Distributional Learning and Language Activation: Evidence from L3 Spanish Perception Among L1 Korean–L2 English Speakers

Abstract

This study investigates L3 Spanish perception patterns among L1 Korean–L2 English bilinguals with varying L3 proficiency levels, aiming to test the applicability of traditional L2 perceptual models in multilingual contexts. We conducted two experiments: a cross-linguistic discrimination task and a cross-language identification task. Results revealed unexpected outcomes unique to multilingual contexts. Participants had difficulty reliably discriminating between cross-linguistic categories and showed little improvement over time. Similarly, they did not demonstrate progress in categorizing sounds specific to each language. The absence of a clear correlation between proficiency levels and the ability to discriminate and categorize sounds suggests that input distribution and language-specific activation may play more critical roles in L3 perception, consistent with the distributional learning approach. We argue that phoneme distributions from all three languages likely occupy a shared perceptual space. When a specific language is activated, the relevant phoneme distributions become dominant, while others are suppressed. This selective activation, while not crucial in traditional L1 and L2 studies, is critical in L3 contexts, like the one examined here, where managing multiple phonemic systems complicates discrimination and categorization. These findings underscore the need for theoretical adjustments in multilingual phonetic acquisition models and highlight the complexities of language processing in multilingual settings.

1. Introduction

Multilingualism has become increasingly common, leading to a growing interest in understanding how additional languages are acquired, particularly in L3 contexts. Phonological and phonetic research initially focused on first (L1) and second (L2) languages, operating on the assumption that adult language learning is fundamentally distinct from childhood acquisition and that any language learned in adulthood is treated as an L2. However, growing evidence indicates that additional languages beyond L2 are not equivalent, as each new language introduces unique and complex interactions within multilingual cognitive systems, resulting in diverse developmental patterns and outcomes (De Angelis, 2007; Falk & Bardel, 2010; Flynn et al., 2004).

Current models of phonetic and phonological perception—including Flege’s Speech Learning Model (SLM-r, Flege & Bohn, 2021), Best’s Perceptual Assimilation Model (PAM-L2, Best & Tyler, 2007), and Escudero’s Second Language Linguistic Perception model (L2LP, Escudero, 2005)—support the notion that learning mechanisms persist throughout life, allowing for the acquisition of additional languages. Consequently, L3 studies often adapt tools from models designed for L1 or L2 contexts, with L3 phonetics and phonology exemplifying this trend (e.g., Liu & Lin, 2021; Sypiańska, 2016). Yet, each language pairing presents a distinct context, raising the possibility that adapting models across learning contexts may have unexpected effects.

In terms of theoretical grounding, the three prominent L2 perceptual models—SLM-r, PAM-L2, and L2LP—provide valuable insights for our study, and we adopt their shared assumption that L1 serves as a foundation for L2 acquisition, where learners’ L1 categories shape their perception of L2 sounds, often causing learners to interpret them through the lens of their L1 phonological categories. This cross-linguistic influence can either facilitate or hinder L2 processing depending on the degree of similarity or divergence between the L1 and L2 sounds.

The three models converge on the idea that language learning is driven by learners’ perception of mismatches between the sounds of their L1 and the target language. Specifically, progress is thought to occur when learners detect these differences. While matched sounds require no adjustment, mismatched sounds prompt learners to create new categories, adjust existing category boundaries, or alter how they assimilate sounds across languages. Difficulties in perceiving these differences are often cited as the root of challenges in acquiring the sounds and contrasts of the target language. While effective discrimination is widely recognized as essential, it can be argued that language learning does not necessarily rely on the explicit comparison of sounds across languages. In this study, we argue that learners generally cannot consciously compare sounds between languages. Instead, we propose that the core mechanism is one of distributional learning, aligning with the recent revision of the Speech Learning Model (SLM-r).

Distributional learning (Jurafsky & Martin, 2022; Kerns et al., 1994; Peperkamp et al., 2006; Saffran et al., 1996) posits that learners do not actively contrast or manipulate categories. Instead, each exposure to speech sounds leaves a memory trace that accumulates according to a normal distribution. In this framework, learners lack the agency to compare and identify differences, focusing instead on processing phonemes, which then settle into distributions forming the basis of phonetic categories. This automatic process is thought to generate an internal distribution that mirrors the variation in and distribution of the language encountered during learning.

Beyond learning mechanisms, other aspects make the SLM-r the most suitable model for L3 application. First is the developmental scope: the PAM, which is designed for novice listeners, is well-suited to initial L2 perception stages, which limits its relevance to our focus on the development of phonetic categories. In contrast, the SLM-r and L2LP are applicable across the full span of L2 development, with L2LP offering specific learning trajectories based on varied contexts. However, L2LP’s focus on phonological contrasts and its computational modeling via Stochastic Optimality Theory does not align well with our phonetic emphasis. A second key advantage of the SLM-r is its emphasis on phonetic categories rather than phonological contrasts. Unlike the PAM-L2 and L2LP, which prioritize phonological contrasts based on lexical representations, the SLM-r concentrates on phonetic category formation, making it particularly relevant to our study of phonetic category acquisition within an L3 context. In this study, we operationalize phonetic categories as mental representations of speech sounds characterized by distributions of acoustic cues (here, VOT and F0 are considered for stop consonants).

While the acquisition of phonological contrasts largely depends on lexical access, phonetic categories emerge from the processing of phonetic cues. The concept of distributional learning (Escudero & Boersma, 2004; Maye et al., 2002; McClelland & Elman, 1986) suggests that processing these cues generates a set of traces that are categorized according to their frequency in the input. From this perspective, phonetic categories can be seen as Gaussian distributions, that is, continuous, symmetric, bell-shaped normal distributions, formed by the accumulation of tokens based on their frequency. These emergent categories are then utilized to process incoming linguistic input. Given the focus on the development of phonetic categories in an L3 context, the SLM-r model serves as the most suitable framework for the present study.

