Glossa Psycholinguistics

1.1 The cognitive architecture of explicit imitation

The term phonetic imitation encompasses a wide range of phenomena, and it is used in studies differing in the target of imitation (e.g., words in isolation vs. natural stretches of speech by an interlocutor), instructions (e.g., with vs. without an explicit directive to imitate), and task (e.g., speech shadowing vs. natural conversation). The current work focuses on explicit imitation of systematic phonetic variation, using a task in which participants are instructed to imitate English sentences characterized by different “accents” that vary minimally from canonical speech in a single phonetic feature: the VOT of voiceless stops.

Faithful imitation requires successful realization of several processes that are each on their own quite complex. As an example, we consider imitation of speech in which the VOTs of voiceless stops are consistently shortened, one of the artificial accents used in the current work. First, the imitator must be able to discriminate the feature(s) characterizing the accent as distinct from canonical speech (i.e., be capable of perceiving the difference between the shortened vs. canonical VOT). They must also be able to articulate the variant (i.e., be capable of producing a voiceless stop with shortened VOT). The imitator must identify which acoustic information is relevant to their imitative goal; if the goal is to imitate an accent, this involves separating the linguistically-relevant feature(s) of the target accent (i.e., the shortened VOT) as distinct from non-linguistically-relevant properties such as those inherent to a speaker’s voice (e.g., a speaker’s raw f0 values). The feature also must be recognized as a systematic property of the accent, rather than specific to a single episode (generalization), and the link between this general feature and the accent must be encoded in the perceptual representation. Finally, the imitator must select the feature in production when imitating the accent.

While this work focuses on explicit imitation of artificial accents, it is also relevant to consider the extent to which the above processes are shared by other types of imitation. Several direct comparisons have found that explicit instructions elicit more imitation than implicit methods. For example, Dufour & Nguyen (2013) showed that participants who were explicitly asked to imitate showed closer approximation to the formant values of a model talker than participants who were simply asked to repeat the words they heard (see also Clopper & Dossey, 2020; Pardo et al., 2010; Sato et al., 2013). To account for these differences, Dufour & Nguyen (2013) proposed that a general automatic alignment mechanism is a shared component of both explicit and implicit imitation, but that explicit instructions to imitate invoke additional “attentional” processes, directing perceptual attention to the specific indexical features of the model talker’s speech. In other words, successful explicit imitation differs from implicit imitation in the extent of attention allocated to the input. We further propose that the explicit imitation may also invoke qualitatively different perception- and production-related processes that are not necessarily required for implicit imitation, as detailed in the following paragraph. We conceptualize explicit imitation as encompassing all processes involved in implicit imitation, but overlaid by additional, “controlled,” perception and production processes.

A prerequisite to any type of imitation is the ability to perceive and produce the relevant variability (discrimination and articulation). Given these capabilities and a model of speech processing including a direct perception-production link (e.g., Goldinger, 1998; Pickering & Garrod, 2013), implicit phonetic convergence can, in principle, occur as an automatic consequence of the interaction between incoming speech and the listeners’ representations (though there is strong evidence that mediating factors play a role, even in implicit convergence; see Coles-Harris 2017 for discussion).¹ On the other hand, explicit imitation also requires additional, controlled processes, whose trajectory cannot be determined automatically based on the combination of input and representation, but which must be guided by other factors. Successful explicit imitation requires not only greater attention to the indexical properties of a talker’s speech (Dufour & Nguyen, 2013), but also qualitative decisions about what to pay attention to, because the choice of which features are relevant (identification) will differ based on the imitative goal. For example, consider an adult speaker of General American English asked to imitate a child who produces systematically shortened VOTs of voiceless stops. The child’s speech will differ in (at least) two ways from the adult’s speech: shorter VOT, and an overall higher pitch due to smaller vocal folds. Which of these differences is imitated depends on the imitative goal: if they aim to imitate the accent of the speaker, the adult might modify their VOT, but if they aim to imitate the age of the speaker, they might instead modify the pitch of their voice. If their goal is to produce the word in a way that was indistinguishable from what they heard, they might modify both. Another decision that must be made by the imitator, in cases where the material to be imitated is not identical to the input, is the scope of generalization. For example, if the imitator in the example above has only heard voiceless stops starting with /p/ (but not /t/ or /k/) in the child’s speech, they must decide whether the shortened VOT is indeed a systematic property of the speaker’s “accent,” and if so, whether this property generalizes to other segments (like /t/ or /k/), or whether it is specific to /p/. Finally, the selection process in production is also controlled by the imitator. For example, the imitator might be reluctant to produce a shortened VOT, even if they are able to articulate it and have recognized it as a property of the accent, perhaps because of social connotations associated with doing so, or perhaps because they are simply uncomfortable diverging from their usual speech norms.

In sum, we conceptualize successful explicit imitation as the combination of automatic processes of phonetic alignment, overlaid with additional controlled processes. Since there is overlap in the processes involved in implicit and explicit imitation (minimally, the ability to discriminate and articulate the relevant variability), findings from the larger body of research on implicit imitation may be relevant to explicit imitation as well, and we therefore consider them in formulating our research questions and hypotheses, and when contextualizing our results within the broader body of work. At the same time, the differences between the two types of imitation, outlined above and supported by the different empirical findings for implicit vs. explicit imitation in past work, highlight the importance of considering the different types of imitation as distinct entities.

1.2 Predictors of variability in imitation

Understanding the sources of variability underlying imitative performance is of both theoretical and practical interest, but it is also a challenge. Because multiple subcomponents are essential to successful imitation, a breakdown at any level will inhibit imitation, so it is often impossible to determine the underlying reason for differences in imitation across groups or individuals simply by looking at the results of an imitation task and how these results correlate with behavioural or neuroanatomical measures. Some previous work has acknowledged this issue: for example, both Nielsen (2011) and Zellou & Brotherton (2021) bring up both perceptual and production-based mechanisms that could underlie the patterns of results found in their work on imitation. Furthermore, Reiterer et al. (2011) found that foreign speech pronunciation aptitude was associated with differences in neural activation in both speech-motor and auditory-perceptual areas.

Perceptual difficulties are often assumed to be the primary cause of lack of imitation (e.g., Olmstead et al., 2013), and theoretical accounts of phonetic convergence include critical roles for perception (e.g., Pickering & Garrod, 2013; Sancier & Fowler, 1997). However, most empirical work on native-language imitation does not directly test the role of perception. One notable exception, Kim and Clayards (2019), examined whether perceptual weighting of spectral and durational cues to the English /ɛ/–/æ/ contrast was related to the imitation of these acoustic dimensions in an explicit imitation task. Results showed that individuals who relied more on spectral cues in perception showed greater imitation of durational, but not spectral differences, while individual reliance on durational cues was not predictive of imitation of either dimension. The authors concluded that general perceptual acuity, rather than attention to specific phonetic dimensions, modulates phonetic imitation. On a group level, Nielsen & Scarborough (2015) found evidence that linguistic selectivity in imitation has a perceptual basis: English listeners were better at discriminating artificially-lengthened than artificially-shortened VOT, providing a perception-based explanation for an earlier finding of asymmetrical imitation (Nielsen, 2011; discussed in detail below). Overall, however, there is no clear or straightforward evidence for the role of individual perceptual patterns in predicting native-language imitation.

If there has been little work directly testing the role of perception in imitation, there has been even less exploring the role of articulation. The one exception we are aware of is work by Reiterer et al. (2013), who, in the context of a neuroimaging study, also took a measure of spectro-temporal acoustic variability, which they termed “articulation space,” in German speakers’ productions of L1 and L2 speech. The participants also completed a separate foreign language imitation task, where they were asked to directly repeat sentences of Hindi, an unfamiliar language, with performance assessed perceptually by native speakers of Hindi. Participants who received higher ratings on the imitation task also had a larger articulation space, and the authors concluded that articulatory flexibility facilitates imitative ability.

Compared to the small number of studies exploring perceptual or articulatory predictors of imitation, there has been considerable interest in examining which characteristics of individual speakers, and their attitudes toward the interlocutor or the target speech, might predict imitation. A multitude of factors, including bilingualism (Spinu et al., 2020), musical experience (Coumel et al., 2019), and general cognitive processes (working memory: Reiterer et al., 2011; neurocognitive flexibility: Reiterer et al., 2013) have been shown to correlate with individuals’ extent of imitation. The list of social or personality-related factors proposed to condition imitation expands even further when considering the better-studied domain of phonetic convergence (see Wade et al., 2020 for a review). While intriguing, findings can be inconsistent and often fail to replicate across studies, making it difficult to draw firm conclusions about the individual characteristics governing imitation (Cohen Priva & Sanker, 2020; Wade, 2022).

