Highlights

• • Phonological and semantic similarities both sped up responses.

• • Phonological similarity was dependent on language early in the time course.

• • Semantic similarity facilitated processing similarly in L1 and L2.

• • Cognate word frequency slowed responses late in time in L1 and L2.

• • L2 proficiency sped responses but did not modulate cross-linguistic influence.

Quick and efficient word recognition is central to effective communication. For bi- and multilinguals, this requires recognizing an even greater number of words than for monolinguals. A plethora of evidence has established that the words a person knows are stored together in a single mental lexicon. Words are processed together, non-selectively, and the knowledge of another language can affect how words in just one are understood, even when it is not needed to complete the task. This cross-linguistic influence (hereafter, CLI) can help (e.g., speed) or hinder (e.g., slow) recognition depending on the requirements of the task. Here, CLI refers to the observable consequences of non-selective lexical activation during word recognition – that is, the influence of the non-target language on processing even when only the target language is required. Importantly, CLI is influenced by factors such as the degree of overlap across languages, the context in which it occurs, a bilingual’s proficiency, the language tested, and the point in time at which it is measured. While CLI has been observed in both visual and auditory word recognition, far more evidence comes from the visual domain, leaving several unanswered questions about how CLI affects spoken word processing.

The present study investigates patterns of CLI in spoken word recognition in a seemingly unlikely context. Specifically, unbalanced, different-script (i.e., Japanese–English) bilinguals living in a primarily Japanese-language environment are presented with unprimed words varying in degree of cross-linguistic overlap, and the time course of activation is examined. Individuals are tested in both L1 and L2 to make a direct comparison across languages and to determine whether CLI influences recognition in L1 as it typically does in L2.

1. Background

In bilingual processing, one of the most frequently examined word types when looking for CLI is cognates, or words that share form and meaning, such as restaurant and the Spanish restaurante. Bilinguals usually read cognates faster in sentences (Cop et al., Reference Cop, Dirix, van Assche and Drieghe2017; Duyck et al., Reference Duyck, van Assche, Drieghe and Hartsuiker2007; Libben & Titone, Reference Libben and Titone2009; Schwartz & Kroll, Reference Schwartz and Kroll2006), respond to them faster in lexical decision (Dijkstra et al., Reference Dijkstra, Grainger and van Heuven1999; Dijkstra et al., Reference Dijkstra, Miwa, Brummelhuis, Sappelli and Baayen2010; Lemhöfer & Dijkstra, Reference Lemhöfer and Dijkstra2004; Poort & Rodd, Reference Poort and Rodd2017), and speak them faster in picture and word naming (e.g., Costa et al., Reference Costa, Caramazza and Sebastian-Galles2000; Schwartz et al., Reference Schwartz, Kroll and Diaz2007) compared to non-cognate translation equivalents. The cognate effect is taken as evidence of the non-selectivity of the mental lexicon. If bilinguals are faster or more accurate at recognizing cognates compared to non-cognates, then knowledge of the other language is understood to contribute to the recognition process.

Importantly, however, these effects are not always symmetrical across languages. The cognate effect and other forms of CLI are more robust in L2 compared to L1 (Jared & Kroll, Reference Jared and Kroll2001; Jared & Szucs, Reference Jared and Szucs2002; Van Assche et al., Reference van Assche, Brysbaert, Duyck, Heredia and Cieślicka2019; Weber & Cutler, Reference Weber and Cutler2004), and the size of effects is often larger in the L1-L2 direction as well (Costa et al., Reference Costa, Caramazza and Sebastian-Galles2000; Gollan et al., Reference Gollan, Forster and Frost1997; Jiang, Reference Jiang2019; Poarch & Van Hell, Reference Poarch and van Hell2012). It is important to note that a major contributor to this so-called asymmetry may be that bilinguals tested are most often unbalanced (Van Hell & Tanner, Reference van Hell and Tanner2012). In other words, it is not the L1/L2 status itself, but the bilinguals’ higher proficiency in L1 that is responsible for this difference. While many earlier studies simply divided bilinguals into low- and high-proficient groups or simply compared monolinguals with bilinguals (e.g., Dimitropoulou et al., Reference Dimitropoulou, Duñabeitia and Carreiras2011; Weber & Cutler, Reference Weber and Cutler2004), more recent investigations have included proficiency as a continuous predictor variable (e.g., Bultena et al., Reference Bultena, Dijkstra and van Hell2014; Miwa et al., Reference Miwa, Dijkstra, Bolger and Baayen2014). Doing so may provide a clearer picture of CLI – including any present in L1 – than grouping participants into overly simple categories.