Regarding the two experiments applied in this study, discrimination tasks represent a widely used method for assessing learners’ ability to perceive phonetic contrasts. By controlling lexical, prosodic, phonological, and phonetic variables, these tasks provide valuable insights into perceptual sensitivity. However, the high level of control inherent in these tasks may reduce their ecological validity, as real-world contexts frequently involve the very variables that are excluded in such experimental settings, which can actually assist listeners in perceiving the contrasts of interest. This limitation is one reason why discrimination tasks are typically combined with identification tasks.

Identification tasks provide insight into how learners perceive and categorize L2 sounds. However, performance in these tasks can be influenced by lexical representation rather than solely by perceptual sensitivity. The reliance on top-down preexisting knowledge can obscure genuine phonetic perception. To mitigate this issue, the present study employed pseudo-words in the identification task. Despite their challenges, discrimination and identification tasks remain essential tools for exploring L2 sound acquisition. Moreover, based on the assumption that language learning mechanisms are consistent across L1, L2, L3, and Ln, there should be no inherent issues in applying them to an L3 context.

In summary, this study extends established phonetic acquisition models and methodological tools into an L3 context, building on the assumption that fundamental language learning mechanisms are consistent across L1, L2, and L3. By focusing on phonetic categories rather than phonological contrasts, we utilize the Speech Learning Model-revised (SLM-r) to examine how multilingual learners form and adapt sound categories when managing multiple linguistic systems.

2. Review of the Literature

A common observation in third language acquisition (L3) research is that what makes this context unique—and so intriguing—is the additional complexity of managing more than two languages. The presence of an extra language introduces new dimensions and amplifies interaction possibilities between linguistic systems. When three linguistic systems are involved, tasks that typically yield consistent results in L1 or L2 contexts may start producing unpredictable outcomes. Previous studies on L3 acquisition have consistently shown that managing multiple linguistic systems introduces complexities beyond what is observed in L2 acquisition (Cabrelli Amaro & Wrembel, 2016; Cenoz et al., 2001; De Angelis, 2007; Flynn et al., 2004; Falk & Bardel, 2011; Rothman et al., 2013; Sánchez, 2020; Wrembel et al., 2020).

Several models have been tailored to the L3/Ln context, but they primarily address cross-linguistic influence (CLI) and emphasize morphosyntactic and lexical interactions (e.g., Bardel & Falk, 2007, 2012; Bardel & Sánchez, 2017; Falk & Bardel, 2010, 2011; Flynn et al., 2004; Berkes & Flynn, 2012; Rothman, 2010, 2011, 2013; Rothman et al., 2019; Westergaard, 2019; Westergaard et al., 2017). Recent efforts to adapt L2 speech learning models, such as the SLM (e.g., Liu & Lin, 2021; Sypiańska, 2016) and PAM-L2 (e.g., Nelson, 2020; Wrembel et al., 2019) to the L3 context are emerging, though a specific model addressing phonetic category formation and sound processing across multiple languages has not yet been proposed. Also, most research into L3 phonology has been carried out on the initial stages of L3 development (e.g., Balas et al., 2019; Kędzierska et al., 2023; Liu & Lin, 2021; Nelson, 2020; Onishi, 2016; Parrish, 2022; Wrembel et al., 2019; Wrembel et al., 2020), while the present study’s interest lies in how L3 learners’ ability to discriminate cross-linguistic pairs and categorize L3 targets evolves with increased proficiency.

Previous L2 research consistently shows a link between proficiency and performance on perception tasks. Several studies confirm that higher L2 proficiency leads to better discrimination and identification of non-native speech contrasts across various language pairs (Flege et al., 1997; Bradlow & Bent, 2008; Díaz et al., 2012). This relationship applies to both vowel and consonant contrasts. Experienced learners also demonstrate more categorical perception (Sebastián-Gallés & Díaz, 2012; Guion et al., 2000). However, this relationship between perception and proficiency is nuanced and often influenced by task type (Díaz et al., 2012) or the specific sounds involved (Darcy et al., 2016).

The rest of this section reviews the literature on the main phonetic cue in this study: voice onset time (VOT), which is a measure for stop consonants across languages that has been widely analyzed since Lisker and Abramson’s study (1964). An intriguing combination for analysis involves the triad of Altaic, Germanic, and Romance languages, represented in this study by Korean, English, and Spanish—three typologically distinct languages. This triad is particularly compelling because English and Spanish use different phonetic cues to distinguish /p/ and /b/ (pre-voicing in Spanish and aspiration in English), while Korean employs a unique three-way contrast among /p/, /p*/, and /p^h/, using mainly F0 and aspiration as distinguishing features. Unlike English, which relies heavily on VOT as a primary cue (Kingston & Diehl, 1994), Korean speakers tend to depend on F0. This results in a noticeable influence when Korean learners of English transfer F0 cues to their L2 perception (e.g., M. Kim et al., 2002; Schertz et al., 2015a), while English learners of Korean primarily rely on VOT to perceive Korean stops (Schertz et al., 2015b).