One reason that the search for straightforward predictors of imitative ability may be elusive could be the multiple perception- and production-based processes underlying imitation laid out above, each of which is likely in and of itself affected by different factors. For example, the ability to detect small magnitudes of phonetic difference, the propensity to generalize, and the willingness to diverge from habitual speech norms, which might be expected to most strongly affect discrimination, generalization, and selection respectively, will not necessarily pattern together across individuals. These different patterns of variability could therefore obscure any relationship between overall imitation and each of the processes individually.

1.3 Imitation of VOT in English voiceless stops

Imitation of VOT in English voiceless stops has been well-studied, and previous work has consistently found imitation of lengthened VOT (e.g., Nielsen, 2011; Shockley et al., 2004; Wade et al., 2020). However, results on imitation of shortened VOT are mixed. Flege & Hammond (1982) found reduced VOTs in English speakers asked to spontaneously mimic their idea of Spanish-accented speech, indicating that it is possible for English speakers to imitate reduced VOTs. However, imitation of reduced VOT has not been shown to be elicited in any lab-based shadowing/exposure studies. Notably, in a direct test of lengthened vs. shortened VOT imitation, Nielsen (2011) showed that participants produced longer VOTs after exposure to voiceless-stop-initial words where VOTs had been lengthened by 40 ms, but no such effect was found after exposure to analogous stimuli with VOTs shortened.

One possible explanation for the asymmetry, proposed by Nielsen (2011), is that shortened VOT might be less perceptually salient than lengthened VOT. This was supported by a subsequent perception study by Nielsen & Scarborough (2015), who found that listeners were more accurate in discriminating similar lengthened VOT than shortened VOT stimuli from those with natural VOT values. This could also explain the discrepancy between the results of Nielsen (2011) and those of Flege & Hammond (1982), who did find shortened VOT imitation: the participants in Flege & Hammond (1982) may have been basing their imitations on their familiarity with (or stereotypes of) English spoken by L1 Spanish speakers, which may be characterized by unaspirated stops, consistent with the phonetic realization of phonologically voiceless stops in Spanish. These unaspirated stops would have lower VOT values than the shortened VOT stimuli used in Nielsen (2011) (which were on average 30 ms for /p/-initial words, as compared to a range of 7–18 ms for /p/-initial Spanish words reported in Flege & Eefting 1987). They are also likely to be perceived by English listeners as phonologically voiced stops, further increasing their perceptual salience.

In our study, we tested imitation of voiceless stops with shortened, as well as lengthened, VOT using more extreme values than those used in Nielsen (2011): our shortened VOT values were within the normal range of voiced stops (i.e., shortened /t/ was 15 ms, which is characteristic of, and therefore likely to be perceived as, /d/). Therefore, if anything, we might expect increased salience relative to the lengthened VOT condition, which does not ever result in a change of phonological category. In addition, we used an explicit imitation paradigm, in contrast to the implicit pre/post-exposure paradigm in Nielsen’s study, which we expected would direct more attention to phonetic characteristics of target stimuli (Dufour & Nguyen, 2013). If the lack of shortened VOT imitation was driven by limited perceptual salience, these differences in our stimuli and paradigm should result in imitation of shortened VOT.

To get a full view of the nature of imitation, it is important to consider distributional patterns as well as average values, as the shape of the distributions of imitated VOT values provides important information that can be obscured when only considering differences in mean values across conditions. The VOT distributions pre- and post-exposure to lengthened VOT in Nielsen (2011) were strikingly similar in shape, just shifted such that the post-exposure distribution was characterized by slightly higher values overall (mean 7 ms). This suggests that participants’ imitations were best characterized as small but consistent increases in VOT. By contrast, when participants were asked to mimic Spanish-accented English, Flege & Hammond (1982) found that the distribution of VOT values was characterized by two clear clusters around 30 ms and 90 ms. This bimodal pattern suggested that there are distinct types of productions: some tokens were characterized by broad, categorical changes in VOT, while others remained consistent with a canonical English pronunciation. This distribution suggests that unaspirated stops that often characterize Spanish-accented English may have been perceived as a separate phonological category (voiced stops) in some participants and/or utterances, and were produced as such in imitation. Although Flege & Hammond (1982) and Nielsen (2011) differ substantially from one another in their methodologies, the contrast is consistent with a potential qualitative difference between imitation of shortened VOT, which can potentially be perceived as a different phonological category, and lengthened VOT, which cannot. Our study provides a direct comparison of imitation of lengthened vs. shortened VOT; comparison of the distributions across the two conditions allows us to determine if differences observed across previous studies might be ascribed to qualitative, rather than or in addition to quantitative, differences.

1.4 Artificial accents and talker variability

In order to test explicit imitation of different accents (i.e., systematic, linguistically-relevant phonetic characteristics) as well as perception of these differences, we required a paradigm where the target accents were presented in direct juxtaposition. To preview the procedure, which is described in detail in later sections, participants were asked to 1) listen to two talkers with different “accents” saying the same sentence, 2) to repeat after each talker, imitating the accent of each, and 3) to decide whether a third talker best matched the first or second talker they had heard. This direct juxtaposition of contrasting accents, along with the need for phonetically controlled stimuli that would allow us to isolate imitation of a single feature, led us to use artificial accents, and also led us to directly test the effect of voice-related variability on perception and imitation of systematic phonetic differences, as we describe in this subsection.

Different accents in the real world are generally produced by different talkers, so ideally, to make the task plausible, we wanted the two accents in a given trial to be heard as having two different voices. On the other hand, using two different talkers undermines the experimental control that is critical for isolating imitation of a single phonetic feature: speech naturally produced by two different talkers will always have many acoustic differences, particularly when using sentence-level stimuli, so participants might perceive and imitate these differences, instead of or in addition to our target feature of VOT.

To maintain full control over the acoustics of the material presented to participants, while still allowing for the voice-related variability that is present in naturally-occurring accents, we created artificial accents using sentence-level stimuli that differ minimally in specific phonetic features. Artificial accents can be constructed in a variety of ways: by using a speech synthesizer (e.g., Maye et al., 2008), by manipulating natural speech (Yu et al., 2013), or by recording natural productions produced with systematic phonetic/phonological differences (Adank & Janse, 2010; Spinu et al., 2020), and they have been shown to elicit phonetic imitation in past work. For example, using an explicit imitation task, Spinu et al. (2020) found that bilinguals showed more imitation than monolinguals of an artificial accent of English which varied from a canonical accent in four features (/ɛ/ -> /jɛ/, intervocalic /l/ -> /ɾ/, /ə/-epenthesis in s+stop clusters, and a novel intonation pattern in tag questions), and Yu et al. (2013) showed that participants produced longer VOTs after exposure to a narrative with systematically lengthened VOT of voiceless stops.

In order to directly compare imitation of multiple artificial accents, we manipulated the VOT of voiceless stops from a single set of baseline recordings to create three “accents” varying only in VOT: one set of stimuli had canonical VOT, approximating the talker’s natural production values, while the other two sets were manipulated to have either systematically shortened or lengthened VOT. Apart from the VOT, the three “accents” were acoustically identical. We then created multiple “voices” for each accent by scaling the f0 and formants. Crucially, because only overall f0 and formant scaling were modified, there were no linguistically-relevant differences between the different voices. In sum, we created a bank of stimuli representing three different accents, with multiple voices per accent, while maintaining complete control over all other acoustic characteristics of the sentences.

The inclusion of multiple voices, in contrast to previous work looking at imitation of artificial accents, brought up the question of how the presence of voice-related variability might affect attention to – and therefore imitation and discrimination of – the target VOT differences. In order to determine what effect this might have, we compared performance in a condition characterized by multiple voices, as described in the previous paragraph, to another condition which included no voice-related variability at all: participants imitated and discriminated sentences that differed only in VOT: in other words, with different artificial accents (e.g., shortened vs. canonical VOT), but with the same “voice.”

Findings from several strands of work make different predictions about how performance might differ across conditions where voice-related variability is present vs. absent. First, there is evidence that acoustic variability introduced by the inclusion of multiple talkers invokes a general phonetic processing cost, perhaps attributable to the additional processing resources needed to separate talker-specific from phonetically-relevant information (e.g., Mullennix & Pisoni, 1990; Mullennix et al., 1989). Support for this idea comes from findings that talker variability has been shown to hinder performance on some perceptual learning tasks, particularly those not requiring generalization (see Baese-Berk, 2018, for discussion and examples). Under this view, which assumes that phonetic detail is retained during processing (Goldinger, 1998), we would expect that decreased sensitivity or attention to the target VOT differences in discrimination and imitation would lead to worse performance with multiple voices, as compared to the condition with no voice-related variability.

