier’) leads to a greater /ata/ response. Talker-specific adaptation is a highly replicated phenomenon that underscores the flexible nature of the perceptual system (Cummings & Theodore, 2023; Norris et al., 2003; Zellou et al., 2023).

Perceptual learning can be neatly accounted for through the ideal adapter framework (Kleinschmidt & Jaeger, 2015). An ideal adapter leverages their prior knowledge about acoustic cue distributions to make the most accurate predictions possible about the incoming speech signal. Importantly, these predictions can be updated through experience, allowing the listener to adjust to novel input in the face of high acoustic variability. For example, after exposure to a talker who repeatedly produces /d/ as an ambiguous segment between /d/ and /t/ (e.g., ‘croco?ile’, ‘legen?ary’, etc.), listeners can then develop a talker-specific mental model. The model would predict that segments with ambiguous acoustic cues (e.g., voice onset time between prototypical /d/ and /t/) are more likely to be classified as /d/ for this talker compared to the general population. Listeners can then utilize this talker-specific model to help them make future predictions (e.g., that an ambiguous token between /ada/ and /ata/ is more likely to be /ada/ when produced by this particular talker).

Unlike talker-specific adaptation, cross-talker generalization is much more inconsistent (i.e., after exposure to a phonetic shift or accent, listeners may or may not be willing to apply what they have learned to a novel talker; Weatherholtz & Jaeger, 2016). The ideal adapter framework can explain this relative restraint as striking a balance between being efficient while also maintaining accurate predictions (Kleinschmidt & Jaeger, 2015). On the one hand, if listeners are initially exposed to a talker with a phonetic shift and then hear a novel test talker who likely produces the same shift, it would be more efficient to generalize – with little decrease in prediction accuracy, the same mental model can be applied to both talkers without expending any extra effort in creating a new model. Overgeneralizing, however, can lead to incorrect predictions if the test talker is unlikely to produce the same phonetic shift. Taken together, the ideal adapter framework predicts that listeners somehow formulate expectations about the relationship between exposure talkers and novel talkers, and that these beliefs then modulate cross-talker generalization.

Although it is uncontroversial that cross-talker generalization is constrained by the relationship between the properties of the exposure and test phases, the specific constraints are still under debate (Baese-Berk et al., 2020). The current study manipulates one type of well-studied constraint (the number of exposure talkers), but re-examines prior work from a sociolinguistic perspective.

Bradlow and Bent (2008) proposed that multiple exposure talkers are necessary for cross-talker generalization in perceptual adaptation to L2-accents. Relative to a control condition (exposure to L1-English speakers), transcription accuracy in noise for a novel Mandarin-accented English speaker was only facilitated for listeners with recent exposure to multiple Mandarin-accented talkers, not for participants who had previously heard a single Mandarin-accented talker. Kleinschmidt and Jaeger (2015) attributed this effect to the distinct mental models that listeners develop following single-talker and multi-talker exposure. On the one hand, participants who are exposed to just one talker might assume that the particularities of the accent are idiosyncrasies, thereby resulting in a talker-specific model that does not generalize to novel talkers. However, listeners who are exposed to multiple talkers with the same L2-accent can more readily recognize the covariation in the speech signal between category membership (Mandarin-accented English speaker) and acoustic variation (e.g., devoicing of word-final stop consonants). This recognition ostensibly leads to the development of a talker-general mental model for Mandarin-accented English, which then gives rise to successful generalization (i.e., enhanced comprehension of a novel talker with the same, Mandarin-accented English accent).

In contrast to Bradlow and Bent (2008), however, more recent work has found no effect of the number of exposure talkers on generalization (Xie & Myers, 2017; Xie et al., 2021). Participants in Xie and Myers (2017) completed an exposure phase (auditory lexical decision task) followed by a test phase (cross-modal priming task). Everyone in the experiment was tested on the same (novel) Mandarin-accented English speaker. In the critical conditions, listeners were either exposed to multiple Mandarin-accented English speakers, a single Mandarin-accented English speaker that was acoustically similar to the novel test talker, or to a single Mandarin-accented English speaker that was acoustically different from the test talker. Two major findings emerged: (i) both the multi-talker exposure and the ‘single-talker, acoustically similar’ conditions resulted in cross-talker generalization, with no meaningful difference observed between the two conditions; (ii) generalization was blocked in the ‘single-talker, acoustically different’ condition. Xie and Myers (2017) concluded that acoustic similarity between exposure and test talkers mediates generalization, not simply the number of exposure talkers. Multi-talker exposure only leads to generalization when one of the exposure talkers happens to be acoustically similar to the novel speaker.