Keating (1984) and Kingston and Diehl (1994) categorize languages into “voicing” or “aspirating” groups based on their stop contrasts. Voicing languages like Spanish differentiate between voicing lead and short-lag stops, whereas aspirating languages such as English contrast short-lag stops with long-lag stops. English is typically classified as an aspirating language, featuring long-lag VOT for word-initial stops; however, it also makes use of all three VOT distinctions—voicing lead, short-lag, and long-lag—depending on the phonological context (Hualde, 2005). English speakers produce varying stop types in initial, medial, and final word positions, with notable dialectal and individual variation. Aspiration characterizes English /p, t, k/ in initial positions, making these voiceless stops aspirated, while /b, d, g/ remain unaspirated. Pre-voicing of /b, d, g/ is possible, but its frequency is generally considered low (Beckman et al., 2011; Davidson, 2016), at least in comparison to its presence in Spanish. This pre-voicing is variable and may be influenced by factors like stress, neighboring sounds, phonological context, and speaker variation (Davidson, 2016; Hualde, 2005; Lisker & Abramson, 1964; Winn, 2020). English word-initial /p, t, k/ exhibit positive VOTs ranging from 55 to 100 ms (Caramazza et al., 1973; Lisker & Abramson, 1964; Riney & Takagi, 1999), while /b, d, g/ have short-lag VOTs, typically under 30 ms. Voicing affects both the F0 of the following vowel and the duration of the preceding one, with F0 being higher after voiceless stops and lower after voiced stops (House & Fairbanks, 1953; Peterson & Lehiste, 1961). F0 serves as a secondary cue for distinguishing voiceless stops, although it does not differentiate between short-lag and voicing lead stops (Dmitrieva et al., 2015).

Korean typologically presents a three-way contrast in stops: aspirated, fortis, and lenis (Cho et al., 2002; Lee et al., 2013; Shin, 2011; Song et al., 2018). Studies traditionally use VOT to define these stop types (Lisker & Abramson, 1964; C. Kim, 1965; Han & Weitzman, 1970; Hardcastle, 1973; Silva, 2006), with fortis stops having the shortest VOT, aspirated stops the longest, and lenis stops in between. However, younger speakers in Seoul have blurred the distinction between lenis and aspirated stops, either by raising lenis VOT values or by lowering those of aspirated stops (Jang, 2011; Y. Kang, 2014; K. Kang & Guion, 2008; Oh, 2011; Silva, 2006). Consequently, VOT alone no longer captures the full contrast among Korean stops, making F0 a primary cue instead. Aspirated stops generally show the highest F0 values, followed by fortis stops, with lenis stops having the lowest (Jang, 2011). Accordingly, this study focuses on speakers of the Seoul–Gyeonggi dialect.

Spanish has a two-way voicing contrast, with /p, t, k/ produced as short-lag stops and /b, d, g/ with lead VOT (Rosner et al., 2000). This can lead to confusion, as Spanish short-lag stops /p, t, k/ align with English short-lag stops /b, d, g/ (Morales-Front, 2018). In Spanish, initial voiceless stops after a pause range from 0 to 30 ms for /p, t/ and 0 to 40 ms for /k/, while voiced stops are realized with negative VOT. F0 functions as a secondary cue, differentiating lead VOT stops from short-lag ones, with F0 higher after short-lag and lower after lead VOT stops (Dmitrieva et al., 2015).

The distribution of labial stops in each language is illustrated in Figure 1, which captures the VOT and F0 values across English, Korean, and Spanish. Detailed VOT and F0 values are provided in

Table 1. Descriptive statistics of production values of female participants in the trilingual baseline group.

	VOT (ms)		F0 (Hz)
	M	95%CI	M	95%CI
Kpp	8.85	7.97/9.73	246.71	242.90/250.52
Kph	86.43	81.76/91.09	269.80	266.40/273.20
Kp	88.18	81.44/94.91	181.34	175.88/186.80
Ep	93.72	85.36/102.09	218.19	213.33/223.05
Eb	12.25	10.54/13.96	184.56	179.57/189.56
Sp	9.97	8.99/10.96	206.49	199.06/213.92
Sbn	−80.78	−87.22/−74.34	184.03	178.40/189.66

.

The present study aims to examine how multilinguals discriminate and assimilate cross-linguistic categories in labial stops and how this ability changes with increased L3 proficiency. With the configurations of English, Korean, and Spanish stops in mind, this study seeks to answer the following research questions:

Research Question 1 (RQ1). 

Can multilinguals discriminate between cross-linguistic categories, and does this ability improve with increased L3 proficiency?

Hypothesis 1 (H1). 

Advanced trilingual learners will distinguish cross-linguistic categories, consistent with the SLM-r’s postulate P2, which states that bilinguals strive to maintain phonetic contrast between all consonants and vowels in their common phonological space.

Research Question 2 (RQ2). 

Do the patterns of assimilating L3 sounds to L1 or L2 categories change with proficiency?

Hypothesis 2 (H2). 

According to the SLM-r’s principle of Dynamic Category Formation, which encompasses the perception of phonetic differences between L1 and L2 (H1), the potential for establishing new phonetic categories (H2), the role of perceived phonetic distance (H3), and the influence of experience (H6), assimilation patterns are expected to shift as proficiency increases. Initially, L3 sounds may be mapped onto existing L1 and L2 categories; however, as proficiency improves, learners should increasingly perceive subtle phonetic differences, resulting in different cross-language mapping patterns.

3. Materials and Methods

3.1. Participants

A total of 59 L3 learners of Spanish participated in the present study, encompassing a broad range of proficiency levels. To isolate the impact of L3 proficiency from that of L1 and L2 proficiency, this study specifically targeted learners at similar L1 and L2 proficiency levels, specifically learners at native or near-native L1 and L2 competence and at varying levels of L3 proficiency. This focus allowed for a clearer examination of L3 proficiency effects. Participants were either college graduates or current undergraduate or graduate students and did not have any speech or hearing impairments.

After applying the inclusion criteria (see

Table 2. Inclusion criteria for L3 learner participants.

