Lexically guided retuning of phonemic categories is an example of perceptual learning that is critical to our ability to adapt to variability in speech, including novel dialects, accents, and individual pronunciation differences. When listeners are exposed to sounds with nonprototypical acoustic profiles, for example, fricatives that are acoustically intermediate between /f/ and /s/, embedded into disambiguating lexical items, e.g., ungrateful, they learn to categorize such instances as exemplars of /f/, thus expanding the boundaries of the phonemic /f/ category. This reorganization can be detected via a forced-choice categorical decision test whereby more items on the acoustic continuum between /f/ and /s/ are categorized as /f/, as a result of lexically guided perceptual learning. Such findings are well established in single-language experiments with L1 (presumably largely monolingual) speakers of the language (Kraljic & Samuel, 2006; Norris et al., 2003; Reinisch et al., 2014). Only recently, researchers began examining lexically guided recalibration in participants’ non-L1 languages and across languages in bilinguals. A very limited body of work presently indicates that (a) recalibration is not limited to listeners’ L1 and (b) recalibration can generalize across languages, such that learning to expand phonemic categories in one language (L1 or L2) results in a similar expansion for the comparable category in bilinguals’ other language (L2 or L1). For example, Reinisch et al. (2013) showed that recalibration of /s/ and /f/ categories occurred for L1 speakers of German trained and tested on Dutch, their L2. The same study demonstrated that Dutch listeners recalibrated /s/ and /f/ categories in the Dutch test after training on the Dutch-accented English of the same speaker who provided the test recordings. Importantly, the authors interpreted these findings as evidence that listeners considered non-prototypical pronunciation of /f/ and /s/ to be the intrinsic characteristic of the speaker, thus generalizing this attribute across the languages spoken by the same person. The authors also noted that given the highly similar acoustic properties of /f/ and /s/ across Dutch and German languages, the establishment of language-specific categories is likely unnecessary.

Schuhmann (2016) examined cross-linguistic phonetic recalibration of /s/ and /f/ in L1 German learners of English and L1 English learners of German. Both groups were trained on English but tested on both English and German. Phonetic recalibration was observed in all four cases, including L1 English speakers tested on L2 German and L2 English speakers tested on L1 German, where the training language and the test language did not match. Given that the effect sizes of perceptual recalibration were not identical across exposure and test language conditions, the author concluded that L1 and L2 categories were linked but not merged (separate but interconnected) across languages. Of note again is the fact that /s/ and /f/ are acoustically highly similar in English and German, thus their strong cross-linguistic link is to be expected. Another important aspect of the design is that the same speaker recorded both the training and the testing stimuli in English and German, which may have induced speaker-specific learning effects. Finally, the test items were nonwords, where only the quality of /r/ (approximant or trill) indicated the ‘language’ of the stimuli, making it difficult to evaluate the generalizability of the findings to real words.

Recently, Caudrelier et al. (2023) and Caudrelier et al. (2024) investigated the recalibration of /s/ and /f/ in French-English bilinguals in Montreal, Canada. As observed in the previous two studies, recalibration was found to transfer from the training language (English or French) to the unmatched test language (French or English), but the recalibration effect was smaller in the cross-language than in the within-language condition. The authors concluded that bilingual listeners may have shared /f/ and /s/ phonemic categories across their two languages, but the links may be stronger across closely related (Germanic) languages than in the case of Romance-Germanic language combination. The study also adopted a single-speaker design: The training and test recordings were provided by the same French-English bilingual speaker, raising the possibility of listeners accommodating to a specific speaker, rather than generalizing abstract perceptual learning across languages.

What all these studies share is the exclusive use of /s/ and /f/ for cross-linguistic recalibration, while recognizing that these phonemes are highly phonetically similar across the languages under investigation. Additionally, all previous studies adopted a single-speaker design in recording audio stimuli, with the same speaker recording both the exposure and the training stimuli. Because of these methodological choices, their findings, while possibly indicating shared phonemic categories across languages, could also be due, at least in part, to speaker-specific perceptual learning effects. Yet lexically guided recalibration is not limited to speaker-specific scenarios. For instance, Kraljic and Samuel (2006) showed that recalibration trained on /d/ and /t/ generalized to another speaker and a parallel phonemic contrast, /b/ and /p/, indicating that the process can be speaker-independent and feature- rather than phoneme-based (see also Llompart, 2024, for cross-talker generalization in perceptual retuning of vowels by advanced learners of English). Importantly, Kraljic and Samuel (2006) suggested that the use of fricatives, which are known to provide rich speaker-specific information, can incentivize phoneme-specific and speaker-specific learning (although Kraljic and Samuel (2005) also show partial speaker generalization with fricatives). To minimize the speaker-specific learning effects, they suggested the use of stops, given the weaker speaker-specific spectral cues, compared to fricatives (Repp, 1984). This hypothesis was further supported in Kraljic and Samuel (2007), who reported speaker-specific recalibration for fricatives but more generalized recalibration for stops. In the present study, we recruited different speakers to record the exposure (training) and the testing stimuli across languages. Both our speakers were female, ensuring acoustic similarity that is believed to facilitate cross-speaker generalization (Kraljic & Samuel, 2005; Tamminga et al., 2020). We also investigated recalibration of the place contrast in stop consonants instead of fricatives to mitigate speaker-specific learning effects.