Providing further evidence against the primacy of the number of exposure talkers, Xie et al. (2021) conducted a replication study of Bradlow and Bent (2008) and did not find a meaningful difference in transcription accuracy between the critical single- and multi-talker exposure conditions. Xie et al. (2021) attribute this lack of replication to “the removal of the design confound [in Bradlow and Bent (2008)] coupled with increased statistical power” (p. e37), where the “design confound” refers to “the single- and multitalker conditions employ[ing] different L2 exposure talkers” (p. e36). In other words, the lack of generalization in the single-talker exposure condition of Bradlow and Bent (2008) may have merely occurred because the authors coincidentally selected exposure talkers that were all acoustically different from the test talker (and conversely, the generalization in the multi-talker exposure condition might have been blocked if none of the exposure talkers were acoustically similar to the test talker).

While not downplaying methodological considerations, there could additionally be a social explanation for why the number of exposure talkers has not affected generalization in recent work. Xie and Myers (2017) and Xie et al. (2021) both examine Mandarin-accented English, and in general, the proportion of L2-accented English speakers is rising in the United States (Graddol, 2003; ShareAmerica, 2023). Although Xie and Myers (2017) recruited “monolingual English speakers with…no or minimal prior experience with Mandarin-accented English or the Mandarin language” (p. 33) and Xie et al. (2021) “excluded participants from analysis who reported a high degree of familiarity with Chinese or Chinese-accented English” (p. e27), many listeners likely had some experience with other L2-English accents.

L2-English accents often share acoustic properties, such as a slower speaking rate (Baese-Berk & Morrill, 2015) and reduced usage of spectral cues when producing tense/lax vowel contrasts (Feng & Wang, 2024; Sidaras et al., 2009). Therefore, despite limited self-reported experience with Mandarin-accented English, listeners in Xie and Myers (2017) and Xie et al. (2021) could have initially engaged in accent-independent adaptation (Baese-Berk et al., 2013), having recognized certain features in the exposure phase (e.g., devoiced word-final stop consonants are features of both Mandarin-accented and Dutch-accented English; Eisner et al., 2013; Xie & Fowler, 2013). After this head start, participants may have attuned more to the specific acoustic properties of the exposure talker, with generalization only occurring when the exposure and test talkers were sufficiently similar acoustically.

Given their likely prior experience with L2-English accents, listeners in Xie and Myers (2017) and Xie et al. (2021) presumably did not consider the acoustic properties in exposure to be entirely idiosyncratic, which could make the number of exposure talkers irrelevant. The distinction between single- and multi-talker exposure conditions might only be important when listeners are truly unfamiliar with a variant (e.g., the /p/ to [b] shift discussed in Section 1, where a novel L1-English speaker produces words like ‘poster’ as boster). A novel phonetic shift might be initially thought of as idiosyncratic if it is heard in only one talker. To prove that a variant is socially-mediated and generalizable to novel talkers, it could be necessary to hear the shift from multiple talkers within the same social group.

Thus, it is still unclear whether multiple exposure talkers are necessary for cross-talker generalization. L2-accented speech may not be the most appropriate test case for addressing this question, given the ubiquity of L2-accented speakers. An ideal test case could be a variant that is likely to be treated as an idiosyncrasy when produced by one speaker and as a socially-mediated feature when produced by more than one speaker from an identifiable social group.

The current study examines the covariation of (binary) gender and stop consonant production in L1-accented English, with females producing /p/ as [b] and males producing prototypical /p/.1 There are several reasons for employing this specific test case. For one, gender is a highly salient and recognizable socio-indexical cue (Barreda & Predeck, 2024; Eckert, 1989). Male and female voices are often acoustically distinct (Barreda, 2021), and listeners can use the (on average) lower f0 and formant frequencies of males to identify speaker gender with high confidence and near-ceiling accuracy (Hillenbrand & Clark, 2009). Listeners also know, consciously or not, that gender can be relevant for speech perception in certain cases (e.g., an ambiguous sound between /s/ and /ʃ/ is more likely to be categorized as /s/ if the talker is male; Yu, 2010).

Critically, however, compared to other phonological contrasts, such as fricatives and vowels, comparatively little covariation currently exists between gender and American English stop consonant production (Kleinschmidt, 2019; Morris et al., 2008). Among studies that do find gender effects (e.g., Robb et al., 2005; Swartz, 1992), the result is almost always the exact opposite of the covariation we have constructed for the current study – male speakers usually produce voiceless consonants with a shorter voice onset time than female speakers (i.e., it is male speakers who produce /p/ as a more acoustically [b]-like sound). A further distinctive feature of the current test case is the use of a “bad map”, or a sound that “fall[s] unambiguously into an unintended category” (Sumner, 2011, p. 132). Shifts from /p/ to [b] can be heard in English, but usually in L2-accented English (Flege & Eefting, 1987; Solé, 2018), not L1-accented English.