This does not mean that CLI does not occur in L1; assuming sufficient L2 proficiency, it can be observed in L1 as well (e.g., Kroll & Stewart, Reference Kroll and Stewart1994; Van Hell & Dijkstra, Reference van Hell and Dijkstra2002). Indeed, such effects have been observed in both visual (Costa et al., Reference Costa, Caramazza and Sebastian-Galles2000; Poarch & Van Hell, Reference Poarch and van Hell2012; van Hell & Dijkstra, Reference van Hell and Dijkstra2002; Van Wijnendaele & Brysbaert, Reference van Wijnendaele and Brysbaert2002) and auditory experiments (Dijkstra et al., Reference Dijkstra, Timmermans and Schriefers2000; Lagrou et al., Reference Lagrou, Hartsuiker and Duyck2011; Mishra & Singh, Reference Mishra and Singh2014, Reference Mishra and Singh2016). The L2 is typically examined alone, without regard to L1 performance, but the bilingual mental lexicon is a dynamic system. Instead of writing off such effects as context-specific or non-existent, Van Assche et al. (Reference van Assche, Brysbaert, Duyck, Heredia and Cieślicka2019) point out that researchers may in fact observe valid CLI in L1 if they conduct “careful and powerful” experiments.

An important factor in CLI is the degree of overlap between languages. Cognates that are more similar show evidence of greater co-activation, for example (Dijkstra et al., Reference Dijkstra, Grainger and van Heuven1999, Reference Dijkstra, Miwa, Brummelhuis, Sappelli and Baayen2010; Duyck et al., Reference Duyck, van Assche, Drieghe and Hartsuiker2007; Lemhöfer et al., Reference Lemhöfer, Dijkstra and Michel2004; Lemhöfer & Dijkstra, Reference Lemhöfer and Dijkstra2004; Van Assche et al., Reference van Assche, Duyck, Hartsuiker and Diependaele2009, Reference van Assche, Drieghe, Duyck, Welvaert and Hartsuiker2011). In contrast to earlier studies simply comparing cognates with homophones, Dijkstra et al. (Reference Dijkstra, Grainger and van Heuven1999) tested Dutch-English bilinguals in both lexical decision and progressive demasking tasks, systematically varying orthographic, phonological, and semantic similarities across items. They found different effects of phonological and semantic similarities across tasks, showing that CLI depends on the joint effects of these three factors. Further, cognate effects tend to be additive; three-language cognates, for instance, are processed even faster than those shared by just two, although the degree of facilitation can vary across tasks and lexical characteristics (Foryś-Nogala et al., Reference Foryś-Nogala, Silva, Ambroziak and Otwinowska-Kasztelanic2025; Lemhöfer et al., Reference Lemhöfer, Dijkstra and Michel2004; Otwinowska, Reference Otwinowska2024; Van Hell & Dijkstra, Reference van Hell and Dijkstra2002). While most evidence comes from relatively similar language pairs, these factors raise the question of how such effects may differ in languages that do not share orthographic script, such as Spanish and Korean or Japanese and English. In these languages, cognates are relatively dissimilar as they share only meaning and phonological form. Japanese and English, for example, differ across several typological dimensions relevant to lexical processing. English is largely isolating, with relatively fixed SVO word order and limited inflectional morphology, whereas Japanese is agglutinative, head-final (i.e., SOV), and employs extensive suffixation and overt topic marking. The two languages also differ substantially in phonological structure (e.g., syllable structure constraints, moraic timing in Japanese) and use entirely distinct writing systems (alphabetic vs. kana/kanji). As a result, Japanese–English cognates share no orthographic overlap and comparatively little phonological overlap relative to most same-script language pairs, providing a stringent test case for models that assume non-selective lexical activation.