	Inclusion Criteria
Age	Between 18 and 40 Years of Age
L1 Korean	- The Seoul–Gyeonggi dialect - Native or near-native level EIT (0–120): between 90 and 120 Self-assessed language skills (1–7): between 5 and 7
L2 English	- Acquired English in the U.S. or Canada after age 2 - Native or near-native level EIT (min: 0, max: 120): between 90 and 120 Self-assessed language skills (min: 1, max: 7): between 5 and 7
L3 Spanish	_ Proficiency levels vary (EIT: between 1 and 90) _ ≤2 years living or studying abroad
Additional language (L4)	- Started learning an additional language (i.e., L4) after learning the L3—Spanish - Weaker competence than the L3 - No experience living in countries in which the L4 is an official or majority language - No daily use of the L4

), 19 participants were excluded from the data analysis, bringing the final sample size down to 40 (24 females and 16 males). Six of the original participants were removed from the sample based on the scores of the language assessment (Elicited Imitation Task—EIT; see Section 3.2.3) on their L1 and L2 (below 90 on a scale from 0 to 120) and self-assessments of L1 and L2 skills below 5 (on a scale from 1 to 7), far from native or near-native levels of Korean and/or English. Eleven learners were not part of the final sample because they had acquired an additional language besides Korean, English, and Spanish at an early age or at a higher proficiency level than their Spanish. Finally, two participants were eliminated for being substantially older than the rest of the sample and for not completing the perception tasks, respectively.

The remaining sample (n = 40) consists of sequential multilinguals with L1 Korean, L2 English, and L3 Spanish. All participants acquired Korean, specifically the Seoul–Gyeonggi dialect, as their first language, which is both their home language and the official language of their home country. At the time of this study, they spent a minimum of half an hour each day communicating in Korean with family members and friends. Each participant began meaningfully learning L2 English upon arriving in the United States as young migrants before subsequently taking up Spanish as their L3, and most used English in their college and work environments on weekdays.

Proficiency levels were assessed based on participants’ performance in the Elicited Imitation Task (see Figure 2) and were further validated through informal interviews, self-assessment scores (see

Table 3. Participant background information.

(N = 40)
		Undergrad	MA	PhD/MD/JD
Education level		23	10	7
			Mean (SD)	95% CI (lower/upper)
Age			28.18 (5.242)	26.50	29.85
Age of acquisition	Eng		9.85 (4.704)	8.35	11.35
Age of acquisition	Sp		17.43 (5.168)	15.77	19.08
Self-assessed language skills (min: 1, max: 7)	Kor	listening	6.95 (0.316)	6.85	7.05
		speaking	6.90 (0.379)	6.78	7.02
		reading	6.88 (0.404)	6.75	7.00
		writing	6.65 (0.770)	6.40	6.90
	Eng	listening	6.38 (0.838)	6.11	6.64
		speaking	6.28 (0.816)	6.01	6.54
		reading	6.33 (0.888)	6.04	6.61
		writing	6.30 (0.966)	5.99	6.61
	Sp	listening	2.53 (1.154)	2.16	2.89
		speaking	2.15 (1.099)	1.80	2.50
		reading	2.60 (1.411)	2.15	3.05
		writing	1.90 (1.105)	1.55	2.25

), and a language background questionnaire (See Appendix G). Participants were recruited via social network services and personal connections, including Korean associations and religious institutions affiliated with one of the authors. As a token of appreciation for their time, participants received a small stipend.

3.2. Tasks

Participants completed a discrimination (oddity) task and an identification task which included goodness-of-fit ratings. We chose the oddity task over other options such as the ABX discrimination task because it relaxes cognitive load and reduces bias. The oddity task does not implicate memory, and it circumvents one of the most pervasive limitations of ABX tasks—the fact that results are subject to order effects.

3.2.1. Discrimination (Oddity) Task

To assess how L3 learners distinguish L3 target sounds from corresponding categories in their L1 and L2, an oddity task was developed based on studies by Flege (2003) and Nagle (2021). In this task, participants listened to three items and identified the one that differed from the others. To control for lexical and contextual effects and to ensure precise timing between stimuli, the items were designed as monosyllabic pseudo-words consisting of a labial stop followed by a low central vowel.

Five pair combinations were selected based on VOT or fundamental frequency (F0) values, as well as the similar contexts described in the SLM-r and findings from a pilot study:

• L3 /p/-L2 /p/: Matching at the phonological level but differing phonetically, with positive VOT as the key cue.
• L3 /p/-L2 /b/: Different at both the phonological and phonetic levels, with no reliable cue since neither VOT nor F0 shows sufficient differentiation.
• L3 /p/-L1 fortis /p*/: Different at both levels, with F0 serving as the key phonetic cue.
• L3 /b/-L2 /b/: Matching phonologically but differing phonetically, where negative VOT is the key cue.
• L3 /b/-L1 lenis /p/: Different at both levels, with both positive and negative VOT serving as the key cues.

The stimuli were produced by two native speakers of Seoul Korean (one male and one female), both of whom had near-native proficiency in English and Spanish. Their proficiency was verified through a language background questionnaire, informal interviews, and self-assessment. These speakers, who used Korean at home, were exposed to English and Spanish from an early age through migration and formal schooling. Each speaker read a target item containing the sequence {<p, b>} + <a> within a carrier sentence in all three languages (see

Table 4. Carrier sentences used for the generation of the oddity task and identification task stimuli.

Language	Carrier Sentence
Spanish	Digo, ‘ba’ porque sí (×5)
	Digo, ‘pa’ porque sí (×5)
English	I say, ‘ba’ just because (×5)
	I say, ‘pa’ just because (×5)
Korean	그냥, ‘바’ 라고 말합니다 (×5) /kunyang, ‘pa’ lako malhapnita/ “Just because, ‘pa’ I say” which means “I say, ‘ba’ just because”
	그냥, ‘빠’ 라고 말합니다 (×5) /kunyang, ‘ppa’ lako malhapnita/ “Just because, ‘p*a’ I say” which means “I say, ‘p*a’ just because”

), producing each target five times. The individual one-syllable words were extracted from the complete sentences for use as stimuli. The recorded words were normalized to 70 dB. Each contrast included 12 trials: 6 per speaker (3 same, 3 different), which were used to evaluate participants’ ability to discriminate phonetically relevant variations between categories (Flege, 2003).