In the studies discussed above that found decreased performance in the presence of multiple talkers, the stimuli were indeed produced by different human talkers, and therefore exhibited natural talker-related variability. In our study, on the other hand, the different voices were fully controlled except for scaling of f0 and formants, characteristics not relevant to phonetic perception in English. Based on findings that natural talker variability hindered performance on spoken word identification, but minimal differences in amplitude or overall f0 did not (Bradlow et al., 1999; Sommers et al., 1994; Sommers and Barcroft, 2006), Sommers and Barcroft (2006) proposed the Phonetic Relevance Hypothesis: variability that alters acoustic properties that are important for phonetic identification imposes processing costs on perception, but non-phonetically-relevant variability does not. Under this view, the presence of our highly controlled different “voices”, specifically constructed not to vary in any phonetically-relevant dimension, would not be expected to affect imitation or perception of VOT differences. In this case, we would expect to see identical performance in the conditions with and without voice-related variability.

Finally, there is reason to expect that voice-related variability might actually improve performance on tasks where generalization is required, as is the case in some portions of the current study. Talker variability has been shown to facilitate generalization in both dialect classification (Clopper & Pisoni, 2004) and accent adaptation (Bradlow & Bent, 2008; Schmale et al., 2012). These studies are not directly analogous to the tasks in the current work, since these previous studies compare performance when there are multiple talkers per accent vs. a single talker per accent, whereas the current work compares multiple talkers per accent to a complete lack of talker variability (i.e., the same voice is used in both accents). However, the fundamental prediction from these previous studies still applies: variability inherent in a multi-talker context may provide information about which dimensions are linguistically relevant (and therefore helpful in informing perceptual judgments in different contexts), and which can be attributed to idiosyncratic or indexical properties of a single talker (and therefore only relevant to that talker). We included a combination of exposed and novel sentences in the discrimination phase of our experiment, allowing us to test whether the effect of voice-related variability would facilitate generalization.

1.5 Current study

This work examines explicit imitation of artificial accents varying in a single feature (VOT), using a preregistered design and analysis.² Our paradigm consists of four tasks: artificial accent imitation, artificial accent discrimination, word-level VOT imitation, and word-level VOT discrimination. All tasks involved imitation and discrimination of VOT differences; however, the accent tasks consisted of imitation of sentences, as opposed to individual words. The sentence-level accent tasks differed from the word-level tasks in that they were designed to invoke the goal of imitating an “accent” rather than idiosyncratic properties of a specific token of a word. We did this by including variability in sentences and talkers (such that the tasks could not be done without some level of generalization), and by explicitly invoking the concept of “accent” in instructions to the participants.

The two accent-level tasks, artificial accent imitation and discrimination, were combined into a single task presented to participants. In each of a series of trial sets, participants heard and imitated a pair of sentences with different artificial accents, then completed an ABX task, deciding whether a third sentence matched the first or second accent they had heard. Stimuli for the accent-level tasks were manipulated in three fully crossed conditions (Table 1). First, the AccentType manipulation varied in the nature of the artificial accent: one condition tested imitation and discrimination of lengthened vs. canonical VOT, while the other tested shortened vs. canonical VOT. Second, to test the role of voice-related variability, the TalkerMatch manipulation varied in whether the sentences within each trial were spoken with the same voice or different voices. Third, the ABX discrimination task included two types of X sentences: half were the same as the A and B sentences in terms of linguistic content, while the other half were different sentences (SentenceMatch). This factor was included to test generalization to new words and segments, and to reinforce the instructions that the task should be done by considering differences as a general feature of an “accent” (i.e., a systematic property of the way a person talks), as opposed to an idiosyncratic property of individual words or utterances.

The other two tasks tested word-level VOT imitation and discrimination. Stimuli for these tasks were tokens of a single word, spoken in a single voice, manipulated to differ only in VOT. In contrast to the accent-level tasks, word-level imitation and discrimination were presented as two separate tasks: for imitation, participants heard and were asked to listen to and imitate tokens drawn from a VOT continuum, and for discrimination, participants completed an ABX task with tokens from the same continuum.

The four tasks included in the study were chosen to tap into different sub-processes involved in explicit imitation of artificial accents. Table 2 shows the mapping between the tasks and the five processes laid out above in Section 1.1. The ability to discriminate VOT differences is necessary for successful completion of all tasks. This is in fact the only prerequisite to completion of the word-level ABX discrimination task, which can be done by shallow acoustic matching of tokens of a single word. The accent discrimination task, on the other hand, additionally requires identification of VOT as a relevant feature, while at the same time identifying other variability, such as f0 of the talker’s voice, as not a relevant feature, when assessing similarity. It also requires generalization of the VOT difference to different segments and words in order to accurately discriminate a sentence with different linguistic content (in the DifferentSentence condition).

Turning to imitation, both tasks require the ability to discriminate the differences, and the ability and decision to produce them (articulation and selection). The question of whether identification is involved in the imitation tasks is less straightforward. To the extent that fully automatic alignment processes operate in all imitation tasks, VOT imitation would be expected to occur even without any active identification of VOT as a relevant dimension. However, following the view of Dufour and Nguyen (2013), we expect that explicit instructions to imitate direct additional attention to the properties of the target stimuli and encourage identification of the dimension that is relevant to the target contrast. Furthermore, although the relevant dimension (VOT) is the same in both the word-level and accent imitation tasks, it is important to note that the two tasks differ in imitative goal (imitate differences in the word vs. differences in two accents), in the nature of the stimuli (sentence- vs word-level), and in the fact that the accent task includes additional voice-related variability (in one of the two TalkerMatch conditions); any of these differences might affect the ease of identification of this dimension. Neither imitation task tests generalization to new material.

The aim of the current study is to test the role of several factors expected to influence explicit imitation of artificial accents varying in VOT. First, we explore the extent to which individual performance in artificial accent imitation is correlated with three tasks (accent discrimination, word-level imitation, and word-level discrimination) designed to target different sub-processes of imitation. For those tasks shown to be significant predictors, we also test to what extent they themselves are related. These comparisons provide information about which factors might be the most important predictors of variability in artificial accent imitation. For example, if overall perceptual acuity is the primary predictor of imitative ability, we would expect variability in the accent imitation task to be correlated with all three other tasks (and that all tasks would in fact be correlated), since all require discrimination of VOT differences. On the other hand, if production-based factors (articulation and/or selection) are primary, we would expect to find that the word-level imitation task is predictive of the accent imitation task, and that this is independent of performance on the other tasks. If the ability to recognize and/or generalize VOT as a property of an artificial accent, even in the face of talker variability, is primary, then we would expect to find that the accent discrimination task is predictive, independent of other tasks. Second, we test the effect of voice-related variability on imitation and discrimination of systematic VOT differences by comparing performance in the SameVoice vs. DifferentVoice conditions of the accent-level tasks. We expect it to be more difficult to perceive – and therefore imitate – differences in VOT in the DifferentVoice condition, due to the additional acoustic variability introduced by the different voices. However, since the acoustic variability was controlled to only differ in non-phonetically-relevant characteristics (overall f0 and formant scaling), this voice-related variability may be automatically filtered out (Sommers & Barcroft, 2006), in which case we would expect no difference in the two conditions. It is also possible that the variability of voices in the DifferentVoice condition facilitates generalization in classifying novel items, and we therefore expect that any inhibitory effect of the DifferentVoice condition on the perception task might be smaller – or reversed – when test items are novel sentences. Finally, we compare explicit imitation of shortened vs. lengthened VOT, using more extreme values than used in previous lab-based studies. We also compare distributional patterns across imitations of the two VOT manipulations.

3.1 Accent tasks: Catch trials

Practice/catch trials (presented at the beginning of the experiment and between blocks of the accent tasks) were included to ensure that participants understood the task and were imitating the accent, as opposed to the voice of the talker. These trials consisted of a sentence containing two coda rhotics, with one accent spoken with the rhotics present and the others with the rhotics absent. Participants completed 4 imitation and 16 discrimination trials (11 of the 168 total imitation trials were missing due to recording errors). Imitation trials were coded for presence/absence of a coda rhotic for each target word, based on the perception of a phonetically trained research assistant. We report the accuracy (for discrimination) and percentage of target imitations in which the presence/absence of rhotic matched the stimulus, both in total across the full experiment, and broken down by Session/TalkerMatch condition.

