Convergence is a phenomenon in which an individual shifts their behaviour towards another individual or group. It is observed in many behavioural domains (Mercado III et al., 2014), including speech. Convergence in pronunciation is sometimes referred to as phonetic convergence. A speaker is likely to be influenced by the phonetic variants used by surrounding speakers. Evans & Iverson (2007) provide a good case study of the long-term effects of phonetic convergence. They explored how Northern England English accent speakers modify their speech when in contact with Standard Southern British English speakers. It is well known that the FOOT-STRUT split and TRAP-BATH split did not occur historically in Northern England English (Wells, 1982: 196; 232); i.e., the word cud is homophonous with the word could in Northern England English, and the word bath is produced with the same vowel phoneme as the word trap. Evans & Iverson demonstrated that Northern accent speakers gradually changed their formant values of the STRUT and BATH vowels after they attended universities, where they were likely to interact with Standard Southern British English speakers. Their habitual formant values slightly shifted towards those of the STRUT and BATH vowels in Standard Southern British English.

Such long-term convergence arises as the cumulative effect of online convergence effects across multiple utterances. A speaker can adjust the probability of producing a particular phonetic variant in accordance with an interlocutor. For example, a Norwich speaker is known to stylize their speech by choosing one of the socially meaningful variants such as [t] or [ʔ] in the production of /t/ in word-medial position (e.g., butter and water). Bell (1984) shows that a speaker changes the rate of /t/ glottalization depending on how often the interlocutor produces the glottal variant [ʔ] during the conversation. To take another example, a French speaker chooses one of the two phonetic variants [dʒ] and [ʒ] in the production of an affricate-initial loanword such as jeans. Lev-Ari & Peperkamp (2014) demonstrate that a French speaker changes the rate of selecting the phonetic variant [dʒ] depending on how often their interlocutors produce the variant during a spontaneous language game. There is abundant evidence that a speaker converges towards an interlocutor during interactive tasks such as map tasks (Pardo, 2006; Cao, 2023) and diapix tasks (Kim et al., 2011; Biro et al., 2022).

Phonetic convergence is also observed even when the interlocutor is not physically present. Many experimental studies demonstrate that a speaker adjusts their pronunciation towards a model talker, when and after the speaker shadows (repeats) model speech stimuli (Goldinger, 1998; Shockley et al., 2004; Babel, 2009; Fiasson, 2015; Pardo et al., 2017; Kwon, 2019; MacLeod & Di Lonardo Burr, 2022), and after they listen to model speech stimuli (Goldinger, 2000; Goldinger & Azuma, 2004; Delvaux & Soquet, 2007; Nielsen, 2011; Yu et al., 2013; Hashimoto et al., 2022). In these experimental studies, the participants first read aloud target words, and the recordings are treated as baseline utterances. Following this, the participants are exposed to model speech stimuli through a shadowing task or a listening task. Some studies analyse shadowed utterances in order to assess convergence. On the other hand, some studies employ a postexposure session, where the participants read aloud target words as in the baseline session, and the utterances in the postshadowing session are the target of analysis. Comparing the baseline utterances versus shadowed utterances (or postexposure utterances) allows researchers to explore lab-based speech convergence. Previous studies assess the similarities between model stimuli and shadower’s utterances using either a perceptual experiment (i.e., an AXB experimental paradigm) or an acoustic analysis of a particular phonetic feature (e.g., formant values, spectral peak values, VOT, segmental duration, f0, and intensity). The studies cited above all demonstrate that shadowed utterances or postexposure utterances are more similar to model speech stimuli than baseline utterances are. These results can be interpreted to mean that speakers converge towards variants produced by model talkers once they are exposed to the model speech stimuli of these variants in non-interactive tasks.

These findings suggest that a speaker’s mental representations can be continuously restructured through imitative tasks (Goldinger, 1998; Tilsen 2009). Exemplar Theory provides one straightforward account for lab-based phonetic convergence. Upon exposure to model speech stimuli, unique traces may be stored in the listener’s memory, and the episodic memories or “exemplars” retain phonetically rich information (Johnson, 2006). In speech production, multiple exemplars are chosen as “contributors” and averaged together to form a production target. Exemplars with higher resting activation levels are more likely to be chosen as contributors than those with lower resting activation levels (Pierrehumbert, 2001; Wedel, 2006). The newly stored exemplars have relatively high resting activation levels due to the recency of storage. Hence, they are likely to contribute to the creation of a production target afterwards, and therefore the shadowed and postshadowing utterances are likely to have phonetic features that are similar to the model speech stimuli.

In many of the previous studies about lab-based convergence, the model talker(s) and participants in the experiments share the same language background; i.e., they speak the same language and dialect. Some previous studies explore “cross-dialect convergence,” in which model talkers and participants speak different dialects of the same language (Delvaux & Soquet, 2007; Babel, 2010; Dufour & Nguyen, 2013; German et al., 2013; Walker & Campbell-Kibler, 2015; Gnevsheva et al., 2021; Clopper et al., 2024; Lee-Kim & Chou, 2024). These previous studies have demonstrated that a speaker shifts their production towards a model talker with a different dialectal background when and after shadowing model speech stimuli. Cross-dialect convergence offers an excellent opportunity to develop our understanding of how novel variants are treated by the cognitive system and how they are mentally stored in relation to sociolinguistic categories. The overarching goal of this study is to investigate the mental representations and cognitive mechanisms that are involved in phonetic convergence.

Based on the above convergence literature, we can safely assume that such a cross-dialectal convergence effect will be observed in New Zealanders. As noted at the outset of this study, the DRESS vowel is produced as a high vowel [i] in NZE, whereas it is realized as a mid vowel [ɛ] in AmE. Accordingly, when NZE speakers are asked to shadow AmE DRESS words, they should produce a lower vowel (with higher F1 values) than in their baseline production. Such convergence effects may also persist in postshadowing productions. These input-driven convergence effects are predicted by hypotheses concerning the continuous restructuring of mental representations reviewed above. However, an important open question is whether the effect is limited to the DRESS vowel, or whether shadowing DRESS also affects other vowels. In the next section we review literature suggesting that social knowledge can lead to more general system-wide priming effects.