The choice to use stops allows us to address an additional question. The consensus in the literature on bilingual phonetics and phonology is that for L2 categories that are highly phonetically similar to their L1 counterparts, to the point of being perceptually indistinguishable, creation of distinct categories is either impossible or unnecessary (e.g., PAM-L2; Best & Tyler, 2007). Thus, the inquiry into cross-linguistic phonetic influence has focused on pairs of sounds that are phonetically similar but non-identical across languages, such as aspirated voiceless stops in English versus unaspirated voiceless stops in Spanish. For such pairs, their cross-linguistic relationship was theorized to be more dynamic and subject to developmental changes, including eventual, but not immediate, creation of distinct L2 categories (SLM-r; Flege & Bohn, 2021). For cross-language pairs such as these, the generalization of lexically guided recalibration from the training language to a different test language is not necessarily a foregone conclusion. The use of this methodology allows us to investigate whether phonetically similar but non-identical categories are linked or integrated across languages to such an extent that transfer of phonetic recalibration from one language to another can occur. It is important to note that given the hypothesized dynamic nature of the cross-linguistic relationship between such sounds, the findings may differ depending on the circumstances of bilingualism and the stage of L2 acquisition.

The present study investigates lexically guided recalibration of perceptual boundaries between the voiceless stops /p/ and /k/ in Spanish-English bilinguals residing in the United States. Exposure to sounds acoustically intermediate between /p/ and /k/ was provided via a lexical decision task, and resulting recalibration was tested via a two-alternative forced choice (2AFC) categorization task with real words of English and Spanish. Two different L1 speakers recorded English and Spanish stimuli. Bilingual participants were recruited online and randomly assigned to one of the four groups: Two groups were trained in English, and two in Spanish. Within each language group, some participants were exposed to ambiguous /p/, while others were exposed to ambiguous /k/. Regardless of training language, all participants completed the 2AFC identification task in both English and Spanish. As a result, for two of the groups, the training and test language matched (with the same speaker providing the training and the test recordings), whereas for the other two groups, the training and test languages differed (with a different speaker providing the training and the test stimuli). The research question we are addressing with this experimental design is the following: Do bilinguals who were trained to shift the perceptual boundaries of the place contrast between /p/ and /k/ in one of their languages (e.g., English) demonstrate resulting recalibration in both languages, Spanish and English? Crucially, voiceless stops in English are aspirated in word-initial position and at the onset of syllables with any degree of stress, while in Spanish voiceless stops are unaspirated across the board (Lisker & Abramson, 1964). Thus, for the transfer of recalibration to occur, bilinguals must categorize Spanish voiceless unaspirated and English voiceless aspirated stops together, despite their phonetic differences. We hypothesized that the transfer of recalibration would be observed across languages. This hypothesis was motivated by frequent findings in previous literature that voiceless stops across languages, such as English and Spanish, are subject to cross-linguistic influence in the assimilatory direction, although assimilation can be asymmetric as the more dominant or earlier acquired language tends to exert a stronger influence on the less dominant or later acquired language (e.g., Flege & Eefting, 1987, among others). These findings suggest that voiceless stops can often be linked or integrated across languages in bilingual cognitive grammar.

An important question, however, is what exactly is linked cross-linguistically: context-specific allophones of voiceless stops or abstract phonological representations of voiceless stops. SLM-r assumes the former, while PAM-L2, the latter. Finally, L2LP upholds that L1 and L2 perceptual grammars are separate and independent from each other. These assumptions make different predictions, given the methodology of our investigation, as discussed in the following section.