Beyond structural overlap, the presence, magnitude, and direction of CLI can also differ depending on the point in time during processing at which it is measured, an idea which has been observed in both reading (Miwa et al., Reference Miwa, Dijkstra, Bolger and Baayen2014; Taylor & Mukai, Reference Taylor and Mukai2023) and listening (Marian et al., Reference Marian, Blumenfeld and Boukrina2008; Weber & Cutler, Reference Weber and Cutler2004). Cross-linguistic phonological similarity, for example, facilitated recognition early in time but slowed it later on in a visual study by Miwa et al. (Reference Miwa, Dijkstra, Bolger and Baayen2014), echoing findings from spoken word recognition (Marian et al., Reference Marian, Blumenfeld and Boukrina2008). An examination of not just the overall outcome but also the time course of processing is important to understand how activation occurs in the mind. Masked priming (Forster & Davis, Reference Forster and Davis1984) has been a popular method for examining the time course of single word reading, and interestingly, eye movements have more recently been employed for this purpose as well (Kuperman et al., Reference Kuperman, Schreuder, Bertram and Baayen2009; Miwa et al., Reference Miwa, Dijkstra, Bolger and Baayen2014; Taylor & Mukai, Reference Taylor and Mukai2023). With spoken words, examining the time course can be more challenging. Priming and the visual world paradigm have both been used to do so, but findings from these methods must be interpreted with caution (Forster, Reference Forster1998; Huettig et al., Reference Huettig, Rommers and Meyer2011; Zhang et al., Reference Zhang, Fairchild and Li2017). As both involve presentation of a non-target competitor or other information just before or during processing, more naturalistic data are needed to verify whether CLI is present even when competitors are not purposefully activated.

Context also plays an important role in how words are recognized. Words embedded in sentences (Binder & Rayner, Reference Binder and Rayner1998; Chambers & Cooke, Reference Chambers and Cooke2009; Duyck et al., Reference Duyck, van Assche, Drieghe and Hartsuiker2007; Lagrou et al., Reference Lagrou, Hartsuiker and Duyck2013) and texts (Cop et al., Reference Cop, Dirix, van Assche and Drieghe2017) are processed under the influence of semantic, syntactic, and discourse constraints, which shape lexical activation and further facilitate or inhibit recognition compared to isolated words. Global language context (Elston-Güttler et al., Reference Elston-Güttler, Gunter and Kotz2005), stimulus list composition (Brenders et al., Reference Brenders, Van Hell and Dijkstra2011), and the social environment in which an individual lives (e.g., Schwartz & Kroll, Reference Schwartz and Kroll2006) can all also modulate levels of non-target activation, often reducing, but importantly, not eliminating its effects. The present study examines CLI under minimal contextual constraint, using isolated spoken words without priming or sentence context.

2. The need for integrated models of bilingual word recognition

Taken together, the above findings have motivated formal models of word recognition. Overall, there is strong evidence that the mind stores knowledge of and processes different languages together, a concept deeply integrated into popular bilingual word recognition models such as BIMOLA (Grosjean, Reference Grosjean, de Groot and Kroll1997; Léwy & Grosjean, Reference Léwy, Grosjean and Grosjean2008), BLINCS (Shook & Marian, Reference Shook and Marian2013), BIA+ (Dijkstra & Van Heuven, Reference Dijkstra and van Heuven2002), and the more recent Multilink (Dijkstra et al., Reference Dijkstra, Wahl, Buytenhuijs, van Halem, Al-Jibouri, De Korte and Rekké2019). None of these models, however, explicitly predicts recognition in both reading and listening. This is despite considerable overlap in not only the type of information processed (Ferrand et al., Reference Ferrand, Méot, Spinelli, New, Pallier, Bonin, Dufau, Mathôt and Grainger2018; Müller et al., Reference Müller, ten Bosch and Ernestus2024; Tanenhaus et al., Reference Tanenhaus, Flanigan and Seidenberg1980) but also the brain areas used in understanding the two modalities (Chee et al., Reference Chee, O’Craven, Bergida, Rosen and Savoy1999).

The BIA+ has been widely used to understand visual word recognition and can also be readily applied to auditory processing. According to this account, non-target representations, including those from another language known to an individual, can become co-activated when they share information with a word being processed. Specifically, overlapping orthographic, phonological, and semantic representations lead to co-activation, and the more information is shared, the more becomes co-activated. While the model was developed on evidence from same-script languages (such as Dutch and English), it has successfully predicted visual word recognition with different-script languages (e.g., Miwa et al., Reference Miwa, Dijkstra, Bolger and Baayen2014; Taylor & Mukai, Reference Taylor and Mukai2023), a key population for psycholinguistic models (Mishra, Reference Mishra2019). See Figure 1 for a representation of this model for bilinguals in this study.

Figure 1.

