1

PRE-PROOF ACCEPTED VERSIONPublished in Journal of Phonetics

Title: Influence of L2 English phonotactics in L1 Brazilian Portuguese illusory vowel perception

Jennifer Cabrellia*, Alicia Luquea, Irene Finestrat-Martíneza

*Corresponding author: cabrelli@uic.edu (note: Cabrelli is full surname); other author emails: aluque2@uic.edu; ifines2@uic.edu. Tel: +1 312-996-0925, Fax: +1 312-413-1044a The University of Illinois at Chicago, Department of Hispanic and Italian Studies, 601 S. Morgan St., 1712 University Hall MC 315, Chicago, IL, 60607, USA

Abstract

This study examines potential changes to L1 (Brazilian Portuguese, BP) perception of phonotactic structure as a function of L2 (English) experience. Syllables with a coda stop violate syllable structure constraints in BP, but are licit in English. As a result, BP monolinguals perceive an illusory /i/ after an illicit coda (e.g., ob/i/ter ‘to obtain’). To understand the effects on L1 perception when the L2 phonology allows a syllabic structure that is illicit in the L1, we tested 15 L1 speakers of BP in an L2 English immersion setting. In separate BP and English sessions, participants completed an explicit metalinguistic vowel identification task and two ABX tasks, with one ABX task designed to tap lower-level processing and the other, higher-level phonological processing. Our analysis reveals that, while not fully target-like in L2 English, the bilinguals’ perception data in both languages contrast sharply with existing monolingual BP data. This is the case whether participants employ what we assume here to be metalinguistic strategies, lower-level encoding strategies, or higher-level encoding strategies in perception. Based on this pattern of results, we conclude that the data provide evidence of L2 phonotactic influence in L1 perception.

Keywords: Crosslinguistic influence, phonotactics, perception, Brazilian Portuguese, illusory vowels

2

1.0 IntroductionThe cross-linguistic influence that occurs between a first language (L1) and a language learned in adulthood (L2) has been a primary focus in the field of second language acquisition, with the majority of attention dedicated to the influence of the L1 on the L2. However, it is important to acknowledge the bidirectionality of such a process, inasmuch as the acquisition of an L2 also has shown to affect the L1. While L1 restructuring can affect all domains of language, research on L2 phonetic and phonological influence on the L1 has lagged behind the lexical and morphosyntactic domains (see Schmid, 2016, for a timeline of research in this area across domains). In the present study, we examine potential changes to L1 perception at the suprasegmental level. While L1 phonetic and phonological restructuring has been documented in production at the segmental and suprasegmental level, less is known about perception, particularly at the suprasegmental level (see section 2.1). With the paucity of data addressing suprasegmental perceptual effects in mind, we investigate the perception of illusory vowels by L1 Brazilian Portuguese (BP)/late L2 English bilinguals. Syllables with a coda stop violate syllabic structure constraints in BP, but are licit in English. As a result, BP monolinguals perceive an illusory /i/ after an illicit coda (e.g., ob/i/ter ‘to obtain’) (e.g., Dupoux, Parlato, Frota, Hirose, & Peperkamp, 2011). Parlato-Oliveira et al. (2010) found no effect of L2 BP illusory vowel quality (/i/) on perception of the L1 Japanese illusory vowel (/u/) when the L2 was acquired in adulthood, which could lead to the conclusion that L1 illusory vowel perception is impermeable to L2 influence. However, effects on L1 perception are largely unknown when the L2 phonology allows a syllabic structure that is illicit in the L1, particularly when it comes to phonological processing. In order to further investigate how the acquisition of a less restrictive L2 phonotactic system affects L1 perception, we examine English and BP perception in 15 L1 speakers of BP in an L2 English immersion setting.

2.0 Background

2.1 L2 influence on L1 phonetics and phonologyL1 phonetic and phonological plasticity has been documented primarily at the segmental level and via production data (e.g., de Leeuw, Mennen & Scobbie, 2013; de Leeuw, Tusha, & Schmid, 2017; Major, 1992; Mayr, Price, & Mennen, 2012; Yang & Fox, 2017), with effects observed in as few as six weeks in an L2 immersion setting (Chang, 2012). The comparably fewer studies which examine production of suprasegmental phenomena also report deviations from the native norm (e.g., Alcorn & Smiljanic, 2017b; Cabrelli Amaro, 2017; de Leeuw, Mennen, & Scobbie, 2012; Mennen, 2004). To our knowledge, five1 studies have examined L2 effects on L1 perception (Alcorn & Smiljanic, 2017a; Cabrelli Amaro, 2017; Carlson, 2018; Celata & Cancila, 2010; Parlato-Oliveira et al., 2010), with mixed outcomes. Celata and Cancila report that discrimination of the Lucchese singleton-geminate consonant contrast by L1 Lucchese/L2

1 Major (2010) also examines L2 influence on L1 perception (in this case, of foreign accent perception). However, we consider the domain of foreign accent to fall outside the scope of the segmental- and syllable-level studies reported on here.