A total of 127 Spanish (L1)-English (L2) bilingual speakers participated in the study. They were recruited and compensated on Prolific and performed experimental tasks via Gorilla (Anwyl-Irvine et al., 2020).4 Among these, 17 participants yielded unreliable data in the categorization task, and their data were excluded from analysis, resulting in a total of 110 participants (Age M = 29.5, Age range = 18–48). We identified unreliable data using a combination of statistical analysis (mixed-effects logistic regression on categorization responses for each participant) and visual/manual inspection of response patterns and reaction times. Participants were excluded if their responses were not significantly affected by Step (an acoustic continuum between /p/ and /k/), as this indicated that they were not systematically influenced by the stimuli continuum from /p/ to /k/. Participants were also excluded if they selected the same response at least five consecutive times, despite variation in Step, and if this pattern occurred more than once.

All retained participants reported acquiring Spanish first and English later. All participants were residing in the US at the time of the experiment. The majority of participants could be categorized as heritage speakers born and raised in the US (n = 83), while others were early bilinguals who immigrated before the age of eight (n = 9) or late bilinguals who immigrated after the age of eight (n = 18).5

To assess the linguistic backgrounds of the participants, the present study employed an adapted version of the Bilingual Language Profile questionnaire (BLP, Birdsong et al., 2012). Specifically, we excluded the language attitude module of the BLP, as it has been argued to be a less reliable indication of language dominance compared to other modules due to cultural differences in language attitudes (see Olson, 2023, for more discussion). Therefore, the questionnaire included three modules, including language history, use, and proficiency, yielding bilingual dominance scores ranging from –164 (English-dominant) to 164 (Spanish-dominant), with the score of zero representing balanced bilingualism. The BLP score is calculated by summing up the scores in each section with equal weights.

Participants’ dominance scores, as measured by the modified BLP, are shown in Figure 1. The mean score was –14.3 (SD = 34.9), suggesting that on average participants were slightly English-dominant. A Bayesian one-sample t-test, with the μ value of 0 indicating balanced bilingualism as the test value, showed that participants were highly unlikely to be balanced bilinguals (BF₁₀ = 421.46). Indeed, individuals varied considerably with the scores ranging from –95.9 (English-dominant) to 80.7 (Spanish-dominant).

Each section of the BLP showed response patterns consistent with the overall dominance scores. In the subsections of BLP, higher score indicates greater history, more frequent use, and higher proficiency in the corresponding language. The Language History section assessed participants’ Age of Acquisition (AoA), Length of Residence (LoR), comfortable language, education language, family language, and work language. History scores for Spanish (M = 34.5, SD = 7.4) and English (M = 36.3, SD = 8.9) were not notably different (BF₁₀ = 0.3), suggesting that the timing and contextual distribution of participants’ exposure were relatively balanced across the two languages, but with some degree of variability in each language. For instance, participants differed in how comfortable they felt about using Spanish and English, the years of Spanish or English classes they had taken, LoR, and whether or not they used English or Spanish at work. In contrast, participants were consistent in indicating Spanish as their L1 and their family language. The Language Use section evaluated the frequency of Spanish and English use in different contexts. Use scores for Spanish (M = 16.2, SD = 14.9) were noticeably lower than those for English (M = 26.9, SD = 18.8) (BF₁₀ = 7.13 × 10⁵), which was expected as all participants were residing in the US, although both languages showed substantial variability across participants. The Language Proficiency section measured self-reported proficiency in comprehension, speaking, reading, and writing. Proficiency scores were decisively higher for English (M = 51.4, SD = 5.0) than for Spanish (M = 44.8, SD = 8.2) (BF₁₀ = 1.79 × 10⁹), again with a high degree of individual difference. The BLP reports suggest that the participant pool in this study was diverse in terms of linguistic background, while indicating higher proficiency and use in English than in Spanish. Table 1 provides a summary of participants’ linguistic backgrounds.

Following and adapting the methodologies used in previous research, the study included a training phase and a test phase. The training phase was implemented via a cross-modal lexical decision task. During the training phase, participants listened to words, which were also shown on the screen one by one, in the appropriate orthography, and decided, on each trial, whether the word they heard was real or not, as quickly as possible. The goal of the lexical decision task was to induce phonetic recalibration by exposing participants to a sound which was acoustically ambiguous between [p] and [k]. Embedding these sounds into real words of English or Spanish biased participants to categorize them as certain phonemes despite their ambiguous acoustics.