Visualization of the BIA+ Model for auditory word recognition. Note. Adapted for auditory recognition from Taylor and Mukai’s (Reference Taylor and Mukai2023) representation of the BIA+ model (Dijkstra & Van Heuven, Reference Dijkstra and van Heuven2002) for Japanese–English bilinguals.

Figure 1. Long description

The flowchart is organized into three vertical levels and two horizontal columns.

At the bottom center, an Auditory signal represented by a waveform points upward to the Sublexical phonology box. This box contains phonetic symbols in circles, such as forward slash r forward slash, forward slash o forward slash, and forward slash p forward slash. To its left is the Sublexical orthography box, containing Japanese katakana characters and English letters in circles. A double-headed horizontal arrow connects these two sublexical boxes.

The middle level contains Lexical orthography on the left and Lexical phonology on the right, connected by a double-headed horizontal arrow. The Lexical orthography box contains the Japanese word for program and the English word program. The Lexical phonology box contains the phonetic transcriptions forward slash puroguramu forward slash and forward slash p r o u g r ae m forward slash. Vertical double-headed arrows connect these lexical boxes to their respective sublexical boxes below.

At the top, Semantic information is on the left, containing various shaded circles. On the right is Language membership, containing circles labeled L 1 and L 2.

Complex interactions are shown via double-headed arrows:

- Lexical orthography points to Semantic information and Language membership.

- Lexical phonology points to Semantic information and Language membership.

- Semantic information and Language membership are also interconnected with the lexical layers through diagonal crossing arrows.

The BIA+ easily accounts for the discrepancies in CLI mentioned above. For example, in the case of unbalanced bilinguals, the BIA+ model’s temporal delay assumption supposes that the dominant or more frequently used language (i.e., L1) becomes activated more quickly than L2 due to a higher resting activation level. In other words, it takes less input to reach its activation threshold. This difference can explain the asymmetry observed in cross-linguistic effects between L1 and L2. Additive information leads to increased co-activation, resulting in greater CLI for more similar words or languages. Discrepancies in CLI at different processing stages can also be explained by sequential activation of representations in this model: orthographic information becomes activated first, followed by phonological and semantic representations. As a result, orthographic or phonological overlap may drive early co-activation, while semantic similarity may contribute to later effects. The model also includes separate recognition and task/decision subsystems, so although initial activation based on input is bottom-up, a bilingual can adjust their response based on task requirements, accounting for variations across tasks and contexts.

3. Cross-linguistic influence in spoken word recognition

Mirroring visual word recognition, the primary factors affecting spoken word processing are related to the target stimulus itself. Longer words and those with many phonological neighbors typically take longer to recognize (Goh et al., Reference Goh, Yap and Chee2020; Luce & Pisoni, Reference Luce and Pisoni1998; Ziegler et al., Reference Ziegler, Muneaux and Grainger2003), as do less frequent words (Goh et al., Reference Goh, Yap, Lau, Ng and Tan2016). As in reading, evidence for language non-selective activation is pervasive (Blumenfeld & Marian, Reference Blumenfeld and Marian2007; Lagrou et al., Reference Lagrou, Hartsuiker and Duyck2011; Marian et al., Reference Marian, Blumenfeld and Boukrina2008; Marian & Spivey, Reference Marian and Spivey2003; Weber & Cutler, Reference Weber and Cutler2004). While orthographic co-activation cannot directly occur at the earliest stages of auditory word recognition, parallel activation can occur via shared phonological or semantic information, or even as a result of activation spreading to the orthographic representations from phonological input (see, for example, Qu & Damian, Reference Qu and Damian2017; Türk & Domahs, Reference Türk and Domahs2022; Veivo et al., Reference Veivo, Järvikivi, Porretta and Hyönä2016).

Phonological overlap effects have been observed repeatedly in the auditory modality. For instance, using the visual world paradigm, Weber and Cutler (Reference Weber and Cutler2004) observed that Dutch-English bilinguals listening to English words looked longer at pictures whose Dutch names contained sounds similar to those in the English words they heard, indicating cross-linguistic phonological activation. This effect did not occur, however, when the task was conducted in Dutch (L1). By contrast, Spivey and Marian (Reference Spivey and Marian1999) found similar effects in the L1. Instructing Russian-English bilinguals in their L1 to pick up an item (e.g., Podnimi marku (“Pick up the stamp”)), they noticed that bilinguals looked more often to distractor pictures whose English (L2) name began with a similar sound (e.g., marker) than to unrelated distractors. With Hindi-English bilinguals, Mishra and Singh (Reference Mishra and Singh2014, Reference Mishra and Singh2016) observed similar phonological co-activation in both L1 and L2, with bilinguals looking more often towards phonological cohort members, that is, words that share initial phonemes with the target. These studies provide evidence that non-target phonological activation can affect listening.