3

English speakers evidences L1 changes at the phonological level. On the other hand, Cabrelli Amaro found no evidence of L3 BP vowel reduction on the L1 Spanish of L1 Spanish/L2 English participants. However, Celata and Cancila’s participants were long-term residents of the L2 environment (Myears = 39.38, SD = 9.33), while Cabrelli Amaro’s had been residents in the L2 environment for less time (Myears = 8.14, SD = 2.27) and a subset were tested in the L3 environment after six weeks of exposure. Specific to our focus on L2 syllable structure effects on L1 perception, Parlato-Oliveira et al. found L1 resistance to L2 influence in phonological perception. Their study compared L1 BP/adult L2 Japanese and L1 Japanese/L2 BP speakers with early BP/Japanese bilinguals and BP and Japanese monolinguals. Perception of epenthetic vowels has been documented in BP and Japanese due to prohibition of coda stops, with Japanese speakers primarily perceiving /u/ (e.g., Dupoux, Kakehi, Hirose, Pallier, & Mehler, 1999) and BP speakers primarily perceiving /i/ (e.g., Parlato-Oliveira et al.). In an explicit vowel identification task similar to one we implement in this study, the L1 Japanese/L2 BP participants (who had been in Brazil for 35 years) patterned with Japanese monolinguals while L1 BP/adult L2 Japanese participants identified both /i/ (BP) and /u/ (Japanese). However, the L1 BP/L2 Japanese participants patterned with BP monolinguals (i.e., perceived /i/) in an implicit sequence recall task thought to tap phonological processing (the L1 Japanese/L2 BP group did not complete the recall task). The findings from these adult L2 learner data contrast with the authors’ early bilingual data, which align with the BP monolingual data. Although it is possible that these findings are indicative of invulnerability of L1 phonological processing to L2 influence (at least in the quality of the illusory vowel perceived), it is also possible that the participants had not acquired the relevant representation in the L2. Finally, two recent studies investigate L1 restructuring when the relevant phonotactic restriction in the L1 is absent from the L2 phonotactic system. Data in Alcorn and Smiljanic (2017a) from an explicit vowel identification task completed by L1 BP/L2 English listeners show that L2 English proficiency correlates with sensitivity to the presence or absence of word-medial /i/ in the auditory input. Carlson (2018) examined phonetic processing of word-initial sC clusters, which are illicit in Spanish but not in English, by L1 Spanish/adult L2 English learners. Results from an AX task with a 250 ms interstimulus interval (ISI) indicate a weakened perceptual illusion of word-initial /e/ in L1 Spanish, with a larger effect for learners immersed in L2 English. This pattern of results suggests that acquisition of less restrictive phonotactic structure can influence L1 perception, at least when relying on metalinguistic strategies (Alcorn & Smiljanic, 2018) and in lower-level phonetic encoding (Carlson, 2018).

The data in Carlson (2018) and Alcorn and Smiljanic (2017a) align with early bilingual data that show that perceptual epenthesis is weakened or even absent when a more permissive language is acquired in childhood. This is the case for early Spanish-English bilinguals’ perception of #sC clusters (Carlson, Goldrick, Blasingame, & Fink, 2016) and early Korean-English bilinguals’ perception of heterosyllabic consonant clusters (Lee-Ellis, 2012), as indicated by tasks designed to engage metalinguistic (Carlson et al.), acoustic (Carlson et al.; Lee-Ellis), and phonological processing strategies (Lee-Ellis). According to Freeman, Blumenthal, and Marian (2016), who found facilitative influence of Spanish phonotactic constraints in English

4

comprehension by early bilinguals during a cross-modal phonological priming decision task, this influence is a result of parallel activation of phonotactic constraints between the two languages of a bilingual, and the non-relevant system is accessed in comprehension. In the present study, we examine effects of potential L1/L2 co-activation2 in L1 perception via explicit and implicit measures meant to tap auditory and phonological processing.