The inclusion of orthographic forms in the training phase constitutes a deviation from the previous research tradition. It was implemented to enhance the recalibration effect by leveraging top-down influences on speech perception. Prior research has shown that priming can facilitate phonetic adaptation and word recognition (e.g., semantic priming, Meyer & Schvaneveldt, 1971; subtitles, Mitterer & McQueen, 2009; phoneme restoration effect, Warren, 1970). For instance, Keetels et al. (2016) showed that nonword orthographic information (e.g., showing “VdA” for priming /d/ responses in /a?a/ or “VbA” for inducing /b/) can also readjust a perceptual boundary between the two sounds. A recent study demonstrated that displaying the orthographic representations of stimuli containing an ambiguous sound facilitated a recalibration effect (see Caplan et al., 2021). In our study, the visual representation of the word forms served as an additional top-down cue, increasing the possibility that the ambiguous sound is perceived as a phoneme that completes a real word (e.g., /?/ is heard as /k/ in relocate but as /p/ in disrepair). While this approach differs from previous recalibration studies that used auditory presentation only, it aligns with broader findings on top-down processing in lexically guided speech perception (Ganong, 1980; Meyer & Schvaneveldt, 1971; Warren, 1970).

The lexical decision task included 18 critical stimuli with sounds intermediate between [p] and [k] in English and 18 analogous critical stimuli in Spanish. These ambiguous stimuli were designed to meet the following requirements in order to maximize the likelihood of recalibration (Caudrelier et al., 2023; Reinisch et al., 2013). First, none of the critical stimuli were English-Spanish cognates. Second, the ambiguous sound occurred in a stressed syllable as late in the word as possible (second or third syllable) (e.g., disappear). Third, we strived to avoid /p/ and /k/ and their voiced counterparts elsewhere in the critical stimuli and in fillers, although it proved difficult to exclude them completely. It was also impossible to exclude highly frequent voiceless and voiced alveolar stops /t/ and /d/. The lexical decision task also included 18 words with unambiguous contrasting sounds matching with the critical stimuli in length and frequency, 36 real-word fillers, and 72 nonce-word fillers generated by Wuggy (see Keuleers & Brysbaert, 2010, for the description of functions and instructions). Different L1 speakers of English and Spanish, both females in their 20’s, recorded stimuli in the corresponding languages. The speaker of Spanish was originally from Spain and had resided in the United States for two years and ten months for graduate studies at the time of recording. The speaker of English was born and raised in the Midwestern region of the United States and was also a graduate student. Recording was conducted in a sound-attenuated booth using a cardioid condenser microphone with a sampling rate of 44.1 kHz and a quantization rate of 16 bits.

To create a sound ambiguous between [p] and [k], a custom Python (version 3.10) algorithm for direct interpolation was employed, using NumPy and SciPy for numerical operations and audio processing, respectively. Specifically, the syllable-initial stop portions (burst and aspiration) in pairs of recordings of [p] and [k] words (e.g., disrepair and relocate) were blended into each other, creating 10 progressive steps with the first and the tenth steps being the original [p] and [k] and the intermediate steps progressing gradually from /p/ to /k/. The processed audio signals representing the stop portion were then added to the corresponding word frames, creating a [p] to [k] word continuum (i.e., disappear–disakkear).

Prior research showed that stop bursts contain information that can cue the following vowel (e.g., Liberman et al., 1967). Therefore, efforts were made to match vowels as closely as possible during the splicing procedure; however, perfect matching was not always possible (e.g., disrepair and relocate). While imperfect vowel matching could have introduced some unintended recalibration effects during training, a series of informal perceptual checks confirmed that the continua sounded natural, with listeners consistently perceiving a smooth transition from /p/ to /k/ without noticeable influence from vowel mismatches. If mismatched vowel cues had been the primary driver of stop perception, such a smooth transition would not have occurred. Instead, perception would have been disrupted, as the same burst can be heard as /p/ before /i/ but as /k/ before /a/ (Liberman et al., 1967).

The intermediate sound files (either the fifth or sixth on the continuum) were selected for exposure to an ambiguous sound during the lexical decision task. Since the continuum endpoints were natural recordings of /p/ and /k/ words rather than synthesized sounds, the midpoint was expected to reflect perceptual ambiguity between the two categories. While a separate norming test was not conducted, informal pilot tests demonstrated ambiguity at the intermediate steps.

The test phase consisted of a 2AFC task. The task involved 80 critical stimuli, resulting from the four 10-step continua based on /p/-/k/ minimal pairs in both English and Spanish, such as cat-pat, with critical segments in the word-initial position. The English minimal pairs were monosyllabic, and the Spanish minimal pairs were disyllabic with stress falling on the first syllable. The members of each minimal pair were matched in lexical frequency to avoid a frequency-based Ganong effect (Connine et al., 1993; Ganong, 1980). The intermediate steps in the continua were created by using the same method that was used to create ambiguous training stimuli for the lexical decision task. Minimal pairs were used to create the continua for the test phase, so the following vowels were always matched in the spliced stops. It should be noted that the end points of the continua used in the 2AFC task were natural productions of /p/ and /k/, which is different from previous studies where (re)synthesized stimuli were used as endpoints (e.g., Caudrelier et al., 2024; Reinisch et al., 2013). The minimal pairs used in the task are shown in Table 2. The full lists of training stimuli are provided in Appendix A.