In summary, the current study correlates a highly salient variant (a “bad map” /p/ to [b] shift) with a highly salient social cue (speaker gender) in a way that listeners have likely never experienced in L1-English. Unlike L2-accented English, which has been the focus of many prior experiments on adaptation (Xie & Myers, 2017; Xie et al., 2021), there is a greater chance that the phonetic shift in this study will be truly regarded as an idiosyncrasy upon initial exposure. The current study can therefore more effectively investigate how the number of exposure talkers affects cross-talker generalization, while additionally lending insight into the emergence of sociolinguistic perception.

Two experiments were conducted to examine how the number of exposure talkers and the socio-indexical structure of exposure influence cross-talker generalization. All participants completed an exposure phase (identification of phrase-final keywords; e.g., “The wall needed a new coat of paint”) followed by a test phase (categorization of stimuli along a buy-pie continuum for both a novel female and novel male speaker). Within each experiment, there were two types of exposure conditions that varied in socio-indexical structure: (i) a critical condition with a gender-mediated phonetic shift (/p/ produced as [b]; e.g., ‘paint’ as baint); (ii) a control condition with no shifted keywords. Across experiments, the critical conditions differed in the number of exposure talkers – whereas Experiment 1 only presented one female and one male talker, Experiment 2 exposed listeners to two female and two male talkers. The key question is whether the experimental manipulations induce cross-talker generalization. When comparing the control and critical conditions, is there a difference in buy-pie categorization for the novel test talkers?2

The present study adjudicates between three possible accounts of perceptual adaptation: (i) talker normalization; (ii) sufficient similarity; (iii) numerosity. The predictions of each account are summarized in Table 1 and further explained in 1.4.1–1.4.3.

Table 4 and Figure 1 present the aggregated test phase results and model summary statistics, respectively. There was a consistent effect of Step, which demonstrates that as continuum step increased, listeners selected pie more often. There were no other meaningful main effects or interactions. Notably, no interactions between Speaker Gender and Exposure Condition surfaced, implying that the exposure conditions (Female Shifted and No Shift) did not differ in pie responses by speaker gender.

Experiment 1 placed listeners in either a critical or control condition (labelled as the Female Shifted and No Shift conditions, respectively). Both conditions exposed participants to a single male and a single female talker, but differed in whether the female talker produced a /p/ to [b] phonetic shift (the critical condition) or remained unshifted (the control condition). The male exposure talker was unshifted for all listeners. In the test phases for both the critical and control conditions, no difference in buy-pie categorization was observed between the novel female and novel male talkers, reflecting a lack of cross-talker generalization.

The clearest finding from Experiment 1 is that the results go against a sufficient similarity account (Xie & Myers, 2017; Xie et al., 2021). Generalization did not occur from the phonetically shifted female exposure talker to the novel female test talker, even though both were similar in terms of acoustics (higher f0) and social category (same gender). The absence of generalization is especially notable given the wording of the instructions in the present study. Speaker gender was explicitly mentioned before both the exposure and test phases, which in theory could have biased listeners towards generalization based on shared social category membership.

The Experiment 1 results could be explained by one of two theoretical accounts. First, in accordance with a talker normalization account (Joos, 1948; Liberman & Mattingly, 1985), perhaps listeners in the critical condition overlooked social cues in the speech signal, disregarding the systematic gender covariation in the exposure phase (i.e., that the /p/ to [b] phonetic shift was only produced by the female exposure talker, not the male exposure talker). If speaker gender is ignored, then the exposure phase would be perceived as containing conflicting information, where /p/ is produced as a canonical [p] half of the time and as [b] for the other half of trials. This type of scenario has resulted in a null effect in recent work (e.g., Tzeng et al., 2021), which would be realized in the current study as no difference in test phase categorization across talkers and conditions.

Alternatively, the absence of generalization in Experiment 1 could be explained by a numerosity account (e.g., Bradlow & Bent, 2008). The /p/ to [b] shift was designed to be unattested in L1-English (see 1.3 for details), and since the production of /p/ as [b] was only heard in one exposure talker, listeners may have considered the shift to be a talker-specific trait that does not generalize to a novel talker. According to a numerosity account, generalization should only occur if listeners hear the /p/ to [b] shift in more than one talker, as it would confirm that the shift is not merely an idiosyncrasy.

Although the findings of Experiment 1 challenge a sufficient similarity account and match the predictions of talker normalization and numerosity accounts, (see Table 1), a null effect is not sufficient evidence to make a firm conclusion (Vasishth & Gelman, 2021). This issue is addressed in Experiment 2, which is designed to tease apart the talker normalization and numerosity accounts.

