Because spoken language varies from talker to talker, due to both idiosyncratic differences and accent, listeners have to be adaptable when mapping complex acoustic input onto linguistic representations in the mental lexicon (Bent and Baese-Berk, 2021). Changes to representations (and/or decision processes, see Xie et al., 2023¹) based on listeners' global and recent exposure are supported by multiple leading language models, including exemplar (Johnson, 1997; Pierrehumbert, 2001), non-analytic episodes (Goldinger, 1998; Nygaard and Pisoni, 1998), and Bayesian inference (i.e., “the ideal adaptor”; Kleinschmidt and Jaeger, 2015) models. Under these frameworks, it is posited that listeners create categories systematically linking social groupings and phonetic patterns, including accent-specific representations. On this view, listeners' prior experience with a given accent ought to determine their ability to efficiently and accurately process speech produced with that accent. In the same way that processing speech produced by a familiar talker is faster (Newman and Evers, 2007; Magnuson et al., 2021), processing speech produced in a familiar accent ought to be faster.

Correlational evidence aligns with the supposition that the ability to process an L2 accent accurately and efficiently can be developed over time with sufficient (real world) exposure. For example, Kennedy and Trofimovich (2008) found that both semantically meaningful and semantically anomalous Mandarin Chinese-accented sentences were transcribed with higher accuracy by L1 English listeners who had greater experience with Mandarin Chinese accent. Psychophysiological evidence from pupillometry and eye-tracking has also indicated a benefit of experience: Task-evoked pupil response in Porretta and Tucker (2019) indicated that L1 English listeners who had more experience with Mandarin Chinese accent processed Mandarin Chinese-accented English words (presented in noise) more easily, and gaze behavior in Porretta et al. (2020) demonstrated that greater experience with Mandarin Chinese accent also resulted in faster speech processing for Mandarin Chinese-accented English. Further behavioral evidence from L1 Dutch listeners also suggests that activation of L2-accented words depends on listener experience with the target accent. Witteman et al. (2013) used a cross-modal lexical decision task in which each trial participants were presented auditorily with a German-accented word or non-word in Dutch followed by an orthographic probe word or non-word in Dutch. The listeners' task was to make lexical decisions for the visual probe items. Results indicated that listeners with less experience with German accent were less primed by auditory items that were strongly accented than participants with greater experience.

Altogether, these studies suggest a critical role of prior experience in L2 accent processing, aligning with predictions from episodic models of speech processing. What remains to be empirically determined, however, is the amount and rate of exposure that is necessary to observe a benefit of prior experience. Under all of the theoretical frameworks mentioned above (exemplar, non-analytic episodes, and Bayesian inference), listeners with more experience with a given accent ought to be more adept at processing speech produced with that accent. The present study aims to test these theoretical models of speech processing.

Only a small number of studies have investigated the benefits of prolonged exposure to L2 accent using causal experimental design. Within a single experimental session, rapid improvements to listening performance and reductions of listening effort are well-documented (Clarke and Garrett, 2004; Bradlow and Bent, 2008; Sidaras et al., 2009; Baese-Berk et al., 2013; Brown et al., 2020). Beyond a single experimental session, however, investigations of sustained benefits of L2 accent exposure are limited, and those that exist have produced mixed results.

Examining both transcription accuracy and comprehension of Korean-accented English materials, Lindemann et al. (2016) found a sustained benefit of L2 accent exposure at 1–2 days post-training. The authors presented L1 English listeners with either L2 accent exposure or an explicit linguistic training (i.e., teaching phonemic differences between Korean and English, etc.). Notable for the present study, participants in the L2 accent exposure group completed a speech transcription task in which the correct sentence was presented each trial after submitting the typed response. At test, both training groups had better sentence transcription accuracy than a control group; comprehension scores were the same across groups. Thus, the results of Lindemann et al. demonstrate that a (brief) training session with L2 accent can lead to improvements in perceptual accuracy that last into subsequent day(s).

Sustained adaptation to L2 accent over a half-day (12-h) period—as well as generalization—was also demonstrated in a sleep consolidation study conducted by Xie et al. (2017). L1 English listeners in the study were trained with word-length stimuli from a Mandarin Chinese-accented talker, focusing on a key accented phoneme (/d/). All subjects completed a test with a novel Mandarin Chinese-accented talker immediately after training as well as a second iteration of this test 12-h later, but for half of the subjects this 12-h period spanned the day (e.g., 8 a.m. to 8 p.m.) and for the other half is spanned the night (e.g., 8 p.m. to 8 a.m.). In both groups, retention of the training benefit was observed. Critically, however, the overnight group showed unique generalization of learning to another Mandarin Chinese-accented speaker and phonemic category (/t/), suggesting that sleep consolidation promoted generalization of learning.