In contrast to phonological overlap, the role of semantic information is less often investigated in spoken word recognition. Various semantic influences have been observed with monolinguals, however (e.g., Goh et al., Reference Goh, Yap, Lau, Ng and Tan2016; Wurm et al., Reference Wurm, Vakoch and Seaman2004), and again the effects vary by task (Goh et al., Reference Goh, Yap, Lau, Ng and Tan2016). Bilingual-specific semantic influences are not typically investigated with behavioral studies of listening, but an fMRI experiment by Chee et al. (Reference Chee, O’Craven, Bergida, Rosen and Savoy1999) observed similar brain areas active during semantic processing in both modalities. Therefore, it is very possible that semantic effects found in bilingual reading (Gollan et al., Reference Gollan, Forster and Frost1997; Kim & Davis, Reference Kim and Davis2003; Miwa et al., Reference Miwa, Dijkstra, Bolger and Baayen2014; Nakayama et al., Reference Nakayama, Sears, Hino and Lupker2012; Taylor & Mukai, Reference Taylor and Mukai2023) also affect listening beyond commonly examined word-level variables.

Unprimed L2 word reading studies by Dijkstra et al. (Reference Dijkstra, Miwa, Brummelhuis, Sappelli and Baayen2010) and Miwa et al. (Reference Miwa, Dijkstra, Bolger and Baayen2014) found that semantic similarity influenced recognition above and beyond phonological influences. Taylor and Mukai (Reference Taylor and Mukai2023) extended these findings to L1, finding similar effects in unbalanced Japanese–English (i.e., different-script) bilinguals in Japan, a context and population where CLI may be unexpected given the infrequent use of the L2 and relative distance between the languages. Whether such effects arise in listening, as they do in reading, for this population remains unclear. While the BIA+ suggests it is possible, most auditory studies focus primarily on phonological similarity, classifying meaning as a categorical variable (e.g., congruent vs. incongruent). Phonological and semantic overlap may be somewhat confounded in studies that do not model these two continuous variables separately. If, as the BIA+ model suggests, similar mechanisms underlie visual and auditory word processing, both types of overlap need to be examined separately in spoken word studies.

Echoing behavioral findings, ERP studies (e.g., Bobb et al., Reference Bobb, Von Holzen, Mayor, Mani and Carreiras2020) have, in fact, found L2 semantic influences in L1 word recognition, albeit with same-script bilinguals. It is possible that in the present unlikely context – unbalanced, different-script bilinguals living in an L1 environment and tested in the L1 without using priming or visual world cues, CLI could occur in listening just as it did in the reading study by Taylor and Mukai.

4. Objectives of the current study

The present study examines spoken word recognition in both L1 and L2 with Japanese–English bilinguals, without using priming or visual world methods. It extends Taylor and Mukai’s (Reference Taylor and Mukai2023) findings of bidirectional cross-linguistic effects in visual word recognition to the auditory domain, testing the applicability of the BIA+ with this population. To do so, it addresses the following four questions:

First, do phonological similarity, semantic similarity, and cognate frequency influence spoken word recognition in both languages? As BIA+ supposes that the same general mechanisms take place in reading and listening, the same non-target information (i.e., phonological representations, semantic representations) is expected to affect processing in this auditory study. Indeed, same-script language studies by Lagrou et al. (Reference Lagrou, Hartsuiker and Duyck2011) and Marian et al. (Reference Marian, Blumenfeld and Boukrina2008) found this to be the case. Lagrou et al. (Reference Lagrou, Hartsuiker and Duyck2015) and Garrido-Pozú (Reference Garrido-Pozú2024) even found effects in L1. However, Japanese and English share less overlap than the languages they studied, so co-activation may be weaker or delayed, especially if participants tested are L1-dominant and use the L2 only infrequently. The temporal delay of L2, combined with relatively few shared phonological representations, may limit cross-language activation, reaching the lower limit of non-selectivity. However, because spoken input directly activates phonological representations, CLI may still emerge as non-target phonological information becomes co-activated early on.