Exposure phase accuracy was nearly at ceiling (99.70% across all conditions). The model summary statistics and aggregated test phase results for Experiment 2 are shown in Table 5 and Figure 2, respectively.

A meaningful main effect of Step emerged, indicating that listeners provided more pie responses as step number increased. There was also a consistent main effect of Block Order, such that on average, listeners who heard the female block first made fewer pie responses. Critically, however, the effect of Block Order was modulated by a 3-way interaction between Speaker Gender, Block Order, and Exposure Condition.

To examine this 3-way interaction, the hypothesis function from brms was employed (a method of inspecting interactions without running additional post-hoc models, as in a frequentist analysis; for details, see Chapter 7 of Barreda & Silbert, 2023). The hypothesis compared the 2-way interaction between Speaker Gender and Exposure Condition for each block order. A meaningful interaction with a positive coefficient was found for listeners who heard the female block first [β: 0.26, SE: 0.11, 95% HDI = (0.05, 0.47)], but not for listeners who heard the male block first [β: –0.09, SE: 0.10, 95% HDI = (–0.29, 0.11)]. In other words, relative to the No Shift control condition, participants in the Female Shifted exposure condition gave more pie responses for the female speaker than for the male speaker, but only if they heard the female block first (i.e., Female → Male, not Male → Female).

Looking just at the top-left graph of Figure 2, it becomes visually evident how listeners assigned to the Female Shifted exposure condition and Female → Male test block order provided a greater pie response for the female speaker relative to the male speaker. However, when this top-left cell is compared to the other three cells within Figure 2, the adaptation effect might initially appear to be more nuanced – the three-way interaction seems to be driven by a reduced pie response for the novel male speaker, rather than a greater pie response for the novel female speaker. This was investigated statistically through an additional hypothesis examining the 2-way interaction between Block Order and Exposure Condition for each gender. There is ultimately no clear evidence to support this impressionistic observation of Figure 2, as there were no meaningful interactions found for either the male test speaker [β: –0.22, SE: 0.14, 95% HDI = (–0.49, 0.05)], or the female test speaker [β: 0.13, SE: 0.12, 95% HDI = (–0.11, 0.36)]. (That said, there is a marginal interaction for the male test speaker, which accounts for why the meaningful effect of Block Order in the main model has a negative coefficient.)

Unlike Experiment 1, Experiment 2 found evidence for cross-talker generalization, with listeners in the Female → Male test block order providing more pie responses for the novel female test talker than for the novel male test talker. The only difference between the two experiments was the number of exposure talkers – Experiment 2 presented two female and two male exposure talkers, while Experiment 1 only exposed participants to one female and one male talker. It therefore seems that exposure to multiple talkers of each gender is necessary for generalization in this case, contrary to findings reported in recent work (Xie & Myers, 2017; Xie et al., 2021).

Taken together, the results of Experiments 1 and 2 provide the clearest support for a numerosity account for cross-talker generalization. Since the phonetic shift is heard in multiple female speakers in Experiment 2, listeners might generate a mental model that groups the female exposure speakers together (either through shared acoustic features or shared social category membership) and designate a /p/ to [b] shift as a property of female speech in the context of the experiment. The talker-general nature of this model then results in generalization of the shift to a novel female talker, who is more similar (acoustically and socially) to the female exposure speakers. Overall, the current study finds a critical role of the number of exposure talkers on generalization, bolstering numerosity accounts and challenging both talker normalization and sufficient similarity accounts.

There are two caveats of the analysis and results in Experiments 1 and 2. First, generalization was somewhat short-lived in Experiment 2. A test phase categorization shift only occurs among listeners assigned to the Female → Male block order (i.e., when the female talker was presented first in the test phase), not among those placed in the Male → Female order (i.e., when there is a delay in presenting the novel (female) talker who is more similar to the shifted (female) exposure talkers). There are at least three different, non-mutually exclusive explanations for this order effect.

One explanation involves the particular test case used in the critical condition of the current study. As discussed in 1.3, there is no attested covariation in L1-English between a bad map stop consonant shift and speaker gender, such that female speakers produce /p/ as [b] and male speakers remain unshifted. The lack of familiarity with this type of phonetic shift may have made it more challenging to learn, resulting in more ephemeral generalization. A more robust effect might be observed if one or more aspects of the present test case are adjusted to align with attested L1-English phenomena (e.g., if the male speakers produced /p/ as [b] and the female speakers were unshifted; Robb et al., 2005).