Whether benefits of L2 accent exposure are retained over intervals longer than 1 day, however, remains unclear. Bieber and Gordon-Salant (2021), for example, failed to find evidence of a training benefit in test sessions administered 1 week after training. In their study, the authors examined accent-generalizable learning (i.e., performance on a novel/untrained accent following training with multiple other accents) for speech presented in six-talker babble. They employed a dual-task paradigm (similar to the one used in the present study), which combines a speech transcription task with a simultaneous reaction time-based visual task (it is assumed that with finite cognitive resources, reaction times will slow for the secondary visual task as the demands of the primary speech task increase). L1 English-speaking young adults and older adults with and without hearing loss completed three experimental sessions across approximately 3 weeks, where the beginning portion of Week 2's session and Week 3's session each served as measures of retention. Results indicated that within an experimental session listeners rapidly improved, as in prior work (Bradlow and Bent, 2008; Baese-Berk et al., 2013). However, the benefit of each prior week's training session on transcription accuracy was not retained (i.e., the beginning of the Week 2 and 3 sessions did not demonstrate improvement). Reaction times from the secondary measure, on the other hand, were significantly improved for Week 3's session, in particular, indicating that the cognitive load associated with L2 accent processing may have been reduced.²

Focusing on both listening performance and attitudes toward L2 speakers, Derwing et al. (2002) implemented what appears to be the longest L2 accent training protocol to date, occurring over an 8-week period. The authors sought to train L1 English listeners to better understand L2 (specifically, Vietnamese) accent, comparing the effects of a training with explicit phonetic lectures and a training with cross-cultural awareness lectures. Unfortunately, results of the study indicated no significant benefits of either training for speech transcription or comprehension. Attitude questionnaires, however, did reveal that both training groups showed increased empathy toward immigrants, and participants given explicit phonetic training reported increased confidence in their ability to understand L2 accent.

Altogether, the current body of empirical evidence suggests that benefits of L2 accent training sessions may persist into subsequent days but diminish over longer (week-long) intervals. Additionally, benefits observed for cognitive load may diverge from those observed for listening performance (i.e., recognition/transcription accuracy). Based on these observations, in the present study we sought to examine the benefits of a training protocol administered over multiple consecutive days. From Pre-Test to Post-Test, we also incorporated a measure of cognitive load (similar to Bieber and Gordon-Salant, 2021) to determine whether different benefits may be observed for measures of listening performance vs. cognitive load.

In the present study, we implemented a combination of dual-task paradigms and a speech transcription tasks over a 5-day period. On Days 1 and 5, participants completed a Pre-Test and Post-Test (dual-task paradigm), and on Days 2, 3, and 4 participants completed exposure-based training sessions (speech transcription). Participants were randomly assigned to one of three groups for the training days: Control (no exposure to L2 accent or feedback during training), Exposure (exposure to L2 accent but no feedback during training), and Feedback (exposure to L2 accent and feedback during training). All groups had the exact same Pre- and Post-Test with both L1- and L2-accented speech stimuli.

We predicted that response times to the secondary task in the dual-task paradigm (an index of cognitive load) would be shorter on Day 5 than Day 1 for all groups, indicating improvement on the task itself. Critically, we expected that this improvement would be greater for the Exposure and Feedback training groups—particularly in the L2-accented speech condition—than it would be for the Control group. Additionally, we predicted that the Feedback group would show greater reduction in cognitive load than the Exposure group, given that the feedback manipulation provided lexical context to guide perceptual adaptation.

For listening performance (speech recognition accuracy) in the Pre-Test and Post-Test data, we had similar predictions, although we also anticipated the possibility that subjects may demonstrate reduced cognitive load without gains in listening performance (as in Bieber and Gordon-Salant, 2021). We expected that speech recognition scores from the primary task would be larger on Day 5 than Day 1 for all groups (indicating improvement on the task itself), and that the Exposure and Feedback training groups would improve more than the Control group in the L2-accented speech condition, in particular. We also predicted that the Feedback group would show greater improvement than the Exposure group.