2.2 Illusory vowels and phonotactic structure in BP and English Illusory vowel perception is a phenomenon whereby speakers perceive a vowel that is not present in the input in order for illicit input to conform to their L1’s well-formedness restrictions. As noted in the previous section, this phenomenon has been reported for monolingual speakers of Japanese (Dupoux et al., 1999; Dupoux et al., 2011) and BP (Dupoux et al., 2011; Parlato-Oliveira et al., 2010), as well as a number of other languages with strict phonotactic constraints including Spanish (Cuetos, Hallé Domínguez, & Segui, 2011) and Korean (e.g., Kabak & Idsardi, 2007), among others.

2.2.1 BPIn BP, the only obstruent permitted in coda position is /s/ and stop consonants can only appear in onset position. Input containing an illicit stop coda is repaired via insertion of /i/, which allows the stop to (licitly) occupy a syllable onset position.3 A clear illustration is found in the Portuguese word apto /apto/ ‘apt’, which is perceived (and produced) with an epenthetic vowel (ap/i/to), the result of which is syllabification of /p/ in onset position. In perception, this epenthesis causes difficulties for speakers to distinguish between an input with a stop coda (e.g., apto) and an input repaired via vocalic epenthesis (e.g., ap/i/to), and both outputs (pronounced forms) for BP monolinguals would be predicted to be the same.4 This perceptual effect has been confirmed in BP monolinguals via explicit metalinguistic tasks as well as in implicit ABX and sequence recall tasks (Dupoux et al., 2011; Parlato-Oliveira et al., 2010). Crucially, Dupoux et al. found that European Portuguese (EP) monolinguals do not exhibit a perceptual epenthesis effect despite the structural similarities between EP and BP, which indicates that the effect is language specific. Much of the research on these language-specific adaptations of input that does not conform to L1 phonotactic structure has centered on whether the adaptations are driven by perception, phonology, or both (see e.g., Kang, 2011, for a review). Given the language-specific nature of illusory vowels as shown in experiments such as Dupoux et al., we follow Boersma and Hamann’s (2009) proposal that speech perception (and thus perception of illusory vowels) is phonological. Specifically, linguistic perception (as opposed to auditory perception) depends on

2 Herein, we discuss co-activation in broad terms; as an anonymous reviewer points out, the design of our study does not allow us to isolate the mechanism for L2 influence as system-wide co-activation, co-activation of specific phonotactic constraints, or even perceptual retuning without real-time co-activation. We return to this issue in our discussion of future steps in section 6.3. 3 We assume a traditional generative (e.g., Selkirk, 1982) analysis of the word-medial stop-obstruent sequences as coda-onset sequences.4 For production data, see e.g. John and Cardoso (2017) and studies cited therein.

5

the relationship between the auditory form, which contains a range of cues, and the surface form, which is the abstract representation of surface structure and lacks gradient phonetic detail.

Boersma and Hamann’s (2009) proposal is based on Boersma’s (e.g., 2007) Bidirectional Phonetics and Phonology (BiPhon) framework, which is an extension of Optimality Theory (Prince & Smolensky, 1993[2004)]. The framework employs a set of cue constraints used to account for the relationship in perception between auditory processing and abstract (surface) forms. Optimal surface form candidate selection depends on the interaction of these cue constraints and structural (markedness) constraints, the latter of which evaluate the phonological surface form. In line with Boersma and Hamann’s account of illusory vowel perception in Korean loanwords, the interaction between a) a structural conjoined constraint that militates against stops in coda position and guides perception (specifically, *CODA & *[-cont]) and b) two critical cue constraints determines whether an illusory vowel is perceived in the input.5 The first cue constraint prohibits perception of a stop as a phonological coda (*[burst]/C(.)/), which is strong in a monolingual BP grammar. The second cue constraint militates against perception of a segment that is not part of the auditory signal (*[ ]/i/) and is weaker in a monolingual BP grammar. The tableau in Table 1 illustrates the interaction of these constraints applied to the auditory-phonetic input for the VCCV nonce stimulus abza. The strength of *[burst]/C(.)/ in relation to *[ ]/i/ yields an optimal candidate with an epenthetic /i/. In the narrow auditory transcription in the tableau, the underscore represents silence and superscript b represents a labial release burst.