Another possibility is that perhaps there was not enough exposure to result in prolonged generalization. Listeners in the critical condition of Experiment 2 heard 16 phonetically shifted stimuli from only two female talkers during a several-minute-long exposure phase. Given the novelty of the phonetic shift, greater exposure may be needed for the shift to be retained in memory. A more stable effect could be observed if listeners were presented with more stimuli6 (Cummings & Theodore, 2023), completed several iterations of this task over multiple sessions (Xie & Kurumada, 2024), and/or heard more than two shifted talkers (Bradlow & Bent, 2008).

Finally, the transitory nature of generalized adaptation in Experiment 2 could also be explained by the structure of the exposure phase. Listeners heard a mixture of both filler and critical stimuli from two different male and two different female talkers in a randomized order, rather than in blocks. Perhaps presenting the exposure sentences in a blocked order (either by talker or by gender) would have made it easier to keep track of the speakers and to learn that only the females were phonetically shifted. This account accords with work showing detrimental effects of talker-switching on speech perception (Magnuson et al., 2021) and with the intuition that blocked training may be more helpful at the incipient stages of learning (for discussion, see Raviv et al., 2022, p. 476).7

Besides the order effect in Experiment 2, a second caveat is that so far, only the aggregated results have been shown, not the results by individual test talker. As mentioned in 2.1.3, the three female and three male talkers were fully counterbalanced so that each was presented as a test talker equally often. There is a minor possibility that the results genuinely support a sufficient similarity account (rather than a numerosity account), but that the aggregated analysis is masking this finding. The following section investigates this idea in greater depth and then presents a post-hoc analysis of both experiments.

The aggregated data analysis in 3.2 showed evidence of cross-talker generalization in Experiment 2 (i.e., greater pie response for the novel female than for the novel male speaker), but only among listeners who heard the novel female block first in the test phase. Given this order effect, the post-hoc analysis data only contained participants assigned to the Female → Male block order. Furthermore, there were no meaningful interactions involving continuum step in either experiment, so Step was removed as a predictor variable from the post-hoc models.

As depicted in Equation 2, all post-hoc models contained fixed effects of Speaker Gender (within-subjects; Female, Male), Exposure Condition (between-subjects; Female Shifted, No Shift), and their interaction, by-speaker and by-listener random intercepts, and by-listener random slopes for Speaker Gender. A separate model was fitted for each of the nine possible exposure-test talker combinations across both experiments (six models for Experiment 1, three models for Experiment 2). Although this post-hoc analysis decreases statistical power, the number of subjects analyzed in each model (Experiment 1 Models: n = 33, on average; Experiment 2 Models: n = 62, on average) is still comparable to the sample sizes of prior work with 80% power (Cummings & Theodore, 2023).

If cross-talker generalization has occurred, then there should be a meaningful interaction between Speaker Gender and Exposure Condition with a positive coefficient (i.e., indicating that greater pie response was provided for the novel female talker than the novel male talker, but only in the Female Shifted exposure condition). The credible intervals in each post-hoc model for this interaction are listed in Table 8. The other fixed effects are not directly relevant for the current research question and are thus omitted for clarity (but see the Data accessibility statement to examine all of the statistical models).

Meaningful interactions were only present for two of three multi-talker exposure conditions (Joanna + Salli → Ruth, Joanna + Ruth → Salli). In other words, when there was only one exposure talker in Experiment 1, none of the individual exposure-test talker pairs facilitated cross-talker generalization.

The results of the post-hoc analysis are summarized in Table 9. Overall, clear support is provided for a numerosity account over a sufficient similarity account. No evidence of generalization was observed between any of the six female exposure-test talker combinations in Experiment 1, where only one talker was presented in the exposure phase. A sufficient similarity account would attribute this null effect to a lack of sufficient similarity between all female talkers to each other in the current study (i.e., the Joanna voice is dissimilar to both the Ruth and Salli voices, and the Ruth voice is dissimilar to the Salli voice). In theory, if the exposure phase presents two female talkers as in Experiment 2, and both are dissimilar to the test talker, then generalization should be blocked (Xie & Myers, 2017; Xie et al., 2021).

However, this prediction is not borne out. Experiment 2 found robust evidence of cross-talker generalization, at least for two of the possible exposure-test talker combinations (Joanna + Salli → Ruth, Joanna + Ruth → Salli). This suggests that, as predicted by a numerosity account (Bradlow & Bent, 2008), exposure to multiple talkers per gender is necessary to generalize an unfamiliar phonetic shift to a novel talker.