Participants had very high discrimination accuracy for these catch trials across all Session/TalkerMatch conditions (Session 1 SameVoice: 98%; Session 1 DifferentVoice: 98%; Session 2 SameVoice: 100%; Session 2 DifferentVoice: 98%). Imitation of target sounds in the catch trials was also highly accurate. Target words with rhotics present (i.e., the canonical North American English pronunciation) were imitated with rhotics 97% of the time (Session 1 SameVoice: 97%; Session 1 DifferentVoice: 97%; Session 2 SameVoice: 95%; Session 2 DifferentVoice: 98%). Target words with rhotics absent (i.e., a noncanonical pronunciation for North American English speakers) were imitated without rhotics 85% of the time (Session 1 SameVoice: 79%; Session 1 DifferentVoice: 79%; Session 2 SameVoice: 92%; Session 2 DifferentVoice: 93%).

3.2 Artificial accent imitation results

VOT values for imitations of the target words in the artificial accent imitation task, as well as the VOT values of the stimuli themselves, are shown in Figures 2 and 3, and statistical results are shown in Table 3. For purposes of analysis, we calculated a by-trial measure of VOT-imitation that captures the difference in VOT for each target segment across the two accents in each imitation trial, transformed such that positive values indicate a change in the expected direction of imitation. For trials in the lengthened VOT condition, this was [VOT during imitation of lengthened accent] – [VOT during imitation of canonical accent], and vice versa for the shortened VOT condition. We used a mixed-effects linear regression model to test whether there was imitation, as well as whether the extent of VOT-imitation differed based on AccentType (lengthened vs. shortened VOT), TalkerMatch (same or different voices), and Session (1 vs. 2). The three predictor variables (reference levels in italics above) were included, as well as all interactions, mirroring the structure of the perception model (excluding the factor SentenceMatch, which was not relevant for the production data because the imitated sentences in a given trial were always the same). Random intercepts were included for Participant and Sentence ID, as well as random by-participants slopes for TalkerMatch and AccentType, and by-item slopes for TalkerMatch.⁷ All factors were centered (–0.5, 0.5). In interpreting the model, the estimate of the intercept represents the extent of VOT imitation (in ms) across all conditions, and the estimate corresponding to each fixed factor represents the difference in VOT imitation between the two levels of the factor. Likelihood ratio tests showed that including age as an additional predictor did not significantly improve model fit from the model without age, either when it was included as both a main effect and in interaction with Session and TalkerMatch (χ² = 10.26, p = 0.248), or when it was included as only a main effect with no interactions (χ² = 0.23, p = 0.635).

Figure 2: Raw VOT values produced in imitation of sentences with canonical vs. lengthened VOT (left panel) or canonical vs. shortened VOT (right panel), broken down by place of articulation. Large asterisks indicate the VOT values in the stimuli. Error bars show 95% confidence intervals of by-participant means.

Figure 3: Accent imitation task: VOT imitation (difference in VOT between two accents on a given trial, with positive values indicating a difference in the expected direction) for lengthened (left) and shortened (right) VOT conditions. Participants who heard the SameVoice condition in Session 1 and DifferentVoice in Session 2 are shown in the left panel of each graph, and the other half of the participants are shown in the right panel. Error bars show 95% confidence intervals of by-participant means.

The significant (positive) intercept indicates that participants imitated the VOT differences between the two accents, showing on average an 18 ms VOT difference, in the expected direction, between the two productions imitating the two different accents in a given trial. Main effects indicated more imitation in Session 2 than in Session 1, but no overall difference between the two TalkerMatch conditions or between the two AccentType conditions. However, there was a two-way interaction between Session and TalkerMatch, with follow-up tests indicating that there was a significant effect of TalkerMatch during Session 1, with more imitation in the SameVoice than the DifferentVoice condition, and no effect during Session 2 (rather, there was a trend in the opposite direction, with DifferentVoice being numerically higher than SameVoice) (Session 1: χ² = 6.48, p = 0.011; Session 2: χ² = 3.41, p = 0.065). There was also a significant three-way interaction of Session, TalkerMatch, and AccentType: follow-up tests of simple effects of AccentType at each combination of Session and TalkerMatch indicate that there was an effect of AccentType, with greater imitation for lengthened than shortened VOT in the Session 1 SameVoice condition (χ² = 3.90, p = 0.048), and the Session2 DifferentVoice condition (χ² = 7.19, p = 0.007), but no significant effects in the other two Session * TalkerMatch conditions (both p > 0.1). Finally, we performed follow-ups to test whether participants showed significant imitation in all Session * TalkerMatch conditions. The imitation effect was not significant in the least imitative, Session1-DifferentVoice condition, but was significant in the other three conditions (Session1-DifferentVoice: χ² = 2.49, p = 0.114, all others p < .001).

We also examined the shape of the distributions of imitated VOTs. Density plots of the distribution of the full set of participants’ VOT values are shown in Figure 4, broken down by Session and TalkerMatch condition. Overall, distributions of shortened VOT are shifted lower, and lengthened VOT are shifted higher, than the canonical distributions, although this is much less apparent in the Session 1 DifferentVoice condition than the other three Session/Voice combinations, reiterating the result shown above that imitation was inhibited in the DifferentVoice condition for those participants who completed it during the first session. However, observation of the distributions of the conditions where imitation was found (i.e., all except for the Session 1 DifferentVoice condition) reveals differences that are masked by the group means shown above. Specifically, the distributional patterns are different in the shortened condition, where there is clear evidence of bimodality, than in the lengthened condition, where there is not. This is reminiscent of differences found in previous work: recall that the distributions of post-exposure VOT lengthening in Nielsen (2011) indicated a small-but-consistent modification (i.e., a shifted distribution of the same shape), whereas distributions of shortened VOT in spontaneous mimicry of Spanish-accented English in Flege and Hammond (1982) indicated more categorical behavior (i.e., a bimodal distribution).

Figure 4: Density plots showing the distribution of raw VOTs in imitations of the canonical (grey) and manipulated accents (lengthened: green fill and dotted line; shortened: blue fill and dashed line), broken down by Session and TalkerMatch condition.

In order to look at the distributional patterns of imitation in more detail, we examine the results broken down by item, shown in Figure 5. For these item-based plots, we only include data from Session 2: we exclude Session 1 data because we are interested here in examining distributional patterns when there is imitation, and imitation appeared to be categorically inhibited in the Session 1 DifferentVoice condition (there was imitation in Session 1 for half of the participants, those who heard the SameVoice condition first; however, including this would mean that this half of the participants would be over-represented in the distributional analysis). In Figure 5, imitations of the manipulated (lengthened or shortened) VOT are overlaid on imitations of the canonical VOT, broken down by token. First, we consider the distributions of imitations of shortened VOTs, shown in the top panels. There is clear evidence of bimodality in most of the /p/ and /t/ tokens, with the lower peak indicating that some participants (by-participant distributions are shown in Appendix D) imitated these as unaspirated stops, presumably because they perceived them as voiced rather than voiceless stops. In contrast, there is no evidence of bimodality in imitations of shortened VOT for tokens beginning with /k/, suggesting that these tokens were consistently heard as voiceless stops, despite their low VOT.

Figure 5: Density plots showing the distribution of VOTs in imitations of the canonical (grey) and manipulated (blue for /p/, green for /t/, purple for /k/) accent in each trial, broken down by token, in the shortened (top) and lengthened (bottom) conditions (Session 2 data only). Dashed lines indicate the model’s VOT values for canonical (grey) and manipulated (black) conditions.

Turning to the lengthened imitations, recall that we expected a reflection of the findings of Nielsen (2011): a rightward-shifted version of the canonical imitations. Our data instead show distributions with slightly offset peaks, but notably, with a substantial rightward skew. In contrast to the shortened condition, the shape of the distribution for lengthened imitations is fairly consistent across all words (Figure 5). Observation of by-participant results indicates that this shape arises from substantial heterogeneity both within and across participants: some participants showed small-but-consistent modifications, some showed larger, consistent modifications, and some showed more variable, flatter distributions. Furthermore, as can be seen in Figure 5, some productions were hyper-lengthened, i.e., longer than the value of the model talker.

Observation of the distributions of shortened and lengthened VOT imitation calls into question the validity of the comparison of imitation across the two accent types: while the statistical analysis above suggested a similar degree of imitation, the nature of this imitation is quite different. The bimodality in the shortened condition also brings up the question of whether shortened VOT imitation effect might be solely attributable to a categorically different phonological target. To examine this possibility, and to provide a comparison of shortened vs. lengthened VOT imitation that eliminates this difference, we performed a post-hoc analysis on the subset of words beginning with /k/ (cats, coffee, and kites) in Session 2 (504 trials): these words showed no evidence of bimodality, so any imitation found in this subset is unlikely to be driven by perception of a categorically different segment.