Participants completed a total of 60 trials (5 contrasts × 12 trials). The position of the odd item in different trials was counterbalanced following Nagle (2021) (see Appendix A for all items). The stimuli were prepared using Adobe Premiere Pro, and the task was administered through Qualtrics. Participants had ten seconds to identify the item that was different from the others by pressing “1”, “2”, “3”, or “모든 소리가 같음” (/motun solika kathum/means “every sound is the same”), and the interstimulus interval was 1.3 s.

3.2.2. Identification Task with Rating

Two versions of the identification task were developed—one for Spanish–English and another for Spanish–Korean—both incorporating goodness-of-fit scales to evaluate how L3 learners perceptually categorize L3 targets within the frameworks of their previously acquired languages. To minimize lexical effects, pseudo-words were used in this task as well. However, unlike the discrimination task, interstimulus timing was not a factor, allowing the pseudo-words to consist of two syllables and be presented within carrier sentences.

All stimuli comprised pseudo-words (see Appendix E) pronounced within the same Spanish carrier sentence (Table 4). These Spanish-pronounced pseudo-words were utilized to test identification from Spanish to English and Spanish to Korean (see Appendix C and Appendix D). For example, the Spanish-pronounced pseudo-word “pafe” was presented to participants as if it were a Korean-pronounced pseudo-word extracted from a Korean carrier sentence. In another round, the same pseudo-word was presented as if it were extracted from an English carrier sentence. To control for prosodic and assimilatory effects, all target phonemes were positioned after a pause, away from the nuclear accent, at the beginning of a disyllabic word, and before a low central vowel.

In total, there were 20 target words: 10 for /p/ and 10 for /b/ (see Appendix E). For this study, a female trilingual speaker with native proficiency in Spanish, English, and Korean recorded the pseudo-word stimuli within a carrier sentence in each of the three languages.

For each language, participants were informed that they would listen to sentences containing a made-up word pronounced by a native speaker of that language. During the task, they were asked to identify the target Spanish phonemes from a set of labial stop phoneme options—three options for Korean and two for English. Following this identification, participants rated the degree of match between the heard phoneme and the selected phoneme on a 5-point Likert scale. Specifically, they were asked to evaluate whether the labial stop they heard was a good representation of the labial stop they had just identified.

3.2.3. Elicited Imitation Task

Each EIT includes 30 sentences which range between 7 and 19 syllables. The sentences are presented aurally to test takers. Then, approximately 2.5 s later, test takers are instructed to repeat the sentence they have just heard verbatim. In this study, the EIT evaluated participants’ proficiency in Korean, Spanish, and English, following the methodology established by Ortega et al. (2002). More specifically, the Korean version of the EIT (Y. Kim et al., 2016) was re-recorded by a male speaker of the Seoul dialect to address dialectal discrepancies present in the original recording. Similarly, the second versions of the English (Wu et al., 2022) and Spanish (Faretta-Stutenberg & Morgan-Short, 2018) EITs were re-recorded by male speakers to resolve issues of noise, item clarity, and speed found in the original recordings of the second versions. These revised versions were implemented to avoid semantic overlap with the tasks of the first versions. Sound editing for all recordings was conducted using Adobe Premiere Pro.

The EIT is useful for an L3 perceptual study because all test takers have to depend on their aural and oral linguistic abilities during the test. Additionally, the EIT lets researchers easily compare learners’ proficiency in each of the languages (Wu et al., 2022). In particular, the English and Spanish EITs have two parallel forms, which makes it possible to use the three language versions of the EIT without possible semantic confounds. Moreover, the EIT requires only around 10 min to be implemented per each version, while standardized proficiency tests can be administered approximately in an hour or longer (Wu et al., 2022).

The participants’ L1 and L2 proficiency levels were “very high” or “native speakers”, while their L3 proficiency level ranged from “low” to “intermediate”, according to the comparable DELE (the Diploma of Spanish as a Foreign Language, a widely used Spanish proficiency test) and EIT scores by experience level reported in Solon et al. (2019).

3.3. Procedure

This study is part of a larger research project that also included production data. Data collection involved three sessions, lasting approximately two hours in total, with two 10 min breaks in between. In the first session, participants signed an informed consent form and then completed two production tasks: a picture-naming task and a wordlist read-aloud task. Following this, they participated in the Elicited Imitation Task in each language, with the order of the EITs randomized. After a 10 min break, the second session consisted of several individual differences (ID) tasks. After another 10 min break, the third session began with the discrimination and identification tasks. Finally, participants filled out a language background questionnaire after a brief informal interview. All communication occurred in Korean, the participants’ L1, and the researcher was able to see the participants’ screens at all times. While the production and ID tasks were administered during the experiment, they are not included in the present study.

3.4. Data Analysis

3.4.1. Discrimination Task

Cross-language discrimination ability, as assessed by the oddity task, was operationalized using A-prime (A′), an unbiased measure of signal detection sensitivity (Flege, 2003; Flege et al., 1999). A total of 2400 responses were analyzed to calculate A-prime, derived from {5 pairs × 12 trials) × 40 participants}. A-prime incorporates both hit rates and false alarm rates. Hit rates represent the proportion of correct identifications of the odd item among the triplets, while false alarms refer to the proportion of incorrect identifications within the same triplets. A score of 1.0 indicates perfect discrimination ability, whereas a score of 0.5 signifies no discrimination ability (Flege et al., 1999). The formulas used to calculate A′ are as follows:

H: Hit rate, FA: False alarm

A′ score of 1.0 = perfect discrimination sensitivity

A′ score of 0.5 = lack of discrimination sensitivity

a. If H > FA, then A′ = 0.5 + [(H-FA)(1+H-FA)]/[4H(1-FA)]

b. If H = FA, then A′ = 0.5

c. If H < FA, then A′ = 0.5 + [(FA-H)(1+FA-H)]/[4FA(1-H)]

(Guion et al., 2000, p. 2723)

First, we tested the assumptions for inferential statistics. Given the presence of tied observations and the different scales of the independent and dependent variables, we used Stuart-Kendall’s tau-c for all correlation analyses in this section, as it is more suitable for such data (for more details on Kendall’s tau-c, see Berry et al., 2009). Stuart-Kendall’s tau-c correlations were performed at a significance level of p < 0.05.