Previous literature suggests that exposure to a social concept (or a different dialect) shifts the perception and production of a phoneme. For example, Hay et al. (2006) demonstrate that speakers change their perception of KIT vowels upon seeing the word associated with a dialect. The participants in their experiment listened to continua of the target vowels (KIT, DRESS, and TRAP). The formant values were acoustically manipulated, and they ranged from Australian-English-like variants to NZE-like variants. Their task was to indicate which token in the continuum best-matched the target vowel in the word underlined in their answer-sheet. The word “New Zealander” was written at the top of each answer sheet in one condition, and the word “Australian” was written in the other. They found that the female participants were more likely to choose an NZE-like KIT vowel token when “New Zealander” was written on the answer sheet than when “Australian” was written on it, even though most of the participants believed that the talker was a New Zealander. These results suggest that seeing a regional label associated with a particular dialect influences vowel perception.

Several studies show that activation of a dialect region can affect production. Love & Walker (2013) and Sanchez et al. (2015) explored spontaneous speech, and found that speakers shift their production towards a particular dialect when discussing topics associated with the dialect. More specifically, Love & Walker (2013) demonstrate that English speakers living in the United States (US) produce more AmE-like rhotic variants when talking about American football than when talking about English soccer, and Sanchez et al. (2015) report that NZE speakers produce KIT vowels with Australian-English-like F1 values when talking about Australia than when talking about something neutral. The results of these two studies indicate that talking about something related to a particular dialect leads to the shift towards the dialect.

Wade (2022) provides evidence that AmE speakers shift their speech production in accordance with their expectations regarding linguistic behaviours of a talker. The participants in their research read aloud target words including PRICE /aɪ/ by guessing the target words based on their definitions. The clues were visually displayed on the PC screen in the baseline session, whereas they were auditorily provided in the exposure session. The talker was either a Southern talker or a Midland talker in the exposure session. The participants never heard PRICE vowels in the exposure session. The results of their experiment demonstrate that PRICE is more likely to be produced with Southern-like, glide-weakened tokens, in the exposure session than in the baseline session, if the clues were provided by the Southern talker in the exposure session. These results suggest that exposure to Southern-like features except PRICE vowels influences the phonetic realization of PRICE vowels, providing evidence that activating a dialect which is familiar to the participant can have system-wide effects.

Results like those above, and similar, have led multiple authors to posit that social information and phonetic information are stored together, and are linked (see. e.g., Sumner et al. 2014). An intuitive account of how this could work can be captured by a mechanism we will refer to here as “social category activation.” English speakers may have opportunities to hear a variety of dialects, for example through social media and TV. Hence, the mental lexicon may store exemplars that encode speech from other dialects. Exemplars form an exemplar cloud on the basis of sociolinguistic characteristics (Foulkes & Docherty, 2006), and they may be linked with social categories such as AMERICAN and AUSTRALIAN in the lexicon. The exposure to a social concept raises the activation of the relevant social category, in turn increasing the activation levels of the exemplars associated with that social concept (Drager & Kirtley, 2016). Consequently, the exemplars associated with the social category become more readily available afterwards; i.e., the speech perception and production shift towards the variants associated with that social concept to some extent. Such an effect has not been tested in the context of a shadowing paradigm, but the above results suggest that it should be observed. The current study therefore addresses the following key question in order to test such an effect in a shadowing paradigm.

It may be possible that imitation of a particular vowel from a different dialect generalizes to another vowel phoneme. Nielsen (2011) demonstrated that the imitation of an acoustically manipulated token can be generalized to words and phonemes that participants did not listen to. She ran an experiment in which participants listened to words beginning with [p^h] that were acoustically modified to have extremely long VOT values. The participants read aloud not only the words with word-initial /p/ but also other words with word-initial /p/ or /k/ in the preexposure and postexposure sessions. She compared the preexposure tokens with the postexposure tokens, and found that the unheard words were produced with longer VOT values, along with the heard words, in the postexposure session than in the preexposure session. In addition, the results show that the participants produced /k/ as well as /p/ with longer VOT values in the postexposure session than in the preexposure session. These results provide the evidence that the imitation can be systematically generalized at abstract levels to the words and phonemes that participants did not listen to during the experiment (see also Fiasson, 2015; Kwon, 2019). The current study aims to extend these previous findings to cross-dialect convergence. More specifically, we investigate whether imitation of a particular vowel indexed to a different dialect is systematically generalized to another vowel phoneme that is likewise indexed to that dialect.

The generalization of imitation from one vowel phoneme to another is predicted by the hypotheses concerning the mental mechanism of social category activation reviewed above. Under this account, exemplars are assumed to be associated with social categories in memory, and exposure to a social concept may trigger spreading activation between social categories and their associated exemplars. This mechanism makes a prediction that a shadower may modify the phonetic values of a vowel phoneme to which the shadower was not exposed in the shadowing session. This is because exposure to a particular phoneme from a different dialect in a shadowing session may raise the activation of the relevant social category, and thereby exemplars associated with the social category may become more readily available in the postshadowing session. Consequently, the postexposure production may shift towards the different dialect in general. Testing this prediction will contribute another piece of evidence to the growing body of literature documenting the dialect priming effects on speech, and explicitly test the degree to which there are effects in shadowing beyond the particular shadowed variant.

As noted at the outset of this study, NZE provides an interesting test case to address the key question. Many young NZE speakers are exposed to and familiar with AmE to some extent even if they have never lived in the US. This observation has been made by some previous studies. For example, Walker & Campbell-Kibler (2015) note that “New Zealanders have a large amount of exposure to American accents through popular media,” and they demonstrated that NZE speakers could identify a talker as an AmE speaker when they listened to AmE accents. Gibson (2019, 2024) notes that the use of AmE accent is default for NZ singers when they play pop music and Gibson & Hay (2025) show that music specifically leads NZ listeners to have a lower DRESS-TRAP boundary. These observations suggest that NZE speakers may store exemplar memories that encode AmE speech to some extent, and the exposure to AmE model speech stimuli may thus raise the activation of the social category “American English.” As New Zealanders are familiar with AmE, then exposure to one vowel in a shadowing task would be predicted to have potential effects on the later production of other vowels that differ between the varieties.