All participants signed an electronic consent form and performed a headphone screening task before starting the experiment. The instruction language was always English in both the training and test tasks. Appendix B provides sample task images.

Returning to the representational assumptions of the three models of L2 speech learning, SLM-r posits that L1 and L2 speech sounds interact at the level of context-specific allophones, thus forming composite categories and affecting each other exclusively at this level. Given that our training stimuli were intersonorant, word-medial realizations of voiceless stops, place of articulation boundary recalibration induced for these sounds should be limited to this specific position across languages. Therefore, by extension, SLM-r does not predict generalization of recalibration both across languages and across positions (although generalization across positions within language is not precluded). Since our test words contained critical segments in the word-initial position, we would expect to see recalibration within language but not across languages. PAM-L2, in contrast, assumes that L2 sounds assimilate to L1 phonemes, not only to corresponding position-specific L1 allophones. Therefore, according to PAM-L2, cross-linguistic links could transcend the boundaries of specific allophones and extend to the phoneme as a whole (including other allophones). As a result, we would expect to see recalibration trained on word-medial stops in language 1 generalize to word-initial stops in language 2. Finally, L2LP, which assumes separate L1 and L2 grammars, would not predict cross-linguistic generalization of recalibration under any condition (assuming that such generalization indicates that L1 and L2 categories interact in a shared phonetic space).

The primary goal of the lexical decision task was to induce recalibration by exposing participants to ambiguous /p/ or /k/. Therefore, the present study does not focus on the examination of participants’ performance in the lexical decision task, although we provide the full statistical model results for the lexical decision task in Table 3. Descriptive statistical analysis, complemented with a Bayesian independent t-test, indicated that participants were more accurate in distinguishing real words from nonce words in English (percent correct = 90%) than in Spanish (percent correct = 85%) (BF₁₀ = 21.1). Figure 2 visually confirms this result with the distribution of English data being skewed towards higher scores than that of Spanish data. The results from the Bayesian mixed-effects logistic regression model also indicated that participants were more accurate in English than in Spanish (β = –0.47, 95% CI = [–1.05, 0.09], pd = 94.80), suggesting the possibility that they were somewhat more proficient in English than in Spanish.

The mean proportion correct for English critical words was 0.92 (SD = 0.27), while for Spanish critical words, it was 0.72 (SD = 0.45). Every English critical stimulus had an accuracy rate above chance (minimum = 0.65), whereas 10 out of 36 Spanish critical stimuli had a proportion correct of 0.50 or lower. This discrepancy suggests that some Spanish critical words may have been harder to recognize, presumably due to lower Spanish proficiency of our participants. At the participant level, Spanish word recognition accuracy (M = 0.72, SD = 0.41) was generally lower and more variable than in English (M = 0.92, SD = 0.21). Among the participants tested in Spanish, 10% had a mean accuracy below 0.50 on critical items. In contrast, the minimum accuracy in the English condition was 0.65. On the other hand, 91% of the participants in the English condition had an accuracy rate above 0.80, with many reaching perfect accuracy. This discrepancy suggests that for a subset of participants, Spanish critical words were not consistently recognized as real words, which may have influenced the strength of recalibration effects in the Spanish condition.

This section presents descriptive statistics from the 2AFC test phase. The mean proportions of /k/ responses were calculated across a 10-step continuum for each condition, defined by Test Language, Training Language, and Training Category (Table 4). In the Spanish test, both the English and Spanish training groups showed recalibration effects as participants exposed to /k/ tokens produced more /k/ responses than those exposed to /p/ tokens. The recalibration effect was stronger in the Spanish-trained group (M = .65 for /k/ training vs. M = .41 for /p/ training) than in the English-trained group (M = .58 for /k/ training vs. M = .53 for /p/ training). On the other hand, the recalibration effect manifested only in the test-training-matched condition in the English test. Specifically, while participants trained on English /k/ produced more /k/ responses than those trained on English /p/ (M = .49 for /k/ training vs. M = .40 for /p/ training), no meaningful difference was observed between Spanish /k/-/p/ training groups (M = .48 for /k/ training vs. M = .49 for /p/ training). These descriptive patterns are consistent with the model estimates reported in the next section.