3.4.2. Identification Task

For the identification tasks, a total of 2400 responses (20 identification questions × 3 language versions × 40 participants) were included in the data analysis. Participants’ identification of the target sounds as belonging to existing categories of their L1 and L2 was measured in two ways: by calculating the percentage of correct categorizations and by determining the mean goodness-of-fit rating for each consonant target.

4. Results

Initially, we considered including each individual observation in the dataset and analyzing the data using mixed-effect models given their suitability for handling repeated measures. However, the data did not meet the assumptions required for the application of mixed-effect models.

4.1. Discrimination Task

The results of non-parametric Stuart-Kendall’s tau-c correlations revealed no significant relationship between L3 proficiency and discrimination accuracy across all language pairs (see

Table 5. Results of Stuart-Kendall’s tau-c correlations between L3 proficiency and discrimination ability. K/p*/: Korean fortis; K/p/: Korean lenis.

Pairs	τc	Sig
S/p/—E/p/	−0.052	0.719
S/p/—E/b/	−0.051	0.653
S/p/—K/p*/	−0.062	0.526
S/b/—E/b/	−0.177	0.092
S/b/—K/p/	0.092	0.446
All pairs	−0.097	0.409

).

Consequently, additional analyses were conducted to identify factors beyond L3 proficiency that might significantly influence discrimination accuracy. To achieve this, the discrimination accuracy of the cross-language contrast pairs was examined irrespective of L3 proficiency. Below in Figure 3, we present a linear plot of the oddity task by cross-language contrast pairs for all levels of L3 proficiency (Mean A′ scores (SD): S/p/—E/p/, 0.946 (0.68); S/p/—E/b/, 0.640 (0.212); S/p/—K/p*/, 0.717 (0.146); S/b/—E/b/, 0.654 (0.182); S/b/—K/p/, 0.908 (0.843)).

The descriptive results indicated that the accuracy scores for the two pairs, S/p/-E/p/ and S/b/-K/p/, were notably higher than those of the other three pairs: S/p/-E/b/, S/p/-K/p*/, and S/b/-E/b/. To further investigate these differences, a Friedman’s ANOVA, a non-parametric repeated measures analysis, was conducted, treating the contrast type as the independent variable and A′ scores as the dependent variable. The analysis revealed a statistically significant difference in discrimination ability based on the cross-language pairs (χ²(2) = 100.010, p < 0.01).

To determine the specific pairs that differed from one another, a post-hoc analysis was performed using the Wilcoxon test with a Bonferroni correction applied. The results of this analysis indicated no significant difference in discrimination ability between the S/p/-E/p/ and S/b/-K/p/ pairs, nor between the S/p/-E/b/, S/p/-K/p*/, and S/b/-E/b/ pairs. However, the S/p/-E/p/ and S/b/-K/p/ pairs were found to be significantly different from the other three pairs, aligning with the descriptive statistics that indicated higher accuracy scores for S/p/-E/p/ and S/b/-K/p/.

4.2. Identification Task with Rating

The Stuart-Kendall’s tau-c correlations indicated that there was no effect of L3 proficiency on the percentage of identifications or the goodness-of-fit ratings (see

Table 6. Results of Stuart-Kendall’s tau-c correlations between L3 proficiency and identification in Korean mode. K/p^h/: Korean aspirated; K/p*/: Korean fortis; K/p/: Korean lenis.

Stimulus	Identified As		τc	Sig
S/p/	K/ph/	%	−0.163	0.123
		goodness	0.079	0.768
S/p/	K/p/	%	0.167	0.165
		goodness	−0.171	0.346
S/p/	K/p*/	%	0.032	0.773
		goodness	−0.146	0.275
S/b/	K/p/	%	0.030	0.757
		goodness	−0.001	0.992

and

Table 7. Results of Stuart-Kendall’s tau-c correlations between L3 proficiency and identification in English mode.

Stimulus	Identified As		τc	Sig
S/p/	E/p/	%	−0.118	0.365
		goodness	−0.124	0.287
S/p/	E/b/	%	0.118	0.365
		goodness	0.002	0.991
S/b/	E/b/	%	−0.027	0.857
		goodness	−0.030	0.805

).

Regardless of L3 proficiency, L3 learners predominantly categorized the Spanish /p/ as the Korean fortis /p*/ with a mean categorization rate of 89% and an average goodness-of-fit rating of 4.49 points. They also identified Spanish /b/ as the Korean lenis /p/, achieving a mean categorization rate of 98.4% and a goodness-of-fit rating of 4.12 points. Additionally, learners categorized Spanish /p/ as English /p/ with a mean rate of 66% and a goodness-of-fit rating of 3.17 points, while Spanish /b/ was identified as English /b/ with a mean categorization rate of 98.7% and a goodness-of-fit rating of 4.06 points. Figure 4 illustrates the combined results of the identification tasks, with detailed numerical values presented in

Table 8. Identification means and goodness-of-fit ratings in English and Korean mode. K/ph/: Korean aspirated; K/p*/: Korean fortis; K/p/: Korean lenis.

Stimuli (Sp)	Identification in English				Identification in Korean
	/p/		/b/		/p*/		/ph/		/p/
	%	Rating	%	Rating	%	Rating	%	Rating	%	Rating
Sp /p/	66	3.17	34	2.89	88.8	4.49	3.5	3.8	7.3	4.08
Sp /b/	1.13	0.45	98.7	4.06	0.7	2.66	0.5	3	98.4	4.12

.