In particular, the current study explores whether NZE speakers modify the phonetic values of the TRAP vowels after shadowing DRESS vowels. As discussed above, the TRAP vowels in NZE are generally higher than their correspondences in AmE. Using this test case, the general prediction presented above can be narrowed down. Our specific prediction is that NZE speakers may produce TRAP vowels with a lower tongue position (with higher F1 values) after shadowing the model speech stimuli including DRESS vowels produced by an AmE model talker. In the experiment reported below, NZE speakers were asked to shadow words with DRESS vowels that were produced by an AmE model talker, and never heard TRAP words from the model talker. They were asked to produce TRAP words in the preexposure and postexposure sessions.

It is well known that words with higher neighbourhood density (i.e., more minimal pair competitors) are more difficult to recognize and process in speech perception. Luce & Pisoni (1998) is a classic study that explored minimal pair effects on word recognition. They ran an auditory lexical decision task (Experiment 2). Half of the words employed in their experiment were real words, and the other half were nonce words. The words varied in terms of neighbourhood density. The participants listened to audio stimuli, and decided whether the heard word was real as quickly as possible. Their results demonstrated that real words with many neighbours were recognized more slowly than words with few neighbours. Vitevitch & Luce (1998) ran a word repetition task to study the effects of minimal pair competitors on word recognition. The participants in their experiment were asked to repeat a word as quickly as they heard the word. Their results showed that words with a larger number of minimal pair competitors were repeated less quickly.

In speech production, words with minimal pair competitors are more likely to be hyperarticulated than those without minimal pair competitors. Baese-Berk & Goldrick (2009) ran a speech elicitation experiment (Experiment 2), in which participants read aloud target words beginning with voiceless stop consonants /p, t, k/. Half of the target words had a minimal competitor word with a voiced stop consonant, but the others did not. Their task was to instruct a listener to “Click on the (target word)” while seeing two other dummy words on the PC screen. Sometimes, one of the dummy words was a minimal pair competitor word beginning with a voiced stop. Their experimental results show that speakers produced word-initial /p, t, k/ with longer VOT values in relation to target words that have minimal competitor words in terms of word-initial voicing than those which do not have minimal pair competitor words. In addition, they demonstrated that speakers produced word-initial /p, t, k/ with longer VOT values when seeing a minimal pair competitor word as a dummy word than without seeing a minimal pair competitor word. Stern & Shaw (2023) conducted a replication of this finding. They employed pseudowords instead of real words as target words. The pseudowords included word-initial voiceless stop consonants, some of which had lexical minimal pair competitors in terms of word-initial stop consonant voicing. Their task was to instruct a listener to “type the (target word) number” while seeing a target word along with dummy words. Dummy words included a minimal competitor word beginning with a voiced stop consonant in some trials. Their results show that the VOT values of the word-initial stop consonants were enhanced by a minimal pair competitor word on the screen only when the competitor was a real word.

These minimal pair effects also occur in spontaneous speech production. Wedel et al. (2018) explored natural speech data, and analysed the phonetic realizations of stop consonants /p, t, k, b, d, g/ and six nonback vowels /i, ɪ, e, ɛ, ʌ, æ/. They found that the voiceless stop consonants /p, t, k/ are produced with longer VOT values if there is a minimal pair competitor in initial-stop voicing in the English lexicon. Likewise, the voiced stop consonants /b, d, g/ are produced with shorter VOT values if there is a voiceless stop minimal pair competitor. Furthermore, their statistical results suggest that the six nonback vowels are produced with more contrastive formant values if there is a minimal pair competitor in at least one of the neighbouring vowels.

There have been a variety of hypotheses to account for the fact that words with minimal pair competitors are hyperarticulated in speech production. One of the most straightforward accounts is that a speaker has a bias ensuring the intelligibility of a speech signal, and makes an effort to maintain the distinctiveness of linguistic units such as words and phonemes. A speaker is hypothesized to increase the intelligibility of a speech signal when they think that the message transmission is unlikely to be successful (Lindblom, 1990; Pate & Goldwater, 2015; Hall et al., 2018). This bias is referred to as the intelligibility bias in what follows. As noted in the beginning of this section, words with minimal pair competitors tend to be more confusing than those without minimal pair competitors. Hence, a speaker may hyperarticulate words to differentiate them from minimal pair competitor words, and thereby increase the accuracy of message transmission.

Another account comes from the memory storage process. Previous literature argues that phonetically ambiguous words are occasionally discarded and not stored as new exemplars in memory (Wedel, 2006; Sóskuthy, 2015; Todd et al., 2019). If words with minimal pair competitors are produced with reduced speech signals, which are thereby similar to their competitors, they may be underrepresented in the memory storage process. Consequently, clear speech signals that are distinct from minimal pair competitor words tend to leave unique traces in memory. On the other hand, this kind of confusion does not take place in relation to words without minimal pair competitor words. In consequence, speech signals with various degrees of intelligibility can be stored in memory with regard to words without minimal pair competitor words. Through this memory storage process, lexical categories that have minimal pair competitor words are represented with a larger number of exemplars that encode hyperarticulated forms than those which do not have minimal pair competitor words. We note that these two accounts (i.e., the intelligibility bias account and the memory storage account) are not mutually exclusive, and make a similar prediction regarding the minimal pair effects on hyperarticulation.

As noted in Section 1, one goal of this study is to explore whether minimal pair effects can be extended to cross-dialect phonetic convergence. Dialects may differ in the phonetic realization of a given phoneme (Wells 1982: 73); as a result, certain words said in a particular dialect sound similar to different words in another dialect. If imitating model speech stimuli that are produced by a talker with a different dialect background has a risk of blurring a lexical distinction between minimal pair words, do shadowers nonetheless imitate the model speech stimuli to the same degree? To the best of our knowledge, the effects of minimal pairs have not yet been examined in the context of cross-dialect convergence.

NZE provides an interesting test case to address the key question. As reviewed in the outset of this study, AmE DRESS vowels are perceptually close to NZE TRAP vowels. This proximity suggests that AmE DRESS vowels may be ambiguous between the two vowel phonemes, potentially making it difficult to identify the intended words. The current study explores the extent to which NZE speakers converge towards AmE model talkers in relation to DRESS depending on the existence of a minimal pair competitor word with TRAP. If NZE speakers converge towards AmE speakers in their pronunciation of DRESS vowels, the contrast between DRESS and TRAP vowels may become less clear than usual. This situation may impair the lexical intelligibility of some minimal pair words between the two vowels (e.g., bed vs. bad) to some extent. On the other hand, the imitation of AmE DRESS vowels may not affect the lexical intelligibility of certain words, such as deaf and jet, as they do not have minimal pair competitors with TRAP vowels such as *daf and *jat.