Table 1. BP monolingual ranking[a_b.zɐ] *[burst]/C(.)/ *CODA & *[-cont] *[ ]/i/☞/a.bi.za/ *

/ab.za/ *(!) *(!)

2.2.2 EnglishUnlike BP, English permits coda stops, and there is no evidence to our knowledge that L1 English speakers perceive an illusory vowel at the offset of a coda stop. Data from studies such as Alcorn and Smiljanic (2017a) show that L1 English speakers perceive coda stops with accuracy approaching ceiling, and we replicated their finding with ten L1 English speakers (8 females, Mage = 27.30, SD = 6.96)) that completed the same identification and ABX tasks as the bilinguals in the present study (see Appendix C for results). With that said, it is true that the ABX tasks employed in this study with VC.CV and V.C/i/.CV pairs do not allow us to confirm that L1 English speakers do not hear an illusory vowel other than /i/, at least not when engaging phonetic and phonological processing strategies. However, the L1 English group data from the identification task in this study (the only task in which the quality of the illusory vowel might be

5 See also Azevedo, Matzenauer, and Alves (2017) for an analysis of BP /pn/, /pt/, and /pn/ heterosyllabic clusters within a BiPhon and Harmonic Grammar framework.

6

isolated) is not indicative of perception of an illusory vowel, and these data are in line with Alcorn and Smiljanic (2017a).

With these data in mind, we follow that the English grammar ranks *[ ]/i/ above the other two constraints. This ranking yields a VC.CV surface form that is faithful to the auditory input (Table 2).

Table 2. English monolingual ranking [a_b.za] *[ ]/i/ *[burst]/C(.)/ *CODA & *[-cont]

/a.bi.za/ *!☞/ab.za/ * *

2.3 The L2 learning taskIn order for co-activation to yield observable effects on L1 perception, a minimal condition is that a learner’s L2 system has to be distinct from their L1 for the phenomenon in question. That is, L2 learners have to have at least partially converged on the L2 target. With that in mind, in the following sections we model the specific L2 learning task for L1 BP/L2 English listeners and review the literature on L2 perceptual epenthesis to inform our predictions for L2 convergence and L2 influence on the L1.

For an L1 BP speaker acquiring L2 English, the perceptual learning task will be to rerank the critical constraints such that *[ ]/i/ outranks *[burst]/C(.)/ and *CODA & *[-cont] (Table 2).6 While the L2 initial state is assumed to be a copy of L1 BP system (Schwartz & Sprouse, 1996), as learning proceeds, L2 learners are hypothesized to be able to restructure the initial grammar to resemble the target language. Following Boersma (1997) and Boersma and Hayes (2001), we assume that convergence on the L2 target grammar is the result of the Gradual Learning Algorithm (GLA), which is an error-driven learnability algorithm based on stochastic OT. In this probabilistic model, constraints occupy a range on a continuous scale and the distance between constraints determines ranking strictness. In perception, constraints are ranked based on specific values that are selected from the range that each constraint occupies. A mismatch between the ranking and surface form forces the algorithm to continuously adjust the ranking values towards the ranking that will generate the optimal candidate. While target convergence is predicted to be possible, crucially ranked constraints with ranking values that are close together can result in variable outputs, which are widely attested in the L2 literature (see e.g., Broselow & Kang, 2013 for a review). As we will discuss further in section 4.1, our purpose in examining these listeners’ L2 English is to confirm whether they have perceptual acuity in the L2 that is different from what has been attested in a monolingual BP grammar, and not to determine whether their L2 6 See John and Cardoso (2017) for a usage-based proposal of L1 BP/L2 English illusory vowel perception which assumes that L2 English lexical forms are originally stored with an illusory vowel. At a later stage of acquisition, a second (target) lexical form is stored. As the target lexical form is strengthened and the non-target form is weakened, production of epenthetic /i/ reduces. However, the authors acknowledge that data from more than a single experimental session is needed to evaluate the validity of this account, and it is not clear how the learners overcome the perceptual illusion in order to store a target-like form.

7

perception is native-like. Thus, we prioritize a within-subjects comparison of participants’ perceptual acuity in the L2 and L1, rather than directly comparing their data with L1 English data.