Lastly, we planned to examine listening performance (speech transcription³ accuracy) data from the training sessions on Days 2, 3, and 4. We predicted that, if any differences existed, they would be as follows: Higher scores for the Feedback group than the Exposure group overall, and an interaction with days reflecting greater improvement for the Feedback group over time.

Young adult subjects (age mean = 19.5; age range = 18–30) were recruited from Washington University in St. Louis's Psychology Participants Pool. Inclusion criteria (set via demographic filters in SONA Systems) selected for L1 English speakers with normal hearing and vision (or corrected-to-normal vision). Additional criteria on the SONA listing indicated that subjects should not sign up for the study if they had extensive exposure to Mandarin Chinese (for example, they should not speak Mandarin Chinese, have studied Mandarin Chinese, or have parents or roommates who are fluent in Mandarin Chinese). Subjects who did not complete all 5 days of the study were excluded from analyses.

Due to COVID-19-related recruitment issues, we decided to combine a pilot version of the experiment with the main dataset to reach more desirable sample size (N = 160). We report full details regarding the minor differences between the pilot and primary subject groups below. In brief, the two differences were: (1) During the dual-task sessions for the primary group but not the pilot group, the practice session provided feedback instructing participants to “speed up” if they took longer than 3 s to respond; and, (2) Additional measures of cognitive ability (not analyzed in the present manuscript) were not collected from the pilot participants. Results of all analyses remained the same when accounting for time of participation (i.e., when including a two-level fixed effect denoting “pilot” vs. “main” experiment status); because time of participation did not improve model fits or impact the outcomes for the effects of interest, this factor was dropped from all models.

After combining the two datasets, the sample size by group was as follows: Control n = 54, Exposure n = 53, Feedback n = 53. In the pilot version of the experiment, a total of 43 subjects participated. Two subjects were excluded from this sample for failing to complete all days of the study, one for reporting exposure to Chinese, and three for having an average reaction time in the dual-task paradigm > 3,000 ms (the significance of this cut-off is discussed further in the Procedures section). After exclusions, 37 valid subjects remained (by group: Control n = 11, Exposure n = 14, Feedback n = 12). For the main experiment, 152 subjects participated in total. Twenty-nine of these subjects were excluded for one the following reasons: Failing to complete all 5 days of the study (24), self-reporting too much prior exposure to Mandarin Chinese (four), and, in one case, self-reporting a receptive and productive language disorder. After exclusions, 123 valid subjects remained (by group: Control n = 43, Exposure n = 39, Feedback n = 41).

We report information about participants' language experience by random assignment group in Table 1. All participants reported English as their primary language, and as a language learned from birth. All participants reported at least one additional language (which is to be expected given high school language requirements in the U.S.). As can be seen in Table 1, a fairly large proportion of the sample can be classified as simultaneous (21%) or early (8%) bilinguals; these trends are unsurprising when considering current estimates of bilingualism in the United States (~1 in 5 speak a language other than English at home; Dietrich and Hernandez, 2022). Including bilingual status as an effect in the response time and accuracy analyses did not change the results or improve model fits, and was thus not included as a factor in the final models.

Generalized linear mixed-effects regression was used to model the recognition accuracy data in R (version 4.0.4; R Core Team, 2021) with the glmer() function from the lme4 package (Bates et al., 2015). Likelihood ratio tests were conducted to determine the significance of effects of interest, and p-values for model parameters were estimated using the lmerTest package (Kuznetsova et al., 2017). Recognition accuracy was treated as a grouped binomial, meaning that models predicted performance using two columns of data (number of correct words, number of incorrect/missed words) for each sentence. A logit link function was specified. Fixed effects included: Condition (dummy-coded levels: L1 accent, L2 accent), Session (dummy-coded levels: Pre-Test, Post-Test), Group (dummy-coded levels: Control, Exposure, Feedback), as well as all possible two- and three-way interactions between Condition, Session, and Group. Random intercepts were included by item and by subject. Random slopes of Day and Condition were attempted but ultimately removed from all models due to issues with model singularity. Model syntax is provided in Supplemental materials.

For the response time data, linear mixed-effects regression was implemented with the lmer() function. Fixed effects included: Condition (dummy-coded levels: L1 accent, L2 accent), Session (dummy-coded levels: Pre-Test, Post-Test), Group (dummy-coded levels: Control, Exposure, Feedback), and all two- and three-way interactions between Condition, Session, and Group. In all models, random effects included random intercepts by subject and by item, and random slopes of Condition and Session by subject. Model syntax is provided in Supplemental materials.