Importantly, our prediction with respect to this question is conditional on the results of question (1).1 If shadowing AmE DRESS has no effect on the production of the New Zealander’s TRAP vowel, this suggests that there is not a system-wide effect caused by social category activation. If AmE DRESS vowels and NZE TRAP vowels are simultaneously active, then the AmE DRESS vowels may be ambiguous between the two vowel phonemes, potentially making it difficult to identify the intended words. Lexical ambiguity arises particularly for DRESS words that have minimal pair competitors containing TRAP vowels. The imitation of AmE DRESS may affect the lexical intelligibility of some DRESS words (e.g., bed and guess) which have minimal competitor words with TRAP (e.g., bad and gas), whereas it may not affect the lexical intelligibility of some DRESS words which do not have minimal competitor words with TRAP (e.g., deaf and jet). Based on the hypotheses concerning the memory storage process and the intelligibility bias reviewed above, our prediction in this case is that NZE speakers may converge to AmE DRESS vowels in the words that have minimal pair competitor words with TRAP vowels to a lesser extent than in the words that do not have minimal pair competitor words with TRAP vowels. More specifically, NZE speakers may produce DRESS vowels with a higher tongue position (with lower F1 values) in relation to DRESS words that have minimal pair competitor words with TRAP vowels than DRESS words without minimal pair competitor words with TRAP vowels, when and after shadowing AmE model speech stimuli with DRESS.

However, if speakers shift their TRAP vowel in the postshadowing session, then this suggests that exposure to AmE English has activated both AmE DRESS and TRAP vowels. In this case, the speakers have shifted to an overall more AmE-mode, and thus the contrast between the two vowels would be maintained and not in danger. Under this scenario, DRESS imitation would occur independently of the presence of a minimal pair competitor containing TRAP. Specifically, no difference in F1 values would be expected between DRESS words with and without TRAP minimal pair competitors in either the shadowing or postshadowing sessions.

In sum, our predictions for the two key research questions are interlinked. The degree to which this effect is predicted is deeply entangled with the question of whether exposure to the AmE DRESS vowel just affects the speaker’s production of DRESS, or whether it shifts the whole vowel space they are operating with in the direction of AmE vowels. If it is just the DRESS vowel, then the local-variety minimal pair TRAP words are highly overlapping with the target DRESS, and we would expect the intelligibility bias and memory storage process to impede the degree of convergence. But if the whole vowel space is affected and AmE TRAP vowels are thus also activated, then the integrity of the lexical contrast remains intact, and the produced DRESS vowel is not in danger of overlapping with the activated AmE-like TRAP representations. As New Zealanders have reasonable experience with AmE, and the literature shows evidence from other tasks that social category activation can have system-wide effects (see Section 1.2), we regard the latter to be the most likely.

Twenty-four female NZE speakers participated in the current experiment as shadower participants. They identified themselves as NZE speakers. The age of the participants ranged from 18 to 34. We limited our data collection to female speakers to control for any gender effects. It is known that gender plays some role in vocal imitation (Namy et al., 2002; Babel, 2009; Babel et al., 2014; Lee-Kim & Chou, 2024), but our interest does not lie in the gender effects.

Monosyllabic English words including DRESS and TRAP vowels were first collected as target word candidates using CELEX2 (Baayen et al., 1995). Amongst them, 80 words were chosen as target words. There were 40 words containing each of the target vowels DRESS and TRAP that have a minimal pair competitor with the other target vowel, and 40 words including DRESS or TRAP that do not have a minimal pair competitor with the other target vowel. The existence of a minimal pair competitor was identified using CELEX2. There were three exclusion criteria. First, we excluded words with prenasal vowels. This is because prenasal TRAP vowels are known to be produced with a higher tongue position in AmE (Labov et al., 2006: 84). Second, we discarded words with vowels followed by /l/ and /ɹ/. The reason is that vowel-to-approximant boundaries are not always clear, and allophonic variation is triggered by the following /l/ and /ɹ/ (Di Paolo et al., 2010: 88). NZE and AmE can also differ in the degree to which these phonemes are present, as NZE is largely nonrhotic, and can have extensive /l/-vocalization (Hay et al., 2008: 35). We removed words with extremely high word frequency such as have and that. This is because word frequency is known to influence the degree of imitation (Goldinger, 1998; Babel, 2010; Pardo et al., 2017; Hashimoto et al., 2025). The list of target words is provided in Table 1.2

The shadower participants were asked to read aloud the 80 target words in the baseline and postshadowing sessions, and shadow the 40 target words with DRESS vowels in the shadowing sessions (see Section 3.3 for the detail). In the baseline and postshadowing sessions, they were also asked to read aloud 10 filler words (butter, city, dirty, eighty, fighter, party, pretty, shutter, tutor, and water). The filler words largely differ from the target words in terms of the phonological structure. They include neither DRESS nor TRAP vowels, and are disyllabic words. In the shadowing sessions, 16 filler words were added to the word list. Eight of them included LOT vowels (box, fox, pot, shop, shot, socks, stop, top), and the others included FLEECE vowels (beach, cheese, deep, seed, seek, speed, speak, teeth). These words were employed as filler words because shadowing these words may allow a shadower participant to capture the vowel space of a model talker, and distract them from the target vowels. The two vowels are two of the most peripheral vowels in the two varieties of English (i.e., FLEECE is the highest front vowel, and LOT is the lowest back vowel).

The 56 shadowing words were presented to two model talkers. One talker comes from California, the US, and the other comes from New Zealand. They were both young female speakers. They were individually recorded reading aloud the 56 words at a sound-attenuated room in Joetsu University of Education in Japan, equipped with a recorder (TASCAM DR-40X) and a head-mounted microphone (AKG C-520). The sampling rate was set to 44,100 Hz, and the resolution was 16 bits. The recordings of the 56 shadowing words were used as model speech stimuli in the shadowing experiment. The intensity of the sound files was equalized using a Praat script.