A Bayesian mixed-effects logistic regression model was implemented for each test language. Each model examined the effects of Training Category, Step (centered), Training Language and their two- and three-way interactions on the log-odds of /k/ responses. Since Step is centered, the effect of either Training Category or Training Language assumes Step at zero, which is the midpoint of the continuum. That is, the main effects of the categorical factors indicate changes in the log-odds of /k/ responses when auditory stimuli are the most ambiguous. In the subsequent sections, particular focus is placed on the following three effects in examining recalibration: (1) Step, (2) Training Category × Step, and (3) Training Category × Step × Training Language. Tables 5 and 6 report all statistical results from each model.

As suggested by an anonymous reviewer, we conducted additional analyses to examine the potential effects of bilingual dominance (as measured by BLP) on phonetic recalibration. We fit two separate Bayesian linear regression models for the Spanish and English tests, using participants’ proportion of /k/ responses at the midpoint of the continuum (Steps 5 and 6), where recalibration was most prominent, as the dependent variable. The models included Training Category, Training Language, and BLP score, along with their interactions. The categorical factors were sum-coded. However, while both models indicated credible effects of Training Language and the interaction between Category and Training Language, as observed in prior models, none of the BLP or BLP interaction effects were found credible (pd < 95%).

Regarding the question of shared or separate L1 and L2 categories in bilinguals, previous work examining bilingual phonetic recalibration agrees that in cases of high phonetic similarity between L1 and L2 sounds, as exemplified by /s/ and /f/ across languages such as Dutch, German, English, and French, these sound categories could be shared across languages, given robust evidence of cross-linguistic transfer of recalibration (Caudrelier et al., 2023; Reinisch et al., 2013; Schuhmann, 2016). We modified several key methodological parameters in comparison to those in the previous work, which allowed us to formulate a related but novel research question: Does transfer of recalibration occur with sound categories that are phonetically different but phonologically related across languages, such as Spanish voiceless unaspirated and English voiceless aspirated stops, and in the absence of the possible facilitative effects of speaker-specific learning? Concomitantly, we tested generalization of recalibration across phonetic contexts (from word-medial to word-initial), which allowed us to probe the question of the level of representation at which cross-category links are established between languages. To answer these questions, our participants learned to retune the boundaries of the place of articulation contrast between word-medial voiceless stops /p/ and /k/ in English and Spanish through a cross-modal lexical decision task, and while half of them were tested on the same language and the same speaker in a 2AFC task, the second half were tested on a different language and a different speaker. The test presented critical segments in the word-initial position.

The main purpose of the lexical decision task was to induce phonetic recalibration between the /p/ and /k/ categories. Nonetheless, the statistical analysis of the data yielded findings that merit attention. Specifically, the accuracy results suggested that participants were somewhat more proficient in English than in Spanish, as accuracy was higher in English (90%) than in Spanish (85%), aligning with the BLP findings that participants were more English dominant. In addition, these accuracy rates were more consistent with advanced L2 speakers’ results and were somewhat lower than monolingual speakers’ accuracy (95%) reported in previous studies (Reinisch et al., 2013; Schuhmann, 2016).

The primary goal of the 2AFC task was to investigate the potential cross-linguistic transfer of phonetic recalibration, while also testing for generalization across contexts and, in some conditions, across speakers. The statistical analyses indicated both symmetric and asymmetric perceptual behaviors among the participants. First, perceptual categorization was strongly dependent on the steps of the acoustic /p/-/k/ continuum in all models, confirming that acoustic resynthesis aimed at creating progressive /p/-/k/ steps was effective. Second, perceptual recalibration of phonemic boundaries was robustly present when training language and test language (as well as speaker) matched, both for Spanish and English. This finding confirms that the training technique was effective overall. It also demonstrates successful generalization of recalibration from word-medial to word-initial contexts within languages. Additionally, this finding adds to a small body of research indicating that listeners need not be monolingual/L1 speakers of a given language for perceptual learning of this nature to take place (Caudrelier et al., 2023, 2024; Llompart, 2024; Reinisch et al., 2013; Schuhmann, 2016). Specifically, Spanish-English bilingual speakers in the current study who were mostly heritage speakers of Spanish and who differed from those investigated in previous research (diglossic French-English bilinguals in Montreal, Canada, or proficient L2 speakers of closely related languages: Dutch-English or German-English), recalibrated effectively in both of their languages. It should also be noted that in our study, when the training and test languages matched, the speaker also matched, which could have boosted perceptual learning in these conditions.