5. Discussion

Our first research question investigated whether trilingual participants could discriminate between cross-linguistic categories and whether this ability improved with L3 proficiency. Drawing on the SLM-r’s postulate P2, which states that “bilinguals strive to maintain phonetic contrast between all consonants and vowels in their common phonological space”, we hypothesized that advanced trilingual participants would successfully distinguish cross-linguistic categories. However, our findings revealed two key patterns.

First, cross-linguistic phoneme discrimination was poor across most contrasts examined in this study. The notable exception was the discrimination of aspirated/unaspirated pairs, which participants consistently distinguished accurately. This finding about the discriminability of aspiration aligns with previous findings in L2 studies (e.g., Nagle, 2021) and is not surprising given that aspiration is an acoustic feature used in most languages (Lisker & Abramson, 1964) and even chinchillas, quail, and macaques were shown to be able to discriminate VOT contrasts (e.g., Kuhl & Miller, 1975). The fact, though, that other pairs which rely on F0 or negative VOT show low discrimination is harder to understand given that the participants are multilinguals familiar with the tested contrasts.Second, L3 proficiency was not associated with discrimination accuracy in our participant group. This is again unexpected because the general consensus from empirical L2 research is that there is a positive relationship between L2 proficiency and the ability to accurately perceive and discriminate L2-specific sounds. For instance, Wilson and Gick (2011) investigated low-intermediate Japanese learners of English and found a strong correlation between the learners’ ability to discriminate consonant sounds presented in nonsense syllables and their scores on the listening comprehension section of the TOEIC test. This correlation is also found in several studies by Flege and colleagues and makes its way to the SLM set of predictions (e.g., Flege, 1995; Flege & Liu, 2001; Flege et al., 1999).

Starting with the first finding (discrimination limited to aspiration), pairs with aspirated phones (characterized by long-lag VOT) showed high discrimination rates. For example, S/p/-E/p/ and S/b/-K/p/ were well-discriminated because E/p/ and K/p/ are aspirated. Aspiration serves as a highly salient acoustic cue due to perceptible aperiodic noise after oral release (Barzilai, 2020; Silverman, 2003). The extended duration of aspirated stops enhances perception “at an auditory level not only through greater temporal separation between the closure and the voicing onset but also through increasing the rate of auditory nerve firing after a silent period of stop closure” (S. Kim et al., 2012). This aligns with Nagle’s (2021) longitudinal study, which found that L2 learners performed better in discrimination tasks involving pairs with long-lag VOT compared to those without. The first finding aligns also with Wrembel et al. (2020), who highlighted the role of perceptual saliency in L3 perception, particularly for rhotics and final obstruent devoicing. This suggests that acoustic salience plays a crucial role in cross-linguistic phoneme discrimination.

Importantly, the poor discrimination across all contrasts except aspiration appears to contradict the SLM-r’s postulate P2, which formed the foundation of our H1 hypothesis. In H1, we predicted that early-stage L3 learners would struggle with phonetic contrast discrimination, with their abilities improving as L3 proficiency increased. However, our results did not support this expectation.

Interestingly, the same group of learners who failed to show clear evidence of cross-linguistic knowledge of phonetic categories in the current study demonstrated clear developmental progress in their productions of these same phonetic categories, with VOT values aligning more closely with Spanish patterns as proficiency increased (see Mun, 2022). Although L2 phonology research documents instances in which learners are able to produce L2 contrasts they do not reliably perceive (e.g., Flege, 1995), the present results may be attributed to the application of L2 tasks within an L3 context.

Before proposing an interpretation of these findings, it is important to consider whether the observed lack of discrimination may be due to the use of pseudo-words, which do not provide the lexico-semantic cues that listeners might otherwise use to distinguish between sound pairs. We used pseudo-words to test the predictions of the SLM-r, which, unlike the PAM-L2 (which incorporates a phonological level implicitly assuming lexical contrasts) or L2LP (which specifically focuses on perceptual mappings supporting lexical contrast), assumes processing that moves from cues to categories without lexical unit mediation.

While the absence of lexical access may have played a role, the importance of top-down processing should not be overemphasized. Lexical and contextual information enhances bottom-up processing but is not essential for discrimination. If lacking lexical information automatically resulted in poor discrimination, this pattern would have emerged in the many previous studies using pseudo-words, nonce-words, or syllables. Previous research consistently has shown that phonetic discrimination remains possible without lexical access (e.g., Best et al., 2001; Black et al., 2024; Nittrouer & Lowenstein, 2015; Werker & Tees, 1984). Therefore, rather than attributing the results to the lack of lexical access, we propose that the participants in this study may have been relying on domain-general discrimination cues (e.g., universally salient features like aspiration) rather than language-specific cues (e.g., F0 for Korean).

However, this raises two critical questions: why did this pattern emerge in our study, and why is such poor discrimination using pseudo-words not typically reported in previous research? The discrepancy suggests the need for closer examination of the specific factors influencing perception in an L3 acquisition context. One possible explanation may lie in the characteristics of the trilingual learner population examined in this study, whereas previous studies employing similar tasks involved monolinguals or bilinguals. For monolinguals, only one perceptual system is available regardless of whether the stimulus is familiar or novel. For bilinguals, most studies using the oddity task with meaningless syllables do not encounter situations with competing cues across languages. Even when such constraints exist, participants likely default to a “foreign language mode”, that is, a cognitive state where bilingual participants mentally shift to processing sounds according to the phonological system of their non-native language during experimental tasks (Grosjean, 2001). Our situation was more complex, with contrasting cues and three potential decoding systems available to participants. We interpret this indeterminacy as triggering reliance on universal domain-general processing mechanisms rather than language-specific perceptual strategies.