Our post-hoc analysis was designed to test two questions: first, whether we still find imitation of shortened VOT after excluding tokens that were sometimes perceived as voiced, and second, whether the extent of imitation differs between shortened and lengthened VOT in this subset of stimuli. We used a linear mixed-effects model with the same structure as above (excluding the factor of Session) to test how the extent of imitation of /k/ words was predicted by AccentType, TalkerMatch, and their interaction, including random intercepts for Participant and Item, as well as random by-participants slopes for TalkerMatch and AccentType. A significant effect of AccentType showed that there was more imitation of lengthened than shortened VOTs (30 ms for lengthened vs. 8 ms for shortened; β = –22.19, SE = 5.23, t = –4.244, p = 0.003), and a follow-up test indicated that there was indeed significant imitation for the shortened VOT stops (χ² = 7.84, p = 0.003). Effects of TalkerMatch and the AccentType * TalkerMatch interactions were not significant (p > 0.1). We further note that the shape of the distributions for these shortened VOT imitations (Figure 5) shows less variability, and does not show the same type of skew, as the distributions of their lengthened VOT counterparts; rather, they are symmetrical and reflective of the shifted distributions that were found for lengthened VOT in Nielsen (2011). Implications of this will be discussed below.

To summarize the VOT results from the accent imitation task, participants showed imitation of accents differing in lengthened and shortened VOT, but the extent of imitation was not significant when participants heard different voices in the first session. Taking into account data from all words, the mean imitation values between the normal and modified VOT were 6 ms for the least accurate, Session1-DifferentVoice condition, compared with 21 ms in Session1-SameVoice, 28 ms in Session2-DifferentVoice, and 18 ms in Session2-SameVoice. There were no overall effects of AccentType, but in one condition (Session2-DifferentVoice), there was greater imitation for lengthened VOT than for shortened VOT. Although the overall extent of imitation was similar across the two AccentTypes, the shape of the distributions differed. A clear bimodal distribution for shortened VOT imitations was attributable to categorical perception and production of a different phonological category (voiced stops) for some tokens. However, the overall shortened VOT imitation effect cannot be fully attributable to this, as a post-hoc test including only tokens with no evidence of voiced perception (i.e., /k/-initial words) found significant imitation of shortened VOT (despite the lower power due to a smaller sample size), although the extent of imitation was of smaller magnitude than was found for lengthened VOT.

3.2.1 Vowel duration differences

Based on previous work (Nielsen, 2011), we expected that speakers might interpret – and imitate – VOT differences as overall duration/rate differences, and therefore wanted to test whether speakers also produced differences in vowel duration. For each production, we calculated the difference in vowel duration between the production imitating the modified-VOT (lengthened/shortened) sentence and the canonical-VOT sentence. We then tested whether this difference was significantly different than zero (i.e., if speakers showed modification of vowel duration), and whether the extent of difference varied across conditions. We tested this in a parallel way to the VOT difference analysis above, using a mixed-effects linear regression model with the same predictor variables (Session, TalkerMatch, and AccentType) and random effects structure, with estimates representing the average vowel duration difference corresponding to that factor. Figure 6 shows the average vowel duration differences across conditions, and statistical results are given in Table 4. The effect of AccentType is clear in the graph, with speakers showing consistent differences in the lengthened, but not the shortened condition. When imitating lengthened VOT, speakers also increased their vowel duration, while when imitating shortened VOT, there was not a clear pattern of change in vowel duration.

Figure 6: Accent imitation task: By-trial differences in vowel duration between productions imitating modified (lengthened/shortened VOT) vs. canonical values, across all conditions. Positive values indicate that the vowel duration following the modified sentence was longer than that following the canonical VOT sentence. Participants who heard the SameVoice condition in Session 1 and the DifferentVoice condition in Session 2 are shown in the left panel of each graph, and the other half of the participants are shown in the right panel. Error bars show 95% confidence intervals of by-participant means.

Table 4: Statistical results from a linear mixed-effects model predicting vowel duration differences (in ms) in the accent imitation task. Reference levels are in italics, and significant effects are shaded.

Factor	Estimate	Std. Error	t value	p
(Intercept)	6.53	1.599	4.085	< 0.001
Session (1 vs. 2)	0.216	1.275	0.170	0.866
TalkerMatch (same vs. different)	–3.853	1.275	–3.022	0.004
AccentType (lengthened vs. short)	–13.524	2.317	–5.838	< 0.001
Session * TalkerMatch	10.345	5.618	1.842	0.073
Session * AccentType	0.316	2.342	0.135	0.892
TalkerMatch * AccentType	–2.746	2.342	–1.173	0.241
Session * TalkerMatch * AccentType	–5.834	8.087	–0.721	0.475

In the statistical results, the significant (positive) intercept indicates that participants produced overall longer vowel durations for the modified compared to the canonical values, but there were main effects for TalkerMatch (with greater differences for Same vs. DifferentVoice) and AccentType (with greater differences for lengthened vs. shortened AccentType), indicating that this difference may not hold across all conditions. Confirming the patterns seen in the graph, follow-up tests looking at the effect in the two AccentTypes separately show that the vowel duration difference is significantly greater from zero for the lengthened, but not for the shortened, condition (lengthened: χ² = 45.67, p < 0.001; shortened: χ² = 0.01, p = 0.908). Following up on the effect of TalkerMatch, we found that the vowel duration difference was significantly different from zero for both. There was also a trending two-way interaction between Session and TalkerMatch, which, when followed up, mirrored patterns from the VOT analysis above: the vowel duration difference was not significant in the Session1-DifferentVoice condition, but it was in the other three (Session1-DifferentVoice: χ² = 0.66, p = 0.416, all others p < .05).

To summarize, although the only durational difference in the exposure stimuli was VOT, our participants also produced longer vowel durations when imitating lengthened VOT (by 14 ms on average). They did not modify their vowel durations when imitating the difference between shortened and canonical VOT (0 ms difference on average). The vowel duration modification was smaller in the DifferentVoice condition, particularly in Session 1. Increased vowel duration for the lengthened-VOT condition was expected: it has been found in other work (Wade et al., 2020), and it can be plausibly attributed to the idea that participants interpret the VOT differences at least in part as an overall durational difference, and therefore increase global duration in their imitation. However, speech rate differences alone cannot explain the overall findings of VOT imitation: as in Nielsen (2011), there was proportionally more VOT lengthening than vowel duration lengthening. In addition, there was no decrease in vowel duration for the shortened VOT condition.

3.3 Artificial accent discrimination results

Listeners’ discrimination accuracies in the accent discrimination task are shown in Figures 7 and 8, and statistical results are in Table 5. We used a mixed-effects logistic regression model to evaluate the effect of our test variables on listeners’ responses. The binary response variable was listeners’ accuracy in each trial, and the fixed predictors, with reference levels in italics, were Session (1 vs. 2), TalkerMatch (same talker vs. different talker), SentenceMatch (same sentence vs. different sentence), and AccentType (lengthened vs. shortened), as well as the three-way interactions between Session * TalkerMatch * AccentType and between Session * TalkerMatch * SentenceMatch (we did not include interactions involving AccentType and SentenceMatch because we did not have reason to expect that the extent of generalization would differ based on the AccentType). Random intercepts were included for Participant and Sentence ID (i.e., Item), as well as random by-participant slopes for TalkerMatch and AccentType, and by-item slopes for TalkerMatch. All factors were centered (—0.5, 0.5). In interpreting the model, the estimate of the intercept represents the log odds of an accurate response across all conditions, and the estimate corresponding to each fixed factor represents the difference in log odds between the two levels of the factor. Likelihood ratio tests showed that including age as an additional predictor did not significantly improve model fit from the model without age, either when it was included as a main effect and in interaction with Session and TalkerMatch (χ² = 2.03, p = 0.731), or when it was included as only a main effect involved in no interactions (χ² = 0.12, p = 0.728).

Figure 7: Accent discrimination task: Percentage accuracy by Session and TalkerMatch. Error bars show 95% confidence intervals of by-participant means.

Figure 8: Accent discrimination task: Percentage accuracy broken down by all variables. Error bars show 95% confidence intervals of by-participant means.