The two groups who were tested in the nonmatching condition between training and test languages demonstrated asymmetric perceptual patterns. Those who were trained on English and tested on Spanish showed evidence of recalibration generalization from English to Spanish. This finding suggests that bilinguals’ Spanish and English voiceless stops share a connection across languages. Importantly, in this case, recalibration generalized both across positions and across languages. This finding is consistent with the assumption of PAM-L2 that cross-linguistic links are not limited to the level of position-specific allophones, as postulated by SLM-r, but occur at the phonemic level. For recalibration to take place, participants had to generalize from word-medial allophones of English voiceless stops to word-initial allophones of Spanish voiceless stops, suggesting that these sounds are integrated with each other in some way, possibly as parts of a unified interlanguage phoneme. This finding is also at odds with the prediction derived from L2LP, according to our understanding of its assumptions, that cross-linguistic integration among sound categories cannot take place because the two grammars are separate and independent. Of course, these findings can support or contradict relevant theoretical assumptions only to the extent to which they can be interpreted as evidence of cross-linguistic links between L1 and L2 sounds.

Although the magnitude of the effect was decidedly smaller in the unmatched than in the matched conditions, it should be taken into account that the speaker was different between the English training and the Spanish test, possibly diminishing the effect. Recalibration of stop contrasts has also been shown to elicit weaker effects than recalibration of fricatives (Kraljic & Samuel, 2006). At the same time, matching the gender of two different speakers, and the use of stops instead of fricatives were associated with greater likelihood of cross-speaker generalization in previous work, in agreement with our findings (Kraljic & Samuel, 2005, 2006, 2007; Tamminga et al., 2020).

In contrast, the analysis revealed no evidence of recalibration transfer from Spanish to English in participants who were trained in Spanish but tested in English, suggesting that cross-linguistic generalization may not have taken place in this case. There are several possible interpretations or explanations for this asymmetry.

First, the results could be genuinely indicative of L1 and L2 sound categories that are linked across languages in an asymmetric manner, such that a more dominant language exerts a stronger influence on the less dominant one (Caramazza et al., 1973; Kartushina & Martin, 2019; Mack, 1989). In support of this hypothesis, in Caudrelier et al. (2024), cross-linguistic generalization of recalibration was stronger when exposure stimuli were in participants’ dominant language. On the other hand, our exploratory model of dominance failed to support this explanation, given the lack of credible effect for the BLP dominance score. However, since BLP is a self-reported, subjective measure, it may not fully assess language dominance. Without incorporating an external, objective measure of dominance, it remains premature to entirely rule out its potential role in the observed asymmetry in recalibration.

Another possibility relates to the fact that comparable phonetic realizations of stop consonants across English and Spanish do not map onto comparable phonological voicing categories. Specifically, English word-initial voiced stops are typically realized as voiceless unaspirated, similarly to Spanish word-initial voiceless stops (Lisker & Abramson, 1964). This may have compromised the perceptual link from Spanish voiceless unaspirated stops to English voiceless aspirated stops, explaining the asymmetric pattern.

Related to this scenario is the possibility that aspirated stops provide more salient acoustic cues to the place of articulation than unaspirated stops because they are more extended in time. The more salient acoustic nature of aspirated stops may lead to more successful recalibration, including its generalizability to untrained languages (see also Wright, 2001, 2004).9 This possibility needs to be addressed in future research using phonemic categories that do not present such an asymmetry in terms of the robustness of perceptual cues.

The instruction language may have also influenced recalibration in the present study by activating a specific language mode (Grosjean, 2001, p. 49). Since instructions were always given in English, regardless of the training or test language, it is plausible that participants’ English mode remained more strongly activated throughout the experiment, compared to Spanish, modulating recalibration effects.

Finally, it is possible that the Spanish stimuli were more effective at yielding or revealing recalibration than the English stimuli. As pointed out by an anonymous reviewer, we acknowledge that the imperfect match in CV syllables during the splicing procedure may have contributed to the observed asymmetry in recalibration. Since English voiceless stops have longer burst durations than Spanish ones, the residual vowel information in the burst may have been more pronounced in the English exposure stimuli. This could have introduced stronger coarticulatory cues, potentially influencing how ambiguous sounds were categorized and contributing to the asymmetry in recalibration patterns. Thus, a stimuli-related explanation cannot be entirely ruled out. However, it should be noted that the vowel information was unlikely to be the primary driver of perceptual patterns in the 2AFC task as this would have disrupted the observed smooth perceptual transition from /p/ to /k/ (Liberman et al., 1967).