2.4 L2 acquisition of phonotactic structureBefore an L2 phonological grammar can influence L1 perception, learners have to develop an L2 grammar that is distinct from the L1 grammar that we assume is initially transferred. In the following, we review the most relevant literature on the acquisition of L2 phonotactic structure by L1 speakers of different learning backgrounds. Although studies such as Carlson et al. (2016) and Lee-Ellis (2012) indicate that L2 English phonotactic constraints are successfully acquired in childhood, results are mixed in regards to whether late learners achieve target-like L2 phonotactic structure.7 Overall, research shows that late L2 learners at lower levels of proficiency are not able to accurately perceive coda consonants due to the influence of L1 syllable structure constraints on L2 perception (e.g., Cardoso, 2011a; Koerich, 2006; Lee-Ellis, 2012), even after explicit instruction (Silveira, 2011). At higher levels of proficiency, Cardoso reveals that advanced participants’ perceptual accuracy reaches only 59% in a vowel identification task, while Archibald and Yousefi (2017) report 82% accuracy in a similar task administered to L1 Persian/L2 English speakers to test perception of initial sC clusters. Matthews and Brown (2004) report accurate perception by L1 Japanese/advanced L2 English learners in 82% of trials in an AX discrimination task with a 250 ms ISI, although response time data suggests that additional processing steps were necessary to overcome the perceptual illusion in these trials. Finally, although not assessing L2 perception explicitly, Carlson (2018) found a weakened L1 perceptual illusion effect at the level of phonetic processing, which suggests that the L2 phonotactic system had been (at least partially) acquired. These studies assess perception through either explicit metalinguistic identification tasks or discrimination tasks designed to tap acoustic representations; their results speak to L2 perception at lower levels of processing.

To our knowledge, only two studies have examined the effects of L1 phonotactic constraints in advanced L2 learners at a level of phonological encoding. Matthews and Brown (2004) administered a second AX task with a longer ISI (1500 ms) which yielded an accuracy rate of only 63% and slower reaction times in the critical condition than the control condition. This means that learners perceived an illusory vowel in the L2 sequences that could not be accommodated into the L1 phonotactic constraints at the level of phonological processing; once detailed acoustic information decayed and learners had to rely on phonological representation, discrimination became less likely. In contrast to Matthews and Brown, Archibald and Yousefi (2017) reported 87% accuracy in an ABX task with an 800 ms ISI. While the authors propose that the difference is due to availability of relevant phonological elements available for redeployment (in this case, sC clusters which occur in coda position but are illicit in onset

7 L2 illusory vowel perception has been found to be conditioned by linguistic and extralinguistic factors including speech style, segment markedness, preceding segment, word position, and word size, among others (see e.g., Cardoso, 2011; de Jong & Park, 2012). Where relevant, we control for these factors in the study design.

8

position), another potential explanation for the difference in these two outcomes is the testing context. That is, whereas Matthews and Brown’s participants were tested in the L1 environment, Archibald and Yousefi’s were tested in the L2 environment. The difference in the results given the context of testing also follow from recent evidence that shows that L2 English proficiency correlates with d’ sensitivity in L1 BP (Alcorn & Smiljanic, 2017a). We revisit this possibility in our discussion of the results.

In light of Archibald and Yousefi’s (2017) findings, we follow the assumption that late learners are eventually able to perceive input free of epenthetic material when relying on low-level as well as high-level encoding strategies. Thus, advanced L2 learners such as those in our study should have an L2 phonotactic representation that is distinct from the L1 representation available for parallel activation in L1 perception. While this scenario has only recently been investigated with adult L2 learners in an explicit metalinguistic task (Alcorn & Smiljanic, 2017a) and a phonetic processing task (Carlson, 2018), data from early bilingual populations reveal the potential for a weakened illusory effect in the more restricted language even in phonological processing (Lee-Ellis, 2012). The present study thus aims to determine whether these findings can be replicated in adults L2 learners’ L1 perception across different levels of representation.