Along with the 56 shadowing words, they were also asked to read aloud a variety of words with multiple vowels. Based on the recordings, Figure 2 was generated. The left-hand figure shows the vowel space of the AmE model talker, and the right-hand figure shows that of the NZE model talker. The vowel space was defined on the basis of only nonrhotic monophthongs, and the formant values were converted into Bark (Traunmüller, 1990). See Section 2.4 for the formant measurement and Bark scaling. Notice that AmE and NZE are different in terms of their phonemic systems; i.e., AmE has nine non-rhotic monophthongs (KIT, DRESS, TRAP, STRUT, FOOT, FLEECE, LOT, THOUGHT, and GOOSE) while NZE has eleven (KIT, DRESS, TRAP, STRUT, LOT, FOOT, FLEECE, NURSE, START, THOUGHT, and GOOSE). Note that the accent of the AmE talker may have a vowel system with eight monophthongs due to the cot-caught merger (Wells, 1982: 473–475; Labov et al., 2006: 58; Ladefoged & Johnson, 2015: 97). As is clear in the figures, the location of the NZE model talker’s DRESS vowel is relatively higher than that of the AmE model talker’s DRESS vowel, and that of the NZE model talker’s TRAP vowel is relatively higher than that of the AmE model talker’s TRAP vowel. In addition, the acoustic values of AmE DRESS (Bark-scaled F1: 7.6, Bark-scaled F2: 12.7) and those of NZE TRAP (Bark-scaled F1: 6.9, Bark-scaled F2: 13.6) are relatively close to each other. These vowel distributions are in line with the descriptive generalizations provided by the previous studies (see Section 1).

Figure 3 shows the Bark-scaled F1 values of the DRESS vowels of the model stimuli. The vertical axis indicates the Bark-scaled F1 values, and the horizontal axis expresses the model talker. Note that the y axis is inverted to align with vowel height. The white boxes show the distribution of the F1 values of DRESS words that do not have minimal pair competitor words with TRAP vowels, and the grey boxes show DRESS words that have minimal pair competitor words with TRAP vowels. Table 2 provides descriptive statistics of these F1 values. As is clear, the F1 values of the model stimuli produced by the AmE model talker are higher than those produced by the NZE model talker. We fitted a mixed-effects linear regression model with F1 as the dependent variable and dialect (AmE vs. NZE) as a predictor. The model suggests that this difference is statistically significant (p < .001). This suggests that the NZE model talker produces DRESS vowels with a higher tongue position than the AmE model talker. Once again, these descriptions of the F1 values corroborate the previous findings (see Section 1). It seems that the F1 values are slightly higher in relation to the DRESS words that have minimal pair competitor words with TRAP vowels than the DRESS words that do not have minimal pair competitor words with TRAP vowels. For each dialect, we fitted a linear regression model with F1 as the dependent variable and minPair (no vs. yes) as a predictor. The model suggests that the difference is statistically nonsignificant (p > .05), and thus there is no evidence that the F1 values of DRESS vowels in the model stimuli systematically differ depending on the existence of a minimal pair competitor word with TRAP.

The shadower participants were individually invited to a sound booth in the University of Canterbury in New Zealand. The participants were assigned to either Group A or Group B. Group A participants shadowed model speech stimuli produced by the NZE model talker in the first shadowing session and those produced by the AmE model talker in the second shadowing session. Group B participants shadowed the speech stimuli produced by the two model talkers in the reverse order. That is, the shadower participants all shadowed the two model talkers, but the order was counterbalanced between the two groups. The NZE shadowing session (control condition) was necessary because the difference between baseline and shadowed tokens (and postshadowing tokens) could be attributed to extraneous factors such as fatigue and acclimation to the experimental settings (see Martin, 2007: 196–198). To assess the true effect of the AmE shadowing (target condition), we will discuss not only the difference between the baseline and target conditions but also the difference between the baseline and control conditions in what follows. It was also necessary to counterbalance the order of the two conditions (see Martin, 2007: 156–164). Once participants are exposed to one condition, the exposure effects may last afterwards, with the result that the participants’ speech behaviour may be contaminated under the other condition. In addition, participants may feel more tired in the second shadowing session than in the first shadowing session, and the difference between the two sessions could be attributed to extraneous factors. In order to minimize these order effects on the experimental results, Group A and B participants shadowed the two model talkers in the reverse order. One might wonder why we did not employ a pure between-subject design, in which each participant shadowed one model talker and no participant shadowed two model talkers. The biggest disadvantage of the pure between-subject design is that the two groups may not be equivalent to each other accidentally in terms of individual characteristics such as perceptual acuity and experience (see Martin, 2007: 151; 197–198); i.e., the difference between the two shadowing conditions might be confounded with accidental differences between the two groups. In addition, the pure between-subject design limits the number of observations that can be collected from each participant. Our experimental design allows us to collect more data, thereby further strengthening its statistical power. Even though we took extra care of the effects of session orders and individual differences, it is possible that confounding effects may still be present. This is why the statistical analyses reported in Section 3 will take the variable called group into consideration (see Section 2.5).

The experimental program was created using jsPsych (de Leeuw, 2015). The visual stimuli were presented via computer monitor (Microsoft Surface Laptop), and the audio stimuli were given over headphones (audio-technica ATH-AVC 200). The recordings were performed using a recorder (TASCAM DR-40X) and a head-mounted microphone (AKG C-520). The sampling rate was set to 44,100Hz, and the resolution was 16 bits.

In the beginning of the experiment, the participants completed a keyword reading session. They were asked to read aloud eleven keywords while seeing written forms on the PC screen: heat, hid, head, had, hard, hut, hot, horse, hood, whose, and heard. These words include the eleven monophthong phonemes in NZE, and the vowel phonemes follow /h/ and precede alveolar consonants. This consonantal context is chosen to control allophonic variation (Di Paolo, 2010: 88). This initial session is supposed to familiarize the participants with the experimental settings, and the recordings in this initial session were used to capture the individual differences in the vowel spaces.