Why was no generalization observed for the Experiment 2 participants who were exposed to Ruth and Salli and tested on Joanna? One tentative possibility is that effects of acoustic similarity are at play. Even though Joanna has a higher f0 than all three male speakers in the current study (see Table 3), t-tests demonstrate that the test stimuli of Joanna have a lower f0 than both the exposure stimuli of Ruth (mean difference = 22.88 Hz, t = 23.953, p < 0.001) and the exposure stimuli of Salli (mean difference = 18.37 Hz, t = 20.585, p < 0.001). The exposure stimuli of Ruth and Salli, meanwhile, are more similar to each other acoustically. Although Salli does have a consistently greater f0 than Ruth (mean difference = 4.50 Hz, t = 5.58, p < 0.001), the mean difference is reduced compared to the pairwise comparisons between Joanna/Ruth and Joanna/Salli. Perhaps Joanna was not sufficiently similar acoustically to Ruth and Salli, which blocked generalization among listeners in Experiment 2 who were exposed to Ruth and Salli and later tested on Joanna.

However, a sufficient similarity account is not entirely satisfactory. Assuming that Ruth and Salli are more similar to each other, it becomes unclear why no generalization took place in Experiment 1 for listeners who were exposed to Ruth and tested on Salli (Ruth → Salli) or for participants exposed to Salli and tested on Ruth (Salli → Ruth). Additionally, a sufficient similarity account would predict generalization to be weaker among listeners in Experiment 2 who were either exposed to Joanna and Salli and tested on Ruth (Joanna + Salli → Ruth) or exposed to Joanna and Ruth and tested on Salli (Joanna + Ruth → Salli). All participants in the critical conditions across Experiments 1 and 2 heard the same number of phonetically shifted stimuli (n = 16). Whereas all of these /p/ to [b] shifted tokens were produced by a single female exposure talker in Experiment 1, they were evenly distributed across two female exposure talkers in Experiment 2 (n = 8 each). If Ruth and Salli are truly the only two female speakers who share sufficient similarity, then the Experiment 2 listeners in the Joanna + Salli → Ruth and Joanna + Ruth → Salli conditions were exposed to fewer sufficiently similar tokens than the Experiment 1 listeners in the Salli → Ruth and Ruth → Salli conditions (n = 8 in Experiment 2 versus n = 16 in Experiment 1). Reduced exposure generally leads to less robust adaptation (Cummings & Theodore, 2023), so according to a sufficient similarity account, generalization should have been less strongly facilitated in the Joanna + Salli → Ruth and Joanna + Ruth → Salli conditions of Experiment 2 compared to the Salli → Ruth and Ruth → Salli conditions of Experiment 1. Yet, the opposite pattern is depicted in Table 8 – generalization was only facilitated in the Joanna + Salli → Ruth and Joanna + Ruth → Salli conditions of Experiment 2 and blocked in the Salli → Ruth and Ruth → Salli conditions of Experiment 1. Overall, the acoustic differences between Joanna and Ruth/Salli do not appear to adequately explain the current results.

An alternative explanation could involve the properties of the test continua themselves, which are depicted in Figure 3. Although efforts were made to match the continua as closely as possible (see Appendix 2), Figure 3 shows that listeners in the No Shift condition (where no phonetic shift was presented in the exposure phase) are already heavily biased towards a pie response for Joanna compared to Ruth and Salli (Joanna: 67.86% pie response; Ruth: 51.93% pie response; Salli: 51.89% pie response).8 The presence of generalization is also indicated by a greater proportion of pie responses, so it is possible that there was a ceiling effect – among the Experiment 2 listeners who were tested on Joanna, listeners in the Female Shifted condition could not provide a greater pie response than the No Shift condition, resulting in a null effect. The test phase stimuli for Ruth and Salli are not nearly as biased towards a pie response, so participants tested on Ruth and Salli in Experiment 2 showed meaningful generalization.

Although it is true that there are certain acoustic disparities among the female speakers (e.g., Joanna has a lower f0 than Ruth and Salli), the differences in f0 between the female and male speakers are much starker (see Table 2). It therefore seems more probable for listeners to group Joanna with one of the other female speakers than with the male speakers, based on (relative) acoustic and social similarity. All participants in the critical condition of Experiment 2 heard two shifted female and two unshifted male talkers in the exposure phase, and in the Joanna + Salli → Ruth and Joanna + Ruth → Salli conditions, listeners likely placed Joanna within the same mental model as Salli/Ruth. In accordance with a numerosity account, listeners attuned to the number of talkers with the /p/ to [b] shift and, given that the shift was heard in more than one exposure speaker, participants then generalized the shift to the most similar test talker (i.e., the novel female speaker).

Taken together, the results of the post-hoc analysis are more consistent with a numerosity account than a sufficient similarity account. A numerosity account more satisfactorily rationalizes why cross-talker generalization solely occurs with multi-talker exposure in Experiment 2 and never with single-talker exposure in Experiment 1. The absence of generalization in one of the three possible exposure-test talker combinations in Experiment 2 seems to reflect idiosyncratic variation in the test phase continua, not evidence against a numerosity account.