3.0 Research questions and predictions

3.1 Research questionsIn the present study we aim to investigate the effects of acquisition of a less restrictive L2 phonotactic system on L1 perception. To do so, we investigate perception of the illusory vowel /i/ in heterosyllabic obstruent-obstruent clusters by BP native speakers that are late L2 English learners. Specifically, we ask a) to what degree these learners overcome the perceptual illusion and accurately perceive coda stops in their L2, as indicated by perception in a series of tasks designed to invoke metalinguistic, auditory, and phonological processing strategies, and b) whether their L2 perception of coda stops in turn influences their L1 perception in these same tasks administered in BP. Recall that previous research has pointed to differences in perception depending on task type (e.g., Dupoux et al., 2011; Parlato-Oliveira, 2010) and ISI (e.g., Matthews & Brown, 2004; Wayland & Guion, 2003). Thus, in order to establish a holistic picture of the learners’ L2 phonotactic perception and the nature of L2 effects on their L1 perception, we implement an explicit metalinguistic identification task and an ABX task with two ISIs that have been proposed to reflect auditory versus phonological processing (see section 4.3 for details).8

3.2 Predictions8 We recognize that an ABX paradigm is typically considered to be more memory-intensive than other discrimination tasks (e.g., Gerrits & Schouten, 2004) and thus more likely to induce higher-level phonological encoding even with a shorter (e.g., 250 ms) ISI. However, studies including Gerrits (2001) and Wayland and Guion (2003) report differences between ABX or AXB tasks with shorter (500ms) versus longer (1000 ms-1500ms) ISIs and cite a difference in the type of encoding as an explanation. That said, we acknowledge that it is not possible to rule out a number of other variables beyond ISI that influence processing strategies.

9

In what follows, we outline our predictions for our two research questions, dividing our predictions by task type. It is important to note for all tasks that, while our predictions hold for the group level, we expect to find substantial individual variation based on previous studies of phonetic and phonological restructuring.

3.2.1 L2 acquisition

3.2.1.1 English vowel identification taskFollowing Archibald and Yousefi (2017), we predict sensitivity to the absence of /i/ in VC.CV stimuli will approach ceiling but be lower than sensitivity to the presence of /i/ in a V.C/i/.CV stimulus. This would serve as evidence of metalinguistic knowledge of the relevant English phonotactic structure, available to influence L1 perception.

3.2.1.2 English ABX tasksWe predict participants will accurately discriminate between VC.CV and V.C/i/.CV pairs in an ABX task with a 250 ms ISI, although reaction times may be longer compared to discrimination in V.C/i/.CV and V.C/a/.CV or V.C/o/.CV pairs, following Matthews and Brown (2004). This outcome will indicate sensitivity to cues in auditory perception, albeit with additional processing steps, which the BiPhon framework models as the perception of cues without translation to surface phonological forms. The accuracy results in Archibald and Yousefi (2017) lead us to predict a similar outcome in an ABX task with a 1000 ms ISI, which would indicate true linguistic/phonological perception (i.e., mapping of cues to surface forms). In such a case, a distinct L2 grammar should then be available to influence L1 perception.

3.2.2 L2 influence on L1 perception

3.2.2.1 BP vowel identification taskSensitivity in the L2 is expected to correlate with sensitivity in the L1, following Brien and Sabourin’s (2014) report that late L2 learners’ metalinguistic knowledge influences L1 processing of homonyms and the results in Alcorn and Smiljanic (2017a). This outcome would reflect influence of L2 metalinguistic strategies in L1 perception.

3.2.2.2 BP ABX tasksIn light of the results from Carlson (2018), we predict a correlation in sensitivity between the English and BP ABX tasks with a 250 ms ISI. We do not make specific predictions for the ABX task with a 1000 ms ISI due to the conflicting outcomes in the studies of L2 effects on L1 phonological processing we discuss in 2.1.

4.0 Methodology

4.1 Participants

10

The participants were 15 L1 BP/L2 English speakers immersed in the L2 environment at the time of testing. They were all enrolled in a yearlong direct-enrollment program at a large public university in the Midwestern US as part of the Brazilian government’s Sciences without Borders initiative. All participants were BP dominant (according to the Bilingual Language Profile (BLP), Birdsong, Gertken, & Amengual, 2012) and reported using primarily English in an academic context and BP outside of that context and with their Brazilian classmates. In addition to meeting minimum International Test Program TOEFL requirements for the program (550, or a score of 525-549 plus a month-long intensive English course upon arrival), participants completed a 50-point written proficiency assessment adapted from the Oxford Placement Test.