The significant (positive) intercept indicates that listeners were above chance overall, and that performance differed overall by Session and TalkerMatch, as expected, with listeners performing better in Session 2 than Session 1, and better in the SameVoice condition than in the DifferentVoice condition. However, there was a significant interaction between these two variables, and follow-up tests show that the effect of TalkerMatch was only significant in Session 1 (Session 1: χ² = 14.66, p < .001; Session 2: χ² = 0.20, p = 0.654). AccentType and SentenceMatch did not show significant main effects, but there were significant three-way interactions between each of these two variables and Session * TalkerMatch. We performed follow-up tests to determine whether there were significant effects of AccentType and/or SentenceMatch in different Session/TalkerMatch conditions. For AccentType, there was a significant effect in the SameVoice condition in Session 2 (χ² = 7.33, p = 0.007), with the shortened accent condition eliciting higher accuracy than the lengthened condition, but no significant effects in the other three Session * TalkerMatch combinations (all p > 0.1). For SentenceMatch, there were no significant simple effects in any of the four Session * TalkerMatch conditions, although there was a trending effect for higher performance for previously exposed sentences (SameSentence) than novel sentences in the SameVoice condition, in both sessions (Session 1 SameVoice: χ² = 3.85, p = 0.050; Session 2 SameVoice: χ² = 2.90, p = 0.088; Session 1 and Session 2 DifferentVoice conditions, both p > 0.1). We also confirmed via follow-up tests that the effect of TalkerMatch was significant in Session 1, but not Session 2, when broken down by both levels of AccentType and TalkerMatch, respectively. Finally, we performed follow-ups to test whether listeners were above chance in all Session * TalkerMatch conditions (i.e., even in the least accurate Session1-DifferentVoice condition), and they were (Session 1 Different talker: χ² = 6.10, p = 0.014, all others p < .001).

To summarize, listeners were able to classify accents differing in lengthened and shortened VOT at above-chance levels, with performance substantially hindered (albeit still above chance) when presented with different voices in the first session. Mean accuracy rates were 60% in that condition, compared with 76% in Session1-SameVoice, 76% in Session2-DifferentVoice, and 74% in Session2-SameVoice. There was no consistent pattern found for AccentType, but in one condition (Session2-SameVoice), listeners were more accurate on the shortened than the lengthened condition. Listeners generalized to new sentences, with no significant difference between new and familiar sentences (although there was a trending effect for worse performance on new sentences in the SameVoice condition). This ability to generalize to new sentences supports the idea that participants were not simply performing shallow acoustic matching when doing this task, but were basing their responses on their awareness of systematic differences in the phonetic realization of voiceless stops as a category.

3.4 Word-level discrimination

The word-level discrimination task was an ABX discrimination task on pairs of tokens of the word toast that differed by either one or two steps along a five-step VOT continuum (15, 55, 95, 135, and 175 ms). We expected that ease of discrimination would differ based on the acoustic distance between the two (1 or 2 steps), as well as based on whether or not the pair straddles a phonological category boundary (with higher accuracy expected in between-category pairs). As we expected the category boundary to fall between 15 and 55 ms (consistent with previous work on English alveolar stop perception, e.g., 35 ms in Kuhl and Miller, 1978, and approximately 25 ms in Benkí, 2001), we created a factor Shared Category indicating whether the target pair of stimuli was between-category (all pairs including a token with 15 ms VOT), or within-category (all other pairs, since all other tokens had 55 ms VOT or greater). We analyzed the results with a mixed-effects logistic regression model predicting accuracy from the two predictor variables Step Distance and Shared Category, both simple-coded as (–0.5, –.5), as well as their interaction. A random by-subjects intercept and random by-subjects slopes for Step Difference and Shared Category (uncorrelated with the random intercepts) were included. Discrimination accuracy is shown in Figure 9, and statistical results are shown in Table 6.

Figure 9: Word-level discrimination task: Percentage accuracy on an ABX discrimination task, broken down by VOT step distance (1 step = 40 ms; 2 steps = 80 ms) and minimum VOT of the pair. Error bars show 95% confidence intervals of by-participant means.

The significant (positive) intercept indicates that listeners were above chance overall. The effect of Step Difference was not significant, although it trended in the expected direction, with listeners showing numerically higher accuracy for the two-step than for the one-step difference. As expected, listeners showed higher accuracy for pairs that likely straddled a category boundary (i.e., 15 vs. 55 ms, or 15 vs. 95 ms) than those that were likely perceived as the same category (73% vs. 59% accuracy on average). A follow-up test showed that despite having lower accuracy, the within-category pairs were still discriminated significantly above chance (χ² = 15.37, p < 0.001). There was no interaction between Step Difference and Shared Category. Likelihood ratio tests showed that including age as an additional predictor did not significantly improve model fit (χ² = 0.02, p = 0.876).

3.5 Word-level imitation

We tested whether participants modified their VOT and vowel duration when asked to imitate tokens drawn from the same five-step VOT continuum described above. Mean values for VOT and vowel duration as a function of continuum step are shown in Figure 10, and statistical results are shown in Table 7. Statistical analysis was done using two mixed-effects linear regression models, with the response variables being VOT in one model and vowel duration in the other, with a predictor variable of VOT step (scaled to z-scores and treated as a continuous variable) in both cases. Random by-participant intercepts and slopes for VOT step were included in both models. In both models, the effect of VOT step was significant, indicating that participants increased both their VOT and their vowel duration as the VOT of the stimulus increased. To give a concrete example of the effect size, the mean VOT values for imitations at the endpoints of the continuum were 62 ms (following the shortest, 15 ms step) to 142 ms (following the longest, 175 ms step), while mean vowel durations at these endpoints were 224 and 266 ms. As can be seen in the figures, while there were some participants who showed much greater effects of VOT step than others, most participants showed an overall increase. Likelihood ratio tests showed that including age as an additional predictor did not significantly improve model fit (χ² = 1.91, p = 0.385).

Figure 10: Word-level imitation task: VOT (left) and vowel duration (right) of productions following stimuli from a VOT continuum. Dots show raw values; lines show best-fit regression lines by participant.

3.6 Relationship between tasks

To compare performance across tasks on an individual level, we calculated by-participant indices for each of the four tasks: the accent imitation and discrimination tasks (completed over two sessions) and the word-level imitation and discrimination tasks (completed at the end of Session 2 by all participants). Recall that in the accent imitation and discrimination tasks, there were large discrepancies in performance between participants in the different TalkerMatch conditions during Session 1: both discrimination accuracy and extent of imitation were much lower in the DifferentVoice condition than the SameVoice condition. On the other hand, during Session 2, performance was largely similar for all participants, regardless of whether they were listening to the same or different voices. This discrepancy presented a challenge in our task of calculating individual indices of imitative ability. The most straightforward approach, calculating an index based on the full dataset from each participant, would result in systematically lower values for participants who had heard the DifferentVoice condition first, but this would not reflect an actual difference in imitative ability. Therefore, we used only Session 2 data to calculate individual indices for the accent tasks, which we take to be a better indicator of imitative ability. It should therefore be kept in mind that half of the participants’ data is from the SameVoice and the other half is from the DifferentVoice condition; however, as we did not see systematic differences between these conditions in Session 2, we think that this is a reasonable indicator of their overall imitative and discriminative ability.

Indices were calculated as follows. For the perception tasks (both accent and word-level), indices were taken to be each participant’s overall accuracy across target trials. For the accent imitation task, the index was calculated as the mean VOT imitation using the same metric as for the group-level analysis (by-trial VOT difference between the two accents, transformed such that positive values correspond to changes in the expected direction). For the word-level imitation task, the index was the slope of the best-fit regression line predicting an individual’s VOT from the stimulus VOT step (i.e., the slope of one of the lines shown above in Figure 10). Our primary analysis tested how artificial accent imitation was predicted by each of the other three tasks, and we further tested the relationship between 1) word-level discrimination and word-level imitation and 2) accent discrimination and word-level imitation, as discussed below.

Scatterplots of the comparisons of interest are shown in Figure 11, and statistical results are shown in Table 8. Pearson’s product-moment correlations were computed for each of these using the cor.test function in R. All correlations between accent imitation and each of the three other tasks were positive; those with the accent discrimination and word-level imitation tasks were significant, while that with the word-level discrimination task was not. The two significant correlations provide evidence that success in artificial accent imitation was predicted both by subjects’ ability to perceive the difference in the different accents (as evidenced by the significant correlation with the accent discrimination task) and their low-level VOT imitation ability (as evidenced by the significant correlation with the word-level imitation task). It is possible that these are not independent predictors: perhaps talented imitators are high performers in both of these tasks. In this case, we should also see a significant correlation between individual performance on the word-level imitation and the accent discrimination tasks. However, this correlation was not significant. As with any null result, it is possible that the effect was simply not strong enough to be detected with an experiment of our sample size. Based on a post-hoc power analysis, our design had sufficient power to detect correlations of r = 0.45 or higher with the tests we used (i.e, ‘moderate’ or stronger, according to Evans 1996), and therefore may not have been sensitive enough to pick up weaker effects.⁸ Still, if there is an effect, it is likely very weak, indicating a small amount of shared variance, and we therefore take this as evidence of independence between these two predictors of accent imitation. Finally, we tested the relationship between word-level imitation and discrimination, and found no significant correlation.