The present findings address the ongoing debate about whether multi-talker exposure is necessary for cross-talker generalization. In a seminal paper, Bradlow and Bent (2008) offered evidence in support of this claim: comprehension in noise of a novel Mandarin-accented English talker only improved for listeners exposed to multiple talkers of the same accent, not for listeners exposed to a single talker. In a replication of Bradlow and Bent (2008), however, Xie et al. (2021) did not observe any difference in perceptual performance between single and multi-talker exposure conditions. This later finding corroborated Xie and Myers (2017), supporting the hypothesis that the degree of acoustic similarity between the exposure and test talkers mediates generalization, not the number of talkers in exposure. Now, seemingly in contrast to Xie and Myers (2017) and Xie et al. (2021), the current study indicates that multi-talker exposure may be required for generalized adaptation in some instances.

Xie et al. (2021) account for the lack of replicability of the multi-talker exposure benefit by pointing to three methodological issues of Bradlow and Bent (2008): (i) relatively low statistical power (n = 87 across five between-subjects conditions); (ii) the use of different exposure talkers in the single- and multi-talker exposure conditions; (iii) the presentation of only one test talker for all participants. These factors cannot explain the current findings because a large number of listeners were recruited (n = 383 in Experiment 1, n = 386 in Experiment 2), the exposure talkers were drawn from the same pool of talkers for all conditions, and test talker identity was evenly counterbalanced. Therefore, the presence of cross-talker generalization in only Experiment 2 of the current study, when listeners heard multiple talkers per gender in exposure, cannot merely be attributed to the confounds present in Bradlow and Bent (2008).

Even though the empirical result of this work differs from Xie and Myers (2017) and Xie et al. (2021), these apparent contradictions can be reconciled by appealing to a sociolinguistic framework. More specifically, the types of exposure conditions that induce generalization might vary based on the amount of experience listeners have with the presented variants/accents. Both the current work and prior work might be situated along the following timeline: (i) when listeners have little or no prior experience with a variant (e.g., the current study), then multi-talker exposure is necessary for generalization; (ii) when listeners have some prior experience with a variant (e.g., work on L2-accent adaptation; Xie et al., 2017; Xie et al., 2021), then some relevant exposure is necessary for generalization, but either a single talker or multiple talkers can be heard in exposure; (iii) when listeners have extensive prior experience with a variant (e.g., work on sociolinguistic perception; Niedzielski, 1999; Strand & Johnson, 1996), then no exposure is needed for generalization.

On one end of the spectrum, the current study presented a phonetic variant that was intentionally designed to be unattested in L1-English (a “bad map” /p/ to [b] shift covarying with gender; see 1.3 for details). Given their unfamiliarity with this variant, listeners in Experiment 1 likely considered it to be a talker-specific idiosyncrasy when heard in a single female exposure talker, which blocked generalization towards a novel female talker. Generalization was only observed in Experiment 2, when participants heard the /p/ to [b] shift in multiple female exposure talkers and could thus presume that the shift was not merely a talker-specific trait. Even with multi-talker exposure, however, the generalization effect dissipates within only several minutes (see 3.2 and 3.3 for details), suggesting that listeners are reluctant to generalize when they have had little to no prior experience with a variant.

Xie and Myers (2017) and Xie et al. (2021), meanwhile, presented Mandarin-accented English to either undergraduates at a diverse American school (University of Connecticut; Xie & Myers, 2017) or L1 speakers of American English (Xie et al., 2021). Even if participants self-reported a lack of familiarity with Mandarin-accented English, many likely had at least some previous exposure to other accents or dialects, considering the prevalence of L2-accented speakers in the United States (Graddol, 2003). Participants could have therefore discerned accent-independent properties within the speech signal (e.g., slower speaking rate, difficulty producing the tense/lax vowel contrast in English; Baese-Berk et al., 2013), deducing that the acoustic features of the exposure talkers are likely part of some broader L2-accent, not entirely idiosyncratic. Compared to participants in the current experiments, listeners are much less conservative when exposed to L2-accented speech – given the appropriate conditions (e.g., sufficient acoustic similarity between the exposure and novel talkers), generalization can be facilitated with either single- or multi-talker exposure (Xie and Myers, 2017; Xie et al., 2021) and can even occur with a 12-hour delay between the exposure and test phases (Xie et al., 2017).

On the opposite end of the spectrum as the current study are experiments on sociolinguistic perception, in which listeners hear variants that they already have ample experience hearing in their everyday lives (D’Onofrio, 2015; Niedzielski, 1999; Strand & Johnson, 1996). These cases do not require any exposure phase at all for generalization to be triggered. By merely presenting a single cue to social identity, listeners can generalize their prior knowledge of sociolinguistic variation to a novel speaker. For example, participants’ mental models of speaker gender and sibilant production are so robust that, without any prior exposure phase during an experiment, their categorization of tokens along an /s/-/ʃ/ continuum is altered based on whether a gender-ambiguous voice is presented with a stereotypically male or female face (Munson, 2011; Strand & Johnson, 1996).