Table 3. Participant characteristics

M SD

Age (years) 23.13 2.72Age first exposed to L1 0.00 0.00Age first exposed to L2 13.93 2.60Length of residence in US (years) 0.87 0.40English Proficiency Scorea 39.33 2.29Language Dominance -79.19b 22.61

Note. aMaximum score 50 bNegative value indicates BP dominance

As previously discussed, the L1 has generally been reported to be most vulnerable to L2 phonetic and phonological influence when the speaker a) has acquired the L2 at a young age (e.g., Carlson et al., 2016), b) is L2 dominant (e.g., Celata & Cancila, 2010), and/or c) has been in the L2 environment for a long period of time (e.g., Flege, 1987). With this in mind, one might predict that this sample would not evidence changes to their L1 phonotactic system. Nonetheless, there is contrary evidence that suggests that the effects of an L2 on L1 segmental production are also evident in late bilinguals after only a short period abroad (e.g., Cabrelli Amaro, 2017; Chang, 2012), and even among those whose only L2 exposure is through classroom instruction in their home country (Lord, 2008). Of most relevance to the current study are the L2 effects on L1 perception reported in (a) Alcorn and Smiljanic (2017a), which were found for L1 BP/L2 English speakers residing in the US (M = 2.9 years) as well as those in Brazil; and in (b) Carlson (2018), this time in L1 Spanish/L2 English speakers residing in the US (M = 5.54 years) and in Spain.

We use within-subjects analyses to determine learners’ sensitivity to coda stops in VC.CV items in English and in BP, rather than comparing the bilinguals’ data with monolingual BP and English data. We recognize that comparing the bilingual BP data with monolingual BP data would be ideal to confirm the illusory vowel effect in perception of nonce VC.CV items reported in Parlato-Oliveira et al. (2010) and Dupoux et al. (2011). However, a subset of our

11

sample reflects the effect reported for monolingual BP speakers.9 In light of this finding, we refer to the BP monolingual data from Dupoux et al. and Parlato-Oliveira et al., who utilize similar methodologies, as a point of comparison in our discussion. With respect to comparing the bilinguals’ L2 English data with L1 English data, while we tested a group of L1 English participants to determine the absence of perception of an illusory vowel, we take the position that holding the bilingual’s L2 English data to an L1 English standard would be in line with Bley-Vroman’s comparative fallacy (1989, see also recent critical discussions of native vs. non-native comparisons in L2 research, e.g., Birdsong & Gertken, 2013) and would “…move the yardstick of nativelikeness to a point which may, by definition, be out of reach for most bilinguals” (Hopp & Schmid, 2013, p. 354). Moreover, our primary aim is to determine whether L2 perception influences L1 perception, rather than to determine whether learners converge on a native target. Unlike relevant studies such as Carlson (2018) and Alcorn & Smiljanic (2017a), which only tested the participants’ L1, we test participants’ L2 English in great part to confirm that any lack of L2 English influence in BP perception is not because participants’ L2 English still looks like their L1 BP.

4.2 Study procedureThe study was divided in three sessions. During session 1, language history and background information were collected and L2 proficiency was assessed. After providing consent to participate in the study, all participants completed a language background and experience questionnaire (Bilingual Language Profile, Birdsong et al., 2012), and the 50-item English proficiency assessment. Task order was fixed across participants. BP data were collected in session 2 and English data were collected in session 3; this order was fixed in order to capture L2 effects on the L1 without any priming from the L2 English tasks. Sessions 2 and 3, conducted in monolingual mode (including task instructions), consisted of (a) a 15-minute interview to bring participants into the relevant language mode, (b) a syllable concatenation production task (reported elsewhere) and (c) three perception tasks, the order of which was counterbalanced across participants. During the perception tasks, participants were seated in a comfortable chair wearing AKG K240 MKII headphones in a sound-attenuated booth with a Logitech keyboard and a Cedrus RB-844 response box (Cedrus corporation, San Pedro, CA, USA) placed in front of them. Tasks were administered using e-Prime 2.0.10.356 (Psychology Software Tools, Inc.) and auditory stimuli were delivered through a MOTU UltraLite-mk3 audio interface.

4.3 Tasks

4.3.1 Vowel identification tasksFollowing Dupoux et al. (1999) and Parlato-Oliveira et al. (2010), we examined perception of stop codas at an explicit metalinguistic level in English and BP via vowel identification tasks

9 We refer to participant 9 in the identification task, and 3 and 16 in the ABX tasks. Participant 16 was excluded from the final data set because he had negative d’ sensitivity in the identification task. See Appendix B for individual data.

12

completed in each language. Participants were told they would hear a series of invented words in the language being tested that day, and were asked to identify the presence or absence of a vowel between heterosyllabic consonants in a VC.CV or V.CV.CV auditory stimulus.