We acknowledge the possibility that limited statistical power may have encumbered our ability to detect significant training benefits in the present study. In the response time data, in particular, the degree of variance may have reduced our ability to detect effects. We suggest increasing the number of trials per condition in future work when comparing dual-task data across sessions or using a cross-modal matching task in place of a dual-task paradigm (also referred to as a “semantic” or “linguistic” dual-task paradigm; Strand et al., 2018). Although we created 100 novel stimuli in addition to the 200 SNST items, across 5 days this resulted in only 39 items per accent condition per task. In lieu of a larger set of sentence recordings, one solution would be to repeat items, particularly on training days, in future research (see Bradlow and Bent, 2008; Baese-Berk et al., 2013). In Pre- and Post-Test measures, it is critical to include novel stimuli in order to prevent item-specific learning effects from artificially inflating performance; however, analyses of training sessions are typically less crucial, and items could be repeated. There may even be a benefit to repeating items across training sessions, although to our knowledge this has yet to be examined directly. It is our hope that the present study can serve as a benchmark when selecting paradigms and estimating power in future investigations of multi-day accent trainings. We recommend that future studies maximize potential effect sizes via a combination of the following methods: (1) Using more sensitive measures of cognitive load, (2) Increasing the number of trials in test sessions, (3) Increasing the number of trials in training sessions, and (4) Increasing the number of participants.

With regard to the first recommendation, we predict based on prior evidence from speech-in-noise perception (Strand et al., 2018) that cross-modal matching tasks may produce larger effects than dual-task paradigms that involve a non-linguistic secondary task—although direct comparisons of the sensitivity of these paradigms for L2 accent perception/adaptation have yet to be conducted. Comparing the results of Clarke and Garrett (2004) with Brown et al. (2020), however, provides some indication of what types tasks may be most sensitive in the context of L2 accent adaptation: In Clarke and Garrett (2004), rapid (single session) adaptation to L2 accent was robustly demonstrated both across four experimental blocks (each containing 16 trials) and within the early trials. In that study, a cross-modal matching task was used; specifically, participants completed a task where they responded “yes” or “no” to visually-presented sentence-final probe words. In contrast, Brown et al. (2020) used the same non-linguistic dual-task paradigm as the present study and only found evidence of rapid adaptation within the first 20 trials, not across the full 50-trial session. Other differences between the two studies (type of L2 accent, presence of background noise, etc.) may account for the deviating outcomes, but we (cautiously) recommend based on outcomes of these prior studies that researchers may be best served with cross-modal matching tasks in future work. We can also (more confidently) recommend pupillometry as a measure of cognitive load, which proved to be more sensitive than the dual-task paradigms in both Strand et al. (2018) and Brown et al. (2020).

From Pre- to Post-Test, (non-significant) trends in the data indicated larger benefits for the measure of listening performance than the measure of cognitive effort. This outcome runs counter to the findings of Bieber and Gordon-Salant (2021), in which benefits were observed at a test session 1-week after training for cognitive load but not listening performance. One possible explanation for the contrary outcome in the present study may be the design of the middle (training) days, which utilized a speech transcription task rather than the dual-task paradigm from the Pre- and Post-Test sessions. Thus, participants received more extensive training with the linguistic task, but not the non-linguistic (visual) task, from the dual-task paradigm. It may be the case that these training sessions were better situated to promote near-transfer to the more similar (linguistic) task. In future work, matching the designs of training and test sessions may be ideal to remove any potential differential transfer effects by task type.

One strength of the present study was the inclusion of L2-accented stimuli presented in quiet, as opposed to in noise. Although adding noise to stimuli can make it easier to match intelligibilities across conditions (e.g., matching L1 to L2 speakers), prior evidence also indicates that the cognitive and/or perceptual resources recruited to support noisy vs. accented listening conditions may differ (McLaughlin et al., 2018). Thus, when examining questions pertaining to the perception of L2 accent, using L2-accented stimuli presented in noise may not always be suitable. To prevent ceiling effects in the present study, we decided to use semantically-anomalous sentences (e.g., “the wrong shot led the farm”), which pose a different potential issue: Namely, anomalous sentences reduce a listener's ability to use top-down information during speech processing, and are therefore less ecologically valid. Studies that use these types of items thus give a more direct assessment of bottom-up processing at the cost of limited generalizability of the findings. In future work, focusing solely on measures of cognitive load (as opposed to a combination of cognitive load and intelligibility measures) can remove these types of obstacles and allow for more ecological examinations of accent accommodation.