In summary, highlighting the role of listeners’ social experiences not only unifies the current study with Xie and Myers (2017) and Xie et al. (2021), but also ties perceptual adaptation to sociolinguistic perception. This explanation also adds nuance to the literature, shifting from a binary debate about whether multi-talker exposure is necessary for cross-talker generalization (Bradlow & Bent, 2008; Xie et al., 2017; Xie et al., 2021) to a broader discussion about when multi-talker exposure is necessary for cross-talker generalization.

According to a numerosity account, listeners in the critical condition of Experiment 2 grouped the two phonetically shifted female exposure talkers together and generalized the shift towards the test talker who was the most similar (i.e., the novel female test talker, not the novel male test talker). A key question is whether this judgment of similarity is determined by acoustics or by social category membership. The current study cannot tease apart these two possibilities because female and male speakers are acoustically distinct (e.g., through differences in f0; see Table 3) and can also trigger differing categorical judgments about apparent gender (Hillenbrand & Clark, 2009).

Prior work on cross-talker generalization has found evidence for both types of mechanisms. As an example of a socially-mediated effect, Aoki and Zellou (2023a) exposed listeners to Mandarin-accented English and observed that generalization towards a novel Mandarin-accented English speaker was facilitated when both the exposure and novel talkers were presented with an image of an East Asian face (i.e., similarity in apparent ethnicity). Meanwhile, in the absence of visual cues, Xie and Myers (2017) suggested that generalization of Mandarin-accented English stop consonants was based primarily on acoustic similarity, considering that listeners could not identify the accent of the speakers. Both acoustic and social similarity can lead to generalization, and it is thus possible that both are motivating the effects in the current study.

Future work on cross-talker generalization could explore what factors mediate the relative weighting of acoustic versus social similarity. For instance, there might be a critical role played by the types of voices that are presented. The current work deliberately focused on gender as a social category because it is highly salient perceptually and easily recognized by adult speakers (Hillenbrand & Clark, 2009). By contrast, Xie and Myers (2017) and Xie et al. (2021) examined cross-talker generalization of Mandarin-accented speech. While naive listeners generally demonstrate above-chance performance in identification and discrimination tasks for L2-accents, accurate categorization of a speaker’s accent is often lower than for gender (Atagi & Bent, 2017). Therefore, generalization effects for gender and L2-accented speech could be inherently different, with the former recruiting the influence of social categories (since listeners can easily match exposure and test talkers by gender) and the latter relying more heavily on acoustic mechanisms (since listeners are unsure whether the exposure and test talkers actually share the same accent).

Building upon Aoki and Zellou (2023a), however, presenting socially relevant visual cues might modulate effects of acoustic similarity. Generalization between two acoustically different Mandarin-accented English talkers in Xie and Myers (2017) could be enhanced if both speakers share the same apparent ethnicity (and by the same token, a diminished generalization effect might be observed if two acoustically similar talkers do not have the same apparent ethnicity). Socially-mediated perceptual effects also do not necessarily have to be mediated by preconceived social categories – if a phonetically shifted or accented speaker has a pen in their mouth (Liu & Jaeger, 2018), for example, it might trigger generalization towards another individual who similarly has a pen in their mouth.

Another important factor could be the experimental paradigm that is employed. The instructions of the current study mentioned the number and gender of the speakers within each phase, which could have brought more attention to social category membership. Indeed, at the end of the demographic survey following the main task, a handful of listeners in the critical conditions (n = 24 across both experiments) explicitly commented on the /p/ to [b] phonetic shift (e.g., “The female voices in part one sounded like they were saying all of the ‘P’ words with a B instead and I thought that might’ve been intentional”). The phonetic shift in the current study was intended to be clearly noticeable, and the exposure phase task (keyword identification) was designed to draw attention to the phonetic shift. Stating whether stimuli are buy or pie in the test phase patently relates to the exposure phase manipulation, so listeners in the current study may have been influenced by strategic, decision-making biases based on social similarity (Xie et al., 2023). It seems plausible for listeners to take a different approach towards generalization when given a more implicit task with a less obvious phonological contrast (e.g., as in Xie and Myers (2017), who examined devoiced word-final stop consonants in Mandarin-accented speech through auditory lexical decision and cross-modal priming tasks).

Overall, as recent work has noted (Xie et al., 2023), the precise mechanisms of cross-talker generalization are still unclear, and additional work is needed to unpack effects of acoustic versus social similarity.