Next, they did a baseline session, and read aloud the 80 target words in Table 1 plus the ten filler words while seeing written forms on the PC screen. The recordings of the target words were treated as baseline tokens. Next, they did the first shadowing session. The model speech stimuli were NZE tokens for Group A participants, whereas they were AmE tokens for Group B participants. They shadowed the 40 target words with DRESS vowels plus the 16 filler words with LOT or FLEECE vowels. Recall that the participants never heard the TRAP vowels from either model talker. They were not explicitly instructed to imitate the model speech stimuli, but were asked to repeat them as naturally and clearly as possible (see Babel, 2009). The prerecorded model speech stimuli were presented over headphones, and written forms appeared on the PC screen. The presentation of written forms was supposed to make sure that participants recognized which word they heard (Walker & Campbell-Kibler, 2015; Gnevsheva et al., 2021). These recordings were treated as shadowed tokens. Following this session, they did the first postshadowing session. They read aloud the 80 target words plus the ten filler words as in the baseline session. The recordings in this session were treated as postshadowing tokens. After the first shadowing and postshadowing sessions, the participants were asked to have a short break for three minutes. The purpose of this short break was to mitigate the effects of the first shadowing session.

Following the short break, they did the second shadowing session. They shadowed a different model talker from the one they shadowed in the first shadowing session. The model speech stimuli were AmE tokens for Group A participants, whereas they were NZE tokens for Group B participants. Then, they did the second postshadowing session, which is basically identical to the baseline and first postshadowing sessions. The recordings in the second shadowing and postshadowing sessions were treated as shadowed tokens and postshadowing tokens respectively. Note that the order of stimulus presentation was randomized in each of the five key sessions.

Finally, they filled in questionnaires. They answered several questions about their backgrounds (e.g., age, gender, and region where they were born and grew up). In addition, they were asked which dialect the first and second model talkers spoke, and why they thought so. According to the questionnaire data, 17 of the 24 participants correctly identified the AmE model talker as an AmE speaker, five of them identified her as a Canadian English speaker, and the other two participants identified her as a British English speaker. This suggests that the majority of the shadower participants may raise the activation of “North American English” at least. As for the NZE model talker, all the participants correctly identified her as a NZE speaker. The experimental procedure can be summarized as in Table 3.

Our interest lies in the formant frequency values of DRESS and TRAP vowels. We manually annotated the vocalic domains of the target word tokens in the five key sessions using Praat (Boersma & Weenink, 2020). We defined the vocalic domain by taking into consideration the periodicity and intensity of waveform. In addition, we annotated the eleven monophthongs recorded in the keyword reading session. Using a Praat script, the formant measurements were performed. In particular, the central 50% of the vocalic parts were extracted using a Hamming window, and the average formant values of the extracted regions were measured with five formants tracked and the maximum frequency set to 5,500 Hz.

Following this, the raw frequencies were converted into Bark (Traunmüller, 1990). It is well known that human auditory systems are more sensitive to differences in low frequency sounds than high frequency sounds (Johnson, 2012: 88). The Bark scale reflects the fact that the sensation of lower frequency sounds is more minutely divided than that of higher frequency sounds. Bark scaled F1 values were fitted as a response variable in the statistical models reported in Section 3. This variable is called Bark1 in what follows.3

Our general interest lies in whether NZE speakers adjust their production of DRESS to the AmE model talker during the shadowing session, and whether the adjustment is observed after the shadowing session, that is, during the postshadowing session. Furthermore, a key question is whether the convergence effects can be extended to the unheard vowel, TRAP, in the postshadowing session. In order to explore these convergence effects, we coded our dataset depending on the experimental sessions (see Table 3). The tokens produced in the baseline session were coded as baseline tokens. Those in the AmE shadowing session were coded as AmE shadowing tokens, and those in the following postshadowing session were coded as AmE postshadowing tokens. Likewise, the tokens produced in the NZE shadowing session were coded as NZE shadowing tokens, and those produced in the following postshadowing session were coded as NZE postshadowing tokens. This multi-level variable is called task in the following statistical analyses, and fitted as a key variable. Note that this variable is a five-level variable for DRESS vowels (“baseline”, “AmE shadowing”, “AmE postshadowing”, “NZE shadowing”, and “NZE postshadowing”), while it is a three-level variable for TRAP vowels (“baseline”, “AmE postshadowing”, and “NZE postshadowing”). This is because TRAP words were not employed in the shadowing sessions. This variable is treatment-coded, and the reference level is “baseline”. If the participants converged towards the model talkers, statistical significance would be expected for this variable.

Our key interest also lies in whether the imitation of a word with DRESS is influenced by the presence of a minimal pair competitor word with TRAP. To address this research question, we coded a token in our dataset as “yes” if the word has a minimal pair competitor with the other target vowel. For example, the tokens of bad and bed were coded as “yes”, whereas the tokens of fact and deaf were coded as “no”. This binary variable is called minPair in what follows, and fitted as a key variable. This variable is treatment-coded, and the reference level is “no”. If the extent of phonetic convergence were influenced by the presence of a minimal pair competitor, the interaction between minPair and task should reach statistical significance.

The tokens were also coded depending on the group. The tokens produced by Group A participants were coded as “NZE_first”, whereas those produced by Group B participants were coded as “AmE_first”. The reason why this variable is fitted as an explanatory variable in a statistical model is because the first exposure to a particular accent may influence speech production throughout the experiment (see Section 2.3). This binary variable is called group in the following statistical analyses, and fitted as a control variable. This variable is treatment-coded, and the reference level is “NZE_first”. If the exposure effect in the first shadowing session persisted until the end of the experiment, statistical significance would be expected for the interaction between group and task.4

Recall that the 24 NZE speakers read aloud the 80 target words with DRESS and TRAP vowels in the baseline session and the two postshadowing sessions, and shadowed the model speech stimuli of the 40 target words with DRESS vowels in the two shadowing sessions (see Table 1 for the list of target words). With respect to target words, each speaker produced 80 baseline tokens, 80 shadowed tokens, and 160 postshadowing tokens. Hence, we aimed to collect 1,920 baseline tokens (960 DRESS vowel tokens and 960 TRAP vowel tokens), 1,920 shadowed tokens (1,920 DRESS vowel tokens), and 3,840 postshadowing tokens (1,920 DRESS vowel tokens and 1,920 TRAP vowel tokens) in total.