This domain-general processing can explain why cross-linguistic phoneme discrimination was largely limited to pairs involving aspirated sounds, why pairs without aspiration showed poor discrimination, and why L3 proficiency was not associated with discrimination accuracy. In every instance, language-specific knowledge about the categories and cues of each language was not tapped.

The findings suggest that the tasks may not have elicited the intended language-specific processing, leading participants to rely on domain-general perceptual mechanisms. The tasks did not appear to effectively induce a language mode aligned with either Korean or English. This raises the issue of language activation during task performance—specifically, whether participants were operating primarily in their L1, their L2, in a language-neutral state, or with multiple language systems activated simultaneously. If only the L1 or L2 had been activated, participants would be expected to rely on language-specific cues from either Korean or English. However, such patterns were not observed in the present study. Alternatively, if no specific language system had been activated, participants may have engaged in domain-general auditory processing, akin to basic acoustic discrimination mechanisms observed across species. This interpretation aligns with the observed data but diverges from current theoretical accounts of language activation in multilingual individuals, which generally assume some level of language-specific engagement. If in our discrimination and identification tasks our listeners were in an ambiguous language mode that did not support the activation of a specific language, theories like the Inhibitory Control Model (ICM, Green, 1998) would hypothesize that all languages in a multilingual individual could become activated to some degree. This kind of coactivation is what is typically used to explain interference and code switching.

According to the ICM, the active inhibition of non-target languages plays a central role in enabling bilinguals and multilinguals to achieve effective language control. A key principle of the ICM is reactive inhibition. According to this principle, inhibition of non-target languages is not a pre-emptive process but rather occurs in response to the activation of linguistic elements from the other languages. The ICM also posits that inhibition is proportional to activation. This means that the strength of the inhibitory signal applied to a non-target language is directly related to the level of activation of the target language. A corollary of these principles is that if there is no activation or selection of a language in a context, all languages will have similar opportunity to be activated to a certain degree.

The ICM is mainly a model of lexical activation, but for the case at hand, the result of multiple activations would be that the phonetic categories of each language in the multilinguals in this study would mesh together in macro-categories. We assume these categories to be arguably the categories used in processing the target syllables and pseudo-words in the perception tasks. These merged categories would not be very different from the composite categories posited by the SLM-r.

Our basic assumption is that categories are the result of distributional learning (DL). DL has been shown to be at play in category formation during L1 acquisition (e.g., Maye et al., 2002) and in adulthood (e.g., Maye & Gerken, 2001). This mechanism is assumed to be the cognitive process through which learners extract patterns from the statistical distribution of features in their input. In the context of the present study, learners in the process of acquiring each language used the frequency distribution of acoustic–phonetic cues to form phonetic categories for the phonemes and phonetic categories of each language. The SLM-r, PAM-L2, and L2LP all incorporate distributional learning in their accounts of category formation, emphasizing that phonetic categories develop based on the input learners receive. According to this mechanism, as a learner processes tokens with, for instance, specific VOT values, each value leaves a memory trace. These traces accumulate over time, and their frequency produces a normal distribution (bell-shaped curve). From the perspective of DL, categories are essentially clusters of memory traces built through input processing applicable across L1, L2, L3, and Ln. Learners form an initial approximation of the target distribution after processing a few tokens of a given cue. With more experience, these tokens accumulate, and the distribution gradually normalizes, aligning more closely with the target distributions typical of the learning context—in the present case, Spanish. In a given language (let us assume Spanish), two contrastive sounds, such as /p/ and /b/, would normally have separate distributions. In the case of a multilingual and assuming a common space, there are several distributions with the possibility of overlap between phones of different languages. For perception, such overlap would present a challenge if one does not assume that only a single language is activated while others are suppressed through inhibition. If, however, no language was inhibited during task performance, the resulting scenario may involve overlapping density distributions, potentially leading to the formation of broader, composite phonetic categories.

In an L1 context, multiple activation is never an issue. In a bilingual context, this kind of macro-category can, in theory, occur, but it is still unlikely to show effects. In a multilingual context, the level of ambiguity is higher, the crowding of the space is greater, and the effects of coactivation become more apparent.

In summary, while a domain-general interpretation accounts for much of the present study’s findings, the coactivation account, where overlapping language systems give rise to composite macro-categories used in perception but not necessarily in production, offers a more compelling explanation. This perspective also aligns more closely with current understandings of how multilinguals process language.

6. Concluding Remarks

This study investigated Spanish stop perception among L1 Korean speakers with L2 English and varying Spanish proficiency levels. Contrary to SLM-r predictions, increased Spanish proficiency was not associated with enhanced discrimination or identification abilities.

Discrimination task results revealed that participants primarily utilized universally salient acoustic cues (e.g., aspiration) rather than language-specific features. The identification task similarly showed inconsistent application of cues from participants’ known languages. These patterns suggest that participants employed general acoustic processing strategies during perceptual tasks. We propose that these findings align with phonetic perception governed by distributional learning mechanisms and language coactivation phenomena. The present data suggest that phoneme distributions from all three languages likely occupy a shared perceptual space, with selective activation mechanisms playing a crucial role in managing multiple phonological systems. In experimental contexts with decontextualized stimuli lacking clear language cues, multilinguals may experience simultaneous activation of their phonological systems, resulting in composite perceptual categories. This interpretation aligns with the Inhibitory Control Model (Green, 1998), which posits that language selection involves inhibiting non-target languages. When definitive selection cues are absent, these inhibition mechanisms may function less effectively, leading to coactivation across phonological systems—a pattern observed in the present study.

The present findings highlight the need to re-evaluate current conceptualizations of phonetic perception in multilinguals. Models originally developed for L1–L2 contexts may require substantial modification to address the complexities of managing three or more phonological systems. Future research should examine how language mode activation influences perception in multilingual contexts.