Second, if such effects occur, are their size and direction (i.e., if responses are facilitated or slowed) similar in each language? Taylor and Mukai (Reference Taylor and Mukai2023) found weaker effects in L1; this is expected to be the case in this auditory study as well. Taylor and Mukai also found that while semantic similarity and cognate frequency speeded responses in both languages, phonological similarity slowed recognition slightly in L2 but speeded it in L1. Cognate frequency effects, in contrast, did not differ across languages. Listeners, in contrast, may rely more on phonological information than do readers, for whom semantic representations can be directly activated from orthography according to BIA+. If this is true, phonological similarity may be more influential in the present study, or the direction and timing of effects may be different. In English auditory lexical decision tasks (ALDTs) with Spanish–English and English–Spanish bilinguals, for example, Garrido-Pozú (Reference Garrido-Pozú2024) found that phonological similarity speeded L2 responses but slowed them in L1; the same may be true for Japanese–English bilinguals.

Third, do these effects depend on L2 proficiency? Taylor and Mukai (Reference Taylor and Mukai2023) observed greater semantic similarity effects for less proficient readers, but no interaction of proficiency with phonological similarity or cognate frequency. In listening, less-balanced bilinguals may rely more on phonological overlap, consistent with Garrido-Pozú’s (Reference Garrido-Pozú2024) results for L2 with Spanish–English bilinguals.

Finally, does the timing of measurement affect observed patterns of CLI? The time course of auditory recognition is challenging to measure. Brand et al. (Reference Brand, Mulder, ten Bosch and Boves2021) suggest that response times measured from stimulus onset (hereafter, RTonset) reflect different parts of the activation process than those measured from offset (hereafter, RToffset). If a predictor behaves differently in RTonset and RToffset models, its influence likely changes over time. Indeed, stimulus duration, a strong predictor of response times (Ernestus & Cutler, Reference Ernestus and Cutler2015; Ferrand et al., Reference Ferrand, Méot, Spinelli, New, Pallier, Bonin, Dufau, Mathôt and Grainger2018; Tucker et al., Reference Tucker, Brenner, Danielson, Kelley, Nenadic and Sims2019), was found to speed RToffset but slow RTonset in their study. A three-way interaction with language (i.e., L1 versus L2) showed that L1 listeners especially benefitted from information heard during the stimulus itself. CLIs were not considered, however, so it is unclear how they might change over time. In Taylor and Mukai (Reference Taylor and Mukai2023), neither phonological nor semantic similarity was influential early in time, although effects arose later during processing. If this is true in listening, effects may be more pronounced in RToffset models. Still, as orthographic activation is not required for phonological co-activation in listening, phonological and semantic similarities may influence processing from early on. If so, they may be similarly influential in RTonset and RToffset models.

7. General discussion

The intent of this study was to determine whether CLIs similarly affect L1 and L2 auditory lexical decisions, and whether any such effects change by L2 proficiency or by the point in the time course in which activation is measured. The BIA+ model assumes representations from both languages are co-activated during word recognition, such that knowledge of one language can influence processing even when only one is in use. In Japanese and English, which use completely different writing systems and have dissimilar phonological systems, co-activation can occur, but only directly through shared semantic or phonological information (which, in the case of the latter, is often less than for languages such as English and Dutch, on which the BIA+ model was based). As Japanese/English cognates share comparatively less information, and as there may be a large temporal delay of activation for L2 for unbalanced bilinguals living in the present L1 context (i.e., Japan), CLI may be precluded for these items, especially if non-target competitors are not purposefully activated through priming or related methods.

Despite this unlikely context, the present study provides evidence for CLI on spoken word recognition in both Japanese and English. Separate models fitted to RTonset and RToffset revealed that similar task, lexical, and bilingual-specific predictors affect spoken word recognition in the two languages. Further, while English proficiency modulated responses, the overall processes (including the presence of CLI) were similar for high- and low-proficient L2 listeners.

The first major aim of this investigation was to determine whether phonological similarity, semantic similarity, and cognate word frequency affect auditory lexical decisions as they do visual lexical decisions. Shared phonological and semantic information facilitated responses in both L1 and L2, regardless of whether they were measured from stimulus onset or offset. As in Taylor and Mukai (Reference Taylor and Mukai2023), individual phonological similarity ratings were more predictive of performance than group-averaged ratings, suggesting that subjective, item-level similarity judgments may better reflect cross-linguistic phonological activation patterns. By contrast, semantic similarity effects were better captured by the averaged ratings, indicating that meaning-based overlap exerted a consistent, facilitatory influence across participants.