Some tokens were removed from the following statistical analyses. First, we examined the extent to which each participant’s F1 values for the DRESS and TRAP vowels changed during the shadowing and postshadowing sessions compared to the baseline session. We found that one participant was more than 3 SD from the ground mean value in several sessions, indicating that they behaved in a very different manner from the rest of the participants. This participant was therefore removed from the statistical analysis.

Second, a token was removed if its F1 frequency value deviated by 3 standard deviations from the mean F1 value of an individual participant for each of the two target vowels within a session. In the end, 7,299 tokens remained in our dataset. The statistical analyses were performed on the 7,299 tokens in what follows.5

Before proceeding to the main statistical analyses, we present descriptive statistics to explore patterns in the vowel productions of the shadower participants. Figure 4 shows the grand mean formant values of the 11 NZE monophthongs (KIT, DRESS, TRAP, STRUT, LOT, FOOT, FLEECE, NURSE, START, THOUGHT, and GOOSE) across the shadower participants in the keyword reading session. The scales of the horizontal and vertical axes of Figure 4 are identical to those of Figure 2. It exhibits some well-known characteristics of the NZE vowel system, such as DRESS raising, KIT centralization, and GOOSE fronting, suggesting that the shadower participants generally had a typical NZE vowel system. A comparison of the shadower participants’ mean vowel space and the NZE model talker’s vowel space (shown in the right panel of Figure 2) reveals a similar vowel distribution, suggesting that the NZE model talker is also a typical NZE speaker. The figure also indicates that the target vowels, DRESS and TRAP, are distinguished primarily by F1 values.

Next, we would like to examine whether the DRESS vowels of the shadower participants tend to be higher in the vowel space compared to those of the AmE model talker, and whether the TRAP vowels of the shadower participants tend to be close to the DRESS vowels of the AmE model talker. Figure 5 compares the formant values of the shadower participants and the model talkers. The green dot expresses the mean value of the DRESS vowels in the NZE model speech stimuli, and the red dot indicates the mean value of the DRESS vowels in the AmE model speech stimuli. The individual black dots in the left-hand figure express the mean values of the DRESS vowels of the shadower participants in the baseline session, and those in the right-hand figure represent the mean values of the TRAP vowels of the shadower participants in the baseline session. The left-hand figure clearly shows that the DRESS vowels of shadower participants are higher than those of the AmE model talker in vowel space. The right-hand figure suggests that the majority of the shadower participants produced TRAP vowels with formant values that are close to the DRESS vowel of the AmE model talker. These findings corroborate the previous studies (Gordon et al., 2004; Hay et al., 2008, 2015; Ladefoged & Johnson, 2015), and suggest that our data is well suited for addressing the research questions outlined in Section 1.

Finally, we would like to explore the F1 values of DRESS and TRAP vowels in each key session. The top panel of Figure 6 visualizes the distribution of F1 values for DRESS, and the bottom panel shows the distribution for TRAP. The x axis expresses task, and the y axis shows Bark-scaled F1 values. The red horizontal line represents the mean Bark F1 value of the AmE target model stimuli (mean: 7.61), while the green line represents that of the NZE target model stimuli (mean: 4.23). Recall that the shadower participants never shadowed the TRAP vowels from either model talker, and the target model stimuli included only DRESS vowels. Table 4 shows the descriptive statistical values. Clearly, the production of DRESS shifted towards AmE model speech stimuli in the AmE shadowing task. It is worth noting that the mean DRESS F1 value in the AmE shadowing session is still far from that of the AmE model stimuli, suggesting that shadower participants did not completely align their production with an AmE model talker. The other potential changes made in the shadowing and postshadowing tasks are not so clear from these descriptive statistics, and they are further explored using mixed-effects linear regression analyses in the next section.

The statistical analyses were performed using the lme4 library 1.1-35 (Bates et al., 2015) implemented in R (R Core Team, 2024) with the aid of the optimizer optimx 10.21 (Nash, 2014). As discussed in the preceding section, the response variable is a bark-scaled F1 value (Bark1), and the explanatory variables are task, minPair, and group. The DRESS and TRAP vowels were analysed separately. First, we constructed the maximal model with the three-way interaction of the predictor variables, by-participant and by-word random intercepts. Then, backward elimination was performed iteratively, manually removing variables one by one. In the course of the backward elimination, we incorporated by-participant random slopes for the key variables so that the models would converge reliably. The elimination process was based on p values, which were obtained by comparing a fuller model with a reduced model lacking a particular variable or interaction term. Model comparisons were done using anova. If a model comparison shows no significance (p > .05), then the reduced model was adopted; otherwise, the fuller model was adopted. Following this procedure, the two models in (3) were selected as the best-fitted models. The best-fitted models did not have any statistical convergence errors. With respect to the DRESS model, the variable group was removed from further analysis because its main effect and interaction terms with other predictors were not significant (p > .05). With regards to the TRAP model, the variables group and minPair were removed from further analysis because their main effects and interaction terms with other predictors did not improve the model (p > .05). VIF values were checked to avoid overfitting problems. All VIF values are lower than 4, suggesting that there is no multicollinearity in the best fitted models. The model summaries are provided in Tables 5 and 6. As the lme4 library does not provide p values, we used lmerTest library 3.1-3 (Kuznetsova et al., 2017) to retrieve the p values in the tables.

Together, these two best fitted models can answer our research questions. We will outline these results and our background assumption (that DRESS would be affected by AmE), and our two research questions in separate subsections. Recall that the variable group was removed through the backward elimination due to its lack of statistical significance. The statistical nonsignificance suggests that the effects discussed below were observed regardless of the assigned group.

The current study contributed another piece of empirical evidence for cross-dialect convergence. As reviewed in Section 1.1, the previous studies demonstrated that a shadower participant shifts their phonetic articulation in the direction of the model speech, even though the shadower participant and model talker have different dialectal backgrounds (Delvaux & Soquet, 2007; Babel, 2010; Dufour & Nguyen, 2013; German et al., 2013; Walker & Campbell-Kibler, 2015; Gnevsheva et al., 2021; Clopper et al., 2024; Lee-Kim & Chou, 2024). Our experimental results demonstrated that NZE speakers modified the production of DRESS vowels while and after shadowing the model speech stimuli including AmE DRESS vowels. The NZE shadower participants produced DRESS vowels with lower F1 values in the baseline sessions as compared with the AmE model talker (see Figure 5). It was demonstrated that they produced DRESS vowels with significantly higher F1 values in the AmE shadowing session than in the baseline session (see Table 5). In addition, the shadower participants produced DRESS vowels with significantly higher F1 values in the early AmE postshadowing session than in the baseline session (see Section 3.2.1). The shadower participants did not exhibit such shifts in the NZE shadowing and postshadowing sessions, suggesting that the shifts towards AmE DRESS vowels in the AmE shadowing and postshadowing sessions cannot be attributed to extraneous factors such as fatigue and experimental acclimation.