Figure 11: Relationship between individual participants’ performance on the tasks: Accent imitation vs. accent discrimination, word imitation, and word discrimination (top), word-level imitation vs. accent discrimination and word-level discrimination (bottom right). Error bars show 95% confidence intervals surrounding the best-fit regression line. Dotted lines show chance performance (for discrimination tasks) or no imitation (for imitation tasks). Correlation coefficients and p-values are shown; full statistical results are provided in Table 8.

The analyses done above were based on imitation data from all words. However, recall that shortened-VOT imitations appeared to be characterized by categorical segmental substitutions in many words. When considering only the subset of data where this issue did not arise (i.e., /k/-initial words used in the post-hoc analysis for the accent imitation task), we found different imitation behavior for shortened vs. lengthened VOT stimuli, calling into question the validity of an aggregate individual imitation index. We therefore calculated a by-participant imitation index in the same way as above for shortened and lengthened VOT separately, considering the /k/-word subset. Using these indices, we first ran a correlation test to determine participants’ consistency in their extent of imitation across the two accent types; in other words, whether participants with the greatest lengthened VOT imitation also showed the greatest shortened VOT imitation. The correlation was not significant (r = 0.23, t = 1.46, p = 0.151).

We then revisited the same correlations done above: those between the accent imitation task and the three other tasks. However, instead of using the aggregate imitation index based on all words, we ran two sets of post-hoc correlation analyses: one set using the lengthened index and one set using the shortened index, both calculated from the subset of words beginning with /k/. We tested whether individual performance on each of the three tasks was predictive of lengthened and shortened accent imitation separately, within this subset of data, with results shown in Table 8. Participants’ imitation of lengthened VOT patterned the same way as in the original correlational analysis: performance on the accent discrimination and word-level imitation tasks was predictive of performance on the accent imitation task. On the other hand, imitation of shortened VOT was not significantly correlated with either of these sub-tasks, but was significantly correlated with the word-level discrimination task, the one task that had not been correlated with the aggregate value of imitation across all words.

4.1 Summary of results

This work tested the effects of talker variability on explicit imitation of artificial accents with lengthened and shortened VOT, and it explored how individual variability in tasks targeting different perception- and production-based sub-processes predicted imitative performance. Participants imitated artificial accents characterized by shortened, as well as lengthened, VOT, in contrast to previous work showing only lengthened-VOT imitation (Nielsen 2011). Some of the shortened VOT imitation was attributable to the fact that certain tokens were perceived and imitated as a different phonological category (i.e., /p/ perceived and imitated as voiced /b/). Interestingly, even when these tokens were omitted, imitation of shortened VOT was found, albeit to a lesser degree than imitation of lengthened VOT; however, results of this post-hoc analysis need to be interpreted with caution, since it is only based on a subset of data and since our assumptions of perception of the stimuli as voiced vs. voiceless were based on indirect evidence.

Both imitation and discrimination of VOT differences were inhibited by voice-related variability: participants initially presented with a multi-voice condition (DifferentVoice) showed no imitation and had less accurate discrimination performance than participants in a condition with no voice-related variability (SameVoice). However, participants who first completed the condition with no voice-related variability showed equally good performance on the subsequent multi-voice condition, indicating that the inhibitive effect of variability was mitigated by prior experience with the no-variability condition.

Individual imitative performance was independently predicted by two subtasks: accent discrimination and word-level imitation. Performance on the word-level discrimination task was not correlated with overall imitation, but was found to be correlated with the extent of imitation of shortened VOT in a post-hoc analysis. Exploratory analyses did not provide any differences for age-based differences in explicit imitation; however, given that our study was not designed to test this factor, we cannot draw strong conclusions about this and leave more systematic investigation of the effect of age on imitation for future work.

4.2 Sub-processes of imitation

Our first research question focused on the roles of different sub-components of imitation in predicting imitative performance by examining correlations between artificial accent imitation and each of three tasks designed to tap into various sub-processes of variation: accent discrimination, word-level imitation, and word-level discrimination. We found that imitative performance was significantly correlated with the first two of these tasks: individuals who imitated artificial accents more faithfully were more accurate in discrimination of the same artificial accents, and these individuals also showed more faithful imitation of word-level VOT differences. The predictiveness of the accent imitation task indicates that the ability to identify and/or generalize the relevant feature (VOT) as a property of an artificial accent is an important predictor of imitative performance. The predictiveness of the word-level imitation task shows that variability in production-based processes (articulatory precision, flexibility, and/or willingness to diverge from production norms) is similarly important. Furthermore, the fact that these two tasks were independently predictive, uncorrelated with one another, is critical for two reasons. First, it shows that the relevant variability cannot be ascribed to general factors external to the specific tasks (e.g., motivation to complete experimental tasks) or to factors shared by the two tasks (e.g., low-level VOT discrimination acuity). Second, it suggests that performance on the sub-processes targeted by each task patterns differently across individuals, such that an individual might excel at certain sub-components of imitation but not others.

On the other hand, performance on the word-level discrimination task was not predictive of overall performance on the artificial accent imitation task, suggesting that variability in low-level perceptual ability is not a primary driver of differences in accent imitation. This does not mean that low-level perception is irrelevant to accent imitation; in fact, in a post-hoc analysis involving a targeted subset of data, we found that word-level discrimination was correlated with imitation of certain shortened VOT in some cases, as is discussed in more detail below. However, the overall lack of correspondence between the word-level discrimination task and the accent imitation task, as well as the lack of correspondence between the word-level discrimination and word-level imitation task, suggests that sources of variability targeting other sub-processes override differences in low-level perception when accounting for differences in explicit imitation.

The independence of articulatory vs. perceptual predictors of artificial accent imitation highlights the idea that there are multiple, distinct reasons why individuals may vary in imitative ability. While our results are only directly applicable to explicit imitation, we propose that the idea of distinct perception- and production-based predictors may in part account for the difficulty in identifying robust predictors of stable individual differences in imitation more generally (e.g., Cohen Priva & Sanker, 2020; Wade, 2022). While there has been extensive interest in phonetic imitation and its predictors, there has been little research into the level of influence of these predictors. Conceptualizing imitation as a set of sub-processes could provide a clearer framework for testing predictions about characteristics expected to influence imitative ability. It is plausible that there are indeed systematic cognitive traits governing individual variation in imitation, but that simply testing for correlations between the proposed traits and an imitation task may not be the most fruitful way of identifying them. Instead, it is first important to identify which sub-process the trait is expected to influence: some may be predicted to influence perceptual processes (e.g., musicality, Coumel et al., 2019), some to influence low-level production (e.g., “articulation space”; Reiterer et al., 2013), while others may plausibly be expected to influence both (e.g., focus, or lack of attention-shifting; Yu et al., 2013). Making explicit predictions about the level at which a given trait is expected to have an effect, and testing individual sub-processes separately, would result in tighter predictions and higher-powered analyses that would be more likely to result in robust, replicable effects – and facilitate the search of isolating individual predictors amidst the noise of a cognitively complex process.

4.3 The role of talker variability in accent imitation and perception

Based on effects of variability found in past work in accent classification tasks (e.g., Clopper & Pisoni, 2004), phonetic processing (Mullennix et al., 1989; Mullennix & Pisoni, 1990), and accent adaptation/perceptual learning (review in Baese-Berk, 2018), we expected that voice-related variability might hinder performance in both the imitation and discrimination tasks. We did indeed see heavily degraded performance in both tasks for those participants who completed the multiple-voice condition (DifferentVoice) before the no-variability condition (SameVoice). However, we found no disruption at all for those participants who had prior experience with the no-variability condition.

What is the source of the inhibitive effect of voice-related variability? It does not appear to be attributable to a general increased processing cost in the context of greater acoustic variability (e.g., Mullennix et al., 1989; Mullennix & Pisoni, 1990), since a low-level effect of this nature would be expected to persist even after experience with a no-variability condition. Instead, we posit that this occurred at the level of identification: the presence of additional voice-related acoustic variability may have made it more difficult to identify the target feature (VOT) as a relevant feature of the target of imitation. For those participants who first completed the no-variability condition, the acoustic homogeneity may have directed attention to VOT as an appropriate dimension to use as a grouping variable, and this knowledge was retained and used in the subsequent variable-voice session.

The same participants who showed low performance in the variable-voice condition successfully imitated and discriminated catch trials characterized by a rhotic vs. nonrhotic distinction. This indicates that the decreased performance in the presence of voice-related variability is not attributable to general confusion with the task, and also shows that the inhibitory effect is not categorical. Some features appear to be more salient and easier to pick up on as a group-level property of an accent; an interesting topic for future work would be to examine which features are more easily identified as properties of an accent (as with the rhotic/nonrhotic distinction in the current work) and which may need more directing of attention (as with VOT in the current work).