Cognate word frequency (NonTargetWF) was notably also a significant predictor, as in the visual studies by Miwa et al. (Reference Miwa, Dijkstra, Bolger and Baayen2014) and Taylor and Mukai (Reference Taylor and Mukai2023). In those reading studies, the effect emerged early in processing, whereas in the present auditory study, the effect emerged only for responses measured from stimulus offset. This suggests that cognate frequency effects influence auditory word recognition later than visual word recognition, where such effects have been observed from the earliest stages of processing (e.g., first fixations). One possible explanation is that auditory input unfolds incrementally over time, delaying the point at which lexical competition between the target and cross-language competitors becomes fully engaged. In contrast, the full orthographic form is presented immediately during reading, allowing frequency-based competition to influence processing earlier.

The direction of cognate frequency effects also differed from the previous visual studies. While Miwa et al. (Reference Miwa, Dijkstra, Bolger and Baayen2014) observed facilitatory effects in L2, and Taylor and Mukai (Reference Taylor and Mukai2023) observed facilitation in both L1 and L2, the present auditory study observed later-arising inhibitory effects of cognate frequency in both languages, particularly in L2 English. Taken together, these findings suggest that while cross-linguistic activation is evident in both reading and listening, the timing and consequences of cognate frequency differ across modalities.

The second question was whether the direction and size of cross-linguistic effects were similar in the two languages. In contrast to the Taylor and Mukai (Reference Taylor and Mukai2023) study, in which phonological overlap slowed L2 but speeded L1 responses, the present auditory investigation found both semantic and phonological overlap speeded responses in both languages. Because spoken words directly activate phonological representations, it is possible that the relatively early activation of L1 candidates during L2 listening allows enough time for competition to be resolved before a response is made. In reading, by contrast, activation first proceeds via orthography, adding another temporal step. Quick activation of L1 phonological information in reading could mean L1 competitors reach higher activation levels, delaying recognition of L2 targets. In auditory processing, where phonological information is available immediately and unfolds over time, L1 competitors may both arise and be resolved rapidly, leading to the consistently facilitatory effects observed here.

The phonological similarity effect here was dependent on language for the RTonset analysis, in that while it was facilitatory in both languages, it especially speeded responses in L2, echoing findings from L2 English and Spanish (Garrido-Pozú, Reference Garrido-Pozú2024) and supporting the asymmetry predicted by Kroll and Stewart’s (Reference Kroll and Stewart1994) Revised Hierarchical Model. The effect did not differ by language after offset, revealing that this discrepancy is an earlier effect largely resolved during the course of the stimulus itself, which is unsurprising as, according to BIA+, phonological representations would become co-activated early enough to be largely resolved before higher-level processes take place, so any language-specific advantage (or disadvantage) would have been worked out by the time participants plan a response.

Notably, these results are in line with Taylor and Mukai (Reference Taylor and Mukai2023); in that study, language-dependent phonological similarity effects arose at late fixation and flattened out somewhat by response.

The effect of semantic similarity, by contrast, did not differ by language. It did, however, depend on language at RToffset, but only until by-participant random effects were included in the model. This suggests that shared semantic information becomes co-activated, helping auditory processing in a similar way in L1 and L2. For both similarity measures, the magnitude of effects is larger in the RToffset model (b = −0.0127 at onset and b = −0.0283 at offset for PhonologicalRating; b = −0.0112 at onset and b = −0.0412 at offset for meanSemanticRating), suggesting that while similarity effects emerge during the encoding stages of processing, they continue to build and exert a greater impact as recognition unfolds. Taken together, these results make sense from the perspective of the BIA+ in which co-activation is not precluded in either language direction in either visual or auditory word recognition.