Exemplar Theory provides a straightforward account of these cross-dialect convergence effects (see Section 1.1). When a NZE speaker hears AmE model speech stimuli, the auditory experiences may be stored as new exemplars in memory. The newly stored exemplars have relatively high resting activation levels due to the recency of memory storage (Pierrehumbert 2001; Wedel 2006). Hence, the recently experienced AmE DRESS vowels are likely to influence the creation of the production target in the AmE shadowing and postshadowing tasks. Consequently, the shadowed and postshadowing utterances are likely to have phonetic features that are similar to the speech stimuli that are produced by the AmE model talker (e.g., higher F1 values with respect to DRESS vowels).

The observed cross-dialect convergence may be attributed not only to this bottom-up processing, but also to top-down mechanisms. In particular, social category activation may also contribute to the convergence effects observed for DRESS. Exposure to AmE model speech stimuli may raise the activation of the relevant social category “American English,” which in turn increase the activation levels of the preexisting exemplars that encode AmE DRESS vowels. Consequently, exemplars encoding AmE DRESS may become more readily available when speakers form a production target (see Section 4.2.1).

The best fitted model (Table 5) suggests that the convergence effects are larger in the AmE shadowing session than in the AmE postshadowing session. The estimated difference in Bark-scaled F1 between the baseline and AmE shadowing sessions is about 0.581, whereas the difference between the baseline and AmE postshadowing sessions is about 0.018. That is, the shifts towards AmE DRESS vowels were attenuated in the AmE postshadowing session. This attenuation of the convergence effects can be interpreted to mean that the resting activation levels of newly stored exemplars gradually decreased over time since they were stored. Newly stored exemplars may have a strong influence on the formation of a production target at the time of storage, but this influence may diminish over time. Once the activation levels of these exemplars drop significantly, speech production is likely to rely more on the preexisting exemplars that habitually contribute to the creation of a production target in daily speech. Pierrehumbert (2001) refers to this process as “memory decay,” wherein the resting activation levels of exemplars gradually decrease over time. Wedel (2006) demonstrates that this process plays an important role in sound change. Our findings provide further empirical evidence consistent with memory decay.

According to our statistical analyses, there is no robust evidence that the convergence effects differed between Groups A and B. As discussed in Section 3.2, the variable group did not improve the statistical models, as neither its main effects nor its interactions with other predictors were significant. This can be interpreted to mean that the two groups were comparable, and the order of dialect exposure did not affect the extent of phonetic imitation. This finding is consistent with Walker & Campbell-Kibler (2015). In their experiment, the participants shadowed model speech produced by speakers of NZE, Australian English, US Midland English, and US Inland North English. Some participants shadowed the model talkers in this order, while others shadowed them in the reverse order. With respect to DRESS vowels, the participants in the experiment converged towards the AmE model talkers regardless of the order in which the model talkers were presented. The absence of exposure-order effects may suggest that newly stored exemplars are fragile and susceptible to rapid memory decay. Once a shadower is exposed to a new model talker, the exemplars encoding the speech of the previous model talker may quickly become less accessible.6

The best fitted model (Table 5) suggests that the shadower participants exhibited a relatively small degree of convergence towards the AmE model talker. The estimated Bark-scaled F1 value of DRESS in the baseline session is approximately 4.69, whereas in the AmE shadowing session, it is about 5.26; i.e., the shift is only 0.57. The mean F1 value of the AmE model speech stimuli was 7.61 (see Table 2). That is, the difference between the AmE shadowed tokens and model speech stimuli was approximately 2.35, suggesting that shadower participants had room for further phonetic imitation. A similar finding was also reported by Walker & Campbell-Kibler (2015). In their experiment, NZE speakers shadowed AmE model speech stimuli with various vowel phonemes. The NZE speakers shifted their DRESS vowels towards the AmE model talker by around 0.6 in Bark-scaled F1, although there remained room for further phonetic imitation. The limited extent of imitation is not unique to cross-dialect convergence but has also been observed in within-dialect convergence. Babel (2010) investigated the extent to which speakers of Californian English converge towards Californian English model talkers. As in the current study, the participants shadowed model speech stimuli. The vowel plots in their study showed that shadowed utterances moved closer to the model talker’s formant values, although a measurable distance between the two remained. Nielsen (2011) also investigated within-dialect convergence. In their Experiment 1, AmE speakers were exposed to words containing lengthened /p/. The mean VOT value of the lengthened /p/ was about 113 ms. The AmE speakers produced /p/ with about 66 ms in the preexposure session and approximately 73 ms in the postexposure session, indicating substantial room for convergence.

Why phonetic imitation is relatively limited, and why ample room for further imitation remains, may be attributable to several factors. One possible explanation comes from Exemplar Theory. Under this account, multiple exemplars are selected and averaged to form a production target (see Section 1.1). Although a recently heard utterance may exert a relatively strong influence on target formation, exemplars that habitually contribute to the production target continue to shape a production target. As a result, shadowed and postshadowing utterances may reflect both the model stimuli and speakers’ habitual productions, thereby limiting the extent of convergence towards the model talker. An alternative explanation concerns constraints imposed by the articulatory motor system. Even when speakers attempt to shift their productions towards a model talker, articulatory constraints may limit the extent of such adjustments. Supporting this view, previous research has shown that L2 learners often struggle to produce nonnative sounds, even when explicitly instructed on articulatory movements (Saito, 2011). In the case of cross-dialect convergence, social constraints may also play a role. Sociolinguistic variation conveys social meanings associated with particular groups and values (Bell, 2014, 270–271; Eckert, 2016). Shadowers may therefore avoid producing variants that deviate too strongly from their habitual speech patterns, as such variants may convey undesirable social messages.