4.3.1.1 StimuliEach task (BP and English) contained a pseudorandomized set of 48 trials with four conditions: a) eight VC.CV items (ø condition), b) eight VC/i/CV items (/i/ condition), c) 24 VC/a/CV and VC/o/CV control items (/a/-/o/ condition) (six items per vowel with two repetitions of each item), and d) eight VC/i/CV filler items with the [i] segment extracted. The filler items are part of a related project and were excluded from analysis. All items had initial stress and the word-initial and word-final vowels were always /a/. The first consonant was a stop in the set /p b k g/, and the second consonant was a stop or fricative that matched the first in voicing. The stimuli set included the items apka, apsa, abga, abza, akfa, akpa, agba, agza and their V.C/i/.CV and V.C/a/.CV or V.C/o/.CV counterparts. The same set was used in English and BP to maximize comparability between languages (see Appendix A for segmental duration), although between-language phonetic differences (the most robust of which we assume to be vowel and stop quality) were maintained to maximize engagement of the language mode being tested. The BP stimuli were produced by a male phonetically trained native speaker of BP who has near-native proficiency in L2 English and can produce stops in coda position without an epenthetic /i/. The English stimuli for the identification task were produced by a female phonetically trained native speaker of English.

4.3.1.2 ProcedureIn each trial, participants saw a fixation cross on a monitor for 1000 ms followed by a visual prime that followed a VC?CV pattern (e.g., ab?za). The onset of the auditory stimulus occurred 1000 ms after the presentation of the visual prime. Participants then had 2500ms to identify whether they had heard a vowel between the two consonants (i.e., where the question mark was located). Responses were made via button press from a set of six options (a, e, i, o, u, n, where n = none/nenhuma) and failure to respond within 2500ms resulted in exclusion of the trial from analysis.10 Prior to the critical block of 48 trials, participants completed a four-item practice block with no feedback consisting of two /i/ condition items and two /a/-/o/ condition items.

4.3.1.3 Analysisd’ sensitivity scores were calculated from the accuracy data, with a ceiling of 4.01 (taken from Table A5.7 for 6-interval identification in MacMillan and Creelman, 2005, pp. 426-430). Each participant had six d’ scores (Language (2ENG,BP) × Condition (3ø,/i/,/a/-/o/). We also analyzed response time in accurate trials (RT, measured in milliseconds from the offset of the stimulus), excluding RTs < 100ms (see e.g., Luce, 1986). This exclusion resulted in the removal of two trials from the BP task and 15 from the English task; remaining RTs were log transformed to

10 Six trials were excluded from the BP task and 12 from the English task.

13

correct for positive skewness. As in Matthews and Brown (2004), we assume it is possible that L2 English learners that successfully perceive a coda stop incur additional processing burdens while doing so, which we would predict to be reflected in longer RTs.

d’ and RT data were fit to linear mixed effects models with the maximal random effects structure allowed by the data and included by-item intercepts (RT model only) and by-subject slopes across Language, Condition, and Language×Condition. Pairwise comparisons were Bonferroni corrected. Intra-individual Pearson correlations were also calculated to determine the relationship between identification in English and BP. Alpha was set at .05 for all analyses.

4.3.2 ABX discrimination tasksTo examine learners’ ability to discriminate between auditory inputs with and without an epenthetic vowel, we constructed two ABX discrimination tasks in each language. As in the identification task, participants were told they would hear a series of invented words in the language being tested that day. One task had a 250 ms ISI and the other had a 1000 ms ISI. While trial content was identical between tasks, trials in each task were independently pseudorandomized.

4.3.2.1 StimuliThe experimental block of each task included 40 pairs of AB stimuli (eight critical pairs, 16 control pairs, and 16 filler pairs). Critical trials consisted of VC.CV versus VC/i/CV pairs and control trials were made up of VC/i/CV stimuli versus either VC/o/CV or VC/a/CV. Filler pairs (n = 16) consisted of VC/i/CV stimuli with the medial vowel extracted, compared with either VC/i/CV (eight pairs) or VC.CV (eight pairs). Each of the eight pairs in each condition appeared in all four possible stimuli combinations (ABA, ABB, BAA, BAB) for a total of 32 trials each in the critical and control conditions, and 64 in the filler conditions (n=128). Two speakers produced the stimuli in each language. In the English tasks, the AB stimuli were the same as the English identification task stimuli; a male native English speaker recorded the X stimuli. For the BP tasks, a phonetically trained female near