The third goal was to check whether these effects depended on L2 proficiency. While participants with higher LexTALE scores were faster overall in both English and Japanese, the effect was especially pronounced for the highest proficiency listeners, who responded faster to L2 than to L1 targets, whereas the lowest proficiency participants demonstrated the opposite pattern. Within the framework of the BIA+ model, increased L2 experience likely raises the resting activation level of L2 lexical representations and strengthens cross-language connections, thus speeding lexical processing in both languages. Importantly, however, proficiency did not modulate the size of phonological or semantic similarity effects, suggesting that cross-linguistic co-activation operated similarly across the proficiency range of bilinguals tested here. This pattern mirrors previous findings in reading and listening, where proficiency influenced general processing speed but not the magnitude of cross-linguistic facilitation (Garrido-Pozú, Reference Garrido-Pozú2024; Miwa et al., Reference Miwa, Dijkstra, Bolger and Baayen2014; Taylor & Mukai, Reference Taylor and Mukai2023).

Although the present findings demonstrate CLI across a range of L2 proficiencies, the results may not generalize to all bilingual populations. Participants here were primarily L1-dominant, Japanese–English bilinguals residing in an L1 context. Highly balanced bilinguals, bilinguals immersed in an L2-dominant environment, or individuals with substantially earlier and more sustained L2 exposure may exhibit different patterns of co-activation or competition dynamics. In immersion contexts, for example, increased resting activation levels of L2 representations could reduce the asymmetries predicted by the Revised Hierarchical Model or alter the timing of cross-linguistic effects. Future research examining bilinguals across dominance profiles, language-use environments, and modality-specific proficiency measures (e.g., listening comprehension) would help elucidate the lower limits of these auditory cross-linguistic effects.

A final goal was to determine whether cross-linguistic effects differ over time, examining the time course of auditory processing to the extent possible without using priming, the visual world paradigm, or neuroimaging techniques. When RTonset and RToffset analyses are compared, the same effects overall affected processing in both analyses, showing it is possible, at least to some extent, to compare findings from studies that measured response times from stimulus onset (e.g., Ernestus & Cutler, Reference Ernestus and Cutler2015; Ferrand et al., Reference Ferrand, Méot, Spinelli, New, Pallier, Bonin, Dufau, Mathôt and Grainger2018; Tucker et al., Reference Tucker, Brenner, Danielson, Kelley, Nenadic and Sims2019) with findings from those measuring from offset (e.g., Munson et al., Reference Munson, Swenson and Manthei2005; Taler et al., Reference Taler, Aaron, Steinmetz and Pisoni2010).

A few key differences were observed in this study, however. First, the language dependence of phonological similarity at RTonset suggests phonological representations play a larger role early in the process, a finding which is consistent with the predictions of the BIA+ as well as the results of Taylor and Mukai (Reference Taylor and Mukai2023). Further, the marginally significant Language:PreviousRToffset interaction (p = .06) in the RToffset analysis suggests that participants’ later cognitive processes and response patterns may be affected by local speed effects more in Japanese than in English. This may be due in part to differences in post-lexical processes in this task; participants may have greater automaticity in decision making in the dominant L1, compared to the less-dominant L2 in which they might be more cautious before pressing a response button.

Importantly, these findings extend and build upon bidirectional cross-linguistic effects found in isolated word reading with different-script bilinguals (Degani et al., Reference Degani, Prior and Hajajra2018; Kim & Davis, Reference Kim and Davis2003; Miwa et al., Reference Miwa, Dijkstra, Bolger and Baayen2014; Peleg et al., Reference Peleg, Degani, Raziq and Taha2020; Taylor & Mukai, Reference Taylor and Mukai2023; Voga & Grainger, Reference Voga and Grainger2007; Zhou et al., Reference Zhou, Chen, Yang and Dunlap2010) as well as spoken word experiments in which priming was not used (e.g., Garrido-Pozú, Reference Garrido-Pozú2024; Wu et al., Reference Wu, Chen, van Heuven and Schiller2019), corroborating evidence from priming and visual world studies (e.g., Gollan et al., Reference Gollan, Forster and Frost1997; Mishra & Singh, Reference Mishra and Singh2014; Nakayama et al., Reference Nakayama, Sears, Hino and Lupker2012). This study provides strong support for both the predictions of the BIA+ model and its extendibility to the auditory modality. There are substantial differences between reading and listening. In reading, orthography is first activated, while in listening, the process begins directly from sublexical phonology; in reading, all information typically becomes available at once, while in listening, information unfolds serially over a period of time; if the font is held standard, words look the same from one instance to the next during reading, while a spoken signal can vary considerably from instance to instance. Despite these discrepancies, the present findings indicate that the same basic properties laid out by BIA+, including CLI, importantly, affect bilingual activation in the two modalities.