Our findings speak to another issue, although it was not at the core of our investigation. An unresolved debate in the literature concerns the object of recalibration. That is, whether recalibration affects phonemes, phonological features, allophones, acoustics cues, or highly context-specific tokens. In other words, there remains a question regarding the abstractness versus specificity of the object of recalibration. The answer to this question is complicated by the fact that evidence supporting both sides of the spectrum has been reported in the literature (Dahan & Mead, 2010; Eisner & McQueen, 2005; Jesse & McQueen, 2011; McQueen et al., 2006; Mitterer et al., 2016; Reinisch et al., 2014). Generalization of recalibration across contexts and speakers suggests perceptual learning that abstracts away from the specifics of the training set. Thus, on the face of it, our findings appear to support the view that the object of recalibration is rather abstract, such as phoneme or a phonological feature. Our participants were trained to recalibrate the place of articulation based on word-medial, inter-sonorant stops, but tested on word-initial, prevocalic stops. The fact that recalibration occurred indicates that generalization over word positions and phonetic contexts took place. Furthermore, in the English training–Spanish testing condition, recalibration was also generalized to a phonetically different stop consonant (from voiceless aspirated to voiceless unaspirated), as well as to a different speaker and a different language. This suggests an abstract unit, such as phonemes, as the object of recalibration (assuming that Spanish and English voiceless stops belong to the same phoneme, which crosses language boundaries), or possibly a phonological feature of place. The latter predicts that training to recalibrate place on stop consonants should generalize to other segments with the same place features (e.g., nasals). Disputing this prediction, Reinisch et al. (2014) and Mitterer et al. (2016) did not observe generalization of recalibration from stops to nasals along the same place dimension. Moreover, a body of relevant research findings reviewed in Reinisch et al. (2014) suggests that generalization tends to be found in cases where acoustic cues are highly consistent between the training and the testing segments, despite changes in the context and speaker. In support of this view, Mitterer et al. (2016) found that recalibration generalized from tensified stops to highly acoustically similar plain stops in Korean, but not from tensified stops to more acoustically dissimilar aspirated stops or nasals. Kraljic and Samuel (2005) also suggested that patterns of acoustic similarity between male and female voices provide basis for cross-speaker generalization of perceptual learning. Therefore, acoustic similarity of cues to place across aspirated and unaspirated stops, across medial and initial stops, and across different speakers could explain the generalizability of recalibration in our data, without presupposing, but not precluding, a highly abstract object of recalibration, such as a phoneme or similarly abstract underlying unit.

While our findings contribute to this important debate, they are not sufficient to resolve it. In answering this question, further research with a greater variety of phonological contrasts demonstrating distinct patterns of cue variability across contexts and across speakers will be critical.

To conclude, recalibration manifested in both languages of the bilinguals in the training-test match condition and asymmetrically generalized to the untrained language in one of the unmatched conditions. This behavioral asymmetry could have several explanations, including the difference in language dominance, the nature of the phonological contrast tested, or stimuli-related reasons. Thus, our findings paint a complex picture of cross-linguistic organization of sound categories, partially supporting the possibility of category mergers across languages. Another recent report further complicates the situation. Caudrelier et al. (2024) demonstrated that, in addition to generalizing across languages, recalibration can proceed in two opposing directions in the two languages of bilinguals, provided exposure in those languages supplies appropriate perceptual evidence. That is, bilinguals can learn to expand an /s/ category in French while shrinking /s/ in English, suggesting that each category is sufficiently autonomous. Bilinguals’ ability to make such divergent language-specific adjustments may contradict the idea that the manipulated categories are merged across languages. If so, it compromises the validity of cross-linguistic generalization of recalibration as evidence of such mergers, since both divergent and congruent cross-linguistic recalibration was observed for the same group of bilinguals in Caudrelier et al. (2024). Alternatively, it is possible that the merged L1-L2 categories can establish a degree of mutual autonomy, which nevertheless raises the question of what would constitute proof of true and complete cross-linguistic independence between related sound categories, if such true independence is possible. To further illuminate this complex question, more research needs to be conducted with diverse bilingual populations and with a greater variety of phonological categories. Nevertheless, our findings provide the first evidence that indicates that transfer of phonetic recalibration across the two languages of bilinguals is possible even in the absence of the facilitative effect of speaker-specific learning and with phonetically distinct but phonologically related categories. The fact that phonetic learning generalized across speakers, languages, and phonetic differences suggests some degree of integration, however asymmetric, between L1 and L2 sound categories (e.g., Spanish and English /p/), possibly along the lines of underlying phonological representations or unifying acoustic characteristics.