An additional hypothesis was that talker-related variability might facilitate performance in conditions requiring generalization (e.g., Clopper & Pisoni, 2004; Bradlow & Bent, 2008; Schmale et al., 2012). We found little evidence in support of this: while a trending effect of lower accuracy for the generalization sentences in the single-talker condition hints that this may be something to explore further, we cannot make strong claims based on the current findings.

Previous work has posited that, while natural talker-related variability will incur phonetic processing costs, acoustic variability that is purely attributable to properties of a talker’s voice, and not relevant to phonetic identification, is automatically filtered out and therefore does not incur these same processing costs (e.g., overall f0: Sommers & Barcroft, 2006; see also Bradlow et al., 1999; Sommers et al., 1994). The substantial inhibitory effect found in the current work goes against this idea; the voice-related variability clearly inhibited performance on the tasks in this work, despite the fact that the variability was carefully controlled to vary only in properties not relevant for English phonetic identification (overall f0 and formant scaling). One difference between this work and work where no inhibitory effect was found is the nature of the manipulations: while f0 was manipulated in Sommers and Barcroft (2006), none of the previous studies manipulated formant scaling. While it is possible that this additional acoustic variability is the source of the different results, we think it more likely that the discrepancy is due to the different tasks used in the different studies: spoken word identification in the previous studies, compared with imitation and discrimination of group-level features in the current work, which requires more active engagement in deciding which phonetic information is relevant to the task. A direct test of the influence of minimal spectral scaling differences on word identification would be necessary to confirm this.

Finally, our results may point to some practical implications for training paradigms for second language sound training and/or perceptual learning. While high-variability training paradigms have been shown to enhance these kinds of learning, their efficacy is also fairly inconsistent both within and across studies (e.g., Baese-Berk, 2018; Perrachione et al., 2011). As we have seen in the current work, increased variability may make it more difficult to pay attention to, and therefore identify and encode, relevant phonetic detail, presenting a barrier to learning. Ensuring that learners are aware of the relevant dimensions prior to exposure or training, either through explicit instructions or by drawing their attention to the relevant dimensions by including less variability in an early stage of training, as in the no-variability condition in the current work, may provide a relatively efficient way to maximize the effects of high-variability training.

4.4 Explicit vs. implicit imitation of VOT differences

While we did not directly test the difference between explicit vs. implicit imitation in the current study, our findings have implications for the similarities and the differences between them. Recall that previous work has found a greater degree of imitation in explicit compared to implicit tasks (Dufour & Nguyen, 2013; Pardo et al., 2010; Sato et al., 2013). Dufour and Nguyen (2013) posited that the two processes share an automatic general mechanism; however, Sato et al. (2013) found no correlation between the same participants’ performance on explicit and implicit tasks, calling this relationship into question.

These insights from previous work fit in with our conceptualization of the processes underlying explicit imitation and with our empirical results. Under our view, explicit imitation subsumes any automatic processes involved in implicit phonetic convergence, but also requires additional, distinct “controlled” processes that are governed by additional, distinct conditioning factors, including things like willingness to diverge from speech norms, metalinguistic awareness, or experience with imitation. We speculate that there may be greater variability in these traits than there is in those processes governing implicit imitation. If this is the case, then we would expect to see not only more imitation in explicit than implicit tasks (as shown in earlier work), but also more variability, and this is exactly what we found in our lengthened VOT condition. Recall that participants’ imitations of lengthened VOT followed a right-skewed distribution, suggesting that some participants in our experiment were willing and able to make relatively large VOT adjustments, and that there was substantial between-participant variability in the extent of imitation. This skewed distribution stands in contrast to results of Nielsen (2011), in which distributional patterns suggested small but consistent changes in participants’ VOTs after exposure to lengthened VOT. Therefore, in comparison to Nielsen’s (2011) findings of lengthened-VOT imitation, we found both a larger imitative effect (mean 22 ms overall in the current study, vs. 7 ms in Nielsen’s) and a more variable distribution. While the larger magnitude of effect is likely partially attributable to the more extreme VOT values in our lengthened stimuli, we think that differences in the nature of the task augment the effect, and crucially, that the task difference accounts for the greater individual variability found in the distributions.

On the other hand, the distribution for shortened VOT, when considering the subset of stimuli which did not show categorically different perception, showed a shifted distribution, suggesting small-but-consistent changes with little across-participant variability. While any interpretations of this data must remain speculative given the post-hoc nature of the analysis, we propose that a production-based constraint for contrast preservation may be partially responsible for differences in the nature of imitation of shortened, as compared to lengthened, VOT. Given that there is imitation of shortened VOT, this is not a strong, categorical constraint completely inhibiting shortening of VOT. Instead, we propose that the controlled sources of variability that are responsible for individual variation in lengthened VOT imitation are inhibited in the shortened condition, leaving less room for individual variability in imitation.

Our correlation analyses provide some tentative support for this idea. We might expect to find that performance in tasks where low-level perceptual acuity plays a central role is more predictive of more “automatic” types of imitation than it is predictive of imitation where “controlled” processes play a larger role. Consistent with this, we found that performance on a word-level discrimination task (where we assume low-level perceptual acuity is more central) was predictive of variability in shortened, but not lengthened, VOT imitation, while the tasks involving controlled factors (accent discrimination and word-level imitation) were predictive of lengthened, but not shortened, VOT imitation.

4.5 Limitations, future directions, and conclusion

The methodology used in this work diverged from previous work in several ways, such that caution is warranted in the interpretation of our results and in comparison with previous work. First, we designed our tasks to try to tap into different sub-processes of explicit imitation of systematic phonetic variation; however, it is difficult to confirm whether we successfully targeted the intended processes. For example, we designed the accent vs. word-level tasks to try to tap into different targets of imitation (general properties of an accent vs. specific properties of a token) by including talker variability and, in the discrimination stage, sentence variability in the accent discrimination task, and this difference was reinforced in instructions given to participants. Performance on the generalization questions in discrimination suggests that participants were indeed interpreting the variation as we intended, but a direct test of generalization in production would be necessary to confirm this. More broadly, we think that the question of what the intended target is in imitation tasks, and how this might differ based on the instructions and/or presentation context, is one worthy of investigation in and of itself.

Performance on our word-level discrimination task was lower than expected, with 73% accuracy for between-category trials, when we would have expected close to ceiling for these stimuli. While we do not have any clear explanation for this relatively low performance, or the lack of significant effect of Step Difference, it may be due to the specific baseline token we used. Perceptual sensitivity to VOT differences with the same raw values may differ across words (as evidenced by the different imitation strategies shown for the various words in the accent imitation task, Figure 5), or even based on the properties of a particular token of a word. It is possible that the stimuli in the continuum used for our word-level task were particularly difficult to discriminate. This could occur if, for example, secondary acoustic cues in the baseline token contributed to a bias toward /t/ perception, even in the low-VOT range, essentially creating a completely within-category series of stimuli. This possibility points to the importance of using multiple words in this sort of perception task; in future work using the current paradigm, it would be preferable to use the same tokens in both the word- and accent-level tasks, such that only the methodology differs between the two tasks.

While we have drawn tentative conclusions about some aspects of the nature of, and factors conditioning, explicit imitation, as distinct from the better-studied phenomenon of implicit imitation/phonetic convergence, this study did not include a direct comparison between the two types of imitation. Doing so in the future would allow for a test of our predictions about the differences between the two types of imitation outlined above. Another future step, which can be done straightforwardly with this paradigm, would be to expand the range and number of features to further test the role of linguistic selectivity and perceptual salience in imitation. The paradigm also can easily incorporate naturalistic accents, allowing for tests of how factors like familiarity with a specific accent and its social connotations affect phonetic imitation.

Overall, this work explored the nature of explicit imitation of systematic phonetic variation in artificially-constructed accents. Our results provided evidence for the independent roles of perception- and production-based processes in predicting individual imitative ability; this highlights the fact that future work examining how individual differences in social, cognitive, and/or linguistic traits influence imitation should consider the level at which the target trait is expected to exert its influence. Imitation was substantially hindered by voice-related variability during exposure, indicating that even “phonetically irrelevant” variability (f0 and formant scaling) affects the ability to identify features of an accent, and that exposure in the context of less variability can direct attention to the relevant contrast. Finally, we found imitation of shortened as well as lengthened VOT, in contrast to previous work, but saw both qualitative and quantitative differences in imitation of the two manipulations. We hope that the framework provided here can be used in the future to build on and test the generalizability of our results with different linguistic features, with the broader aim of arriving at a fuller understanding of the linguistic, social, and cognitive factors governing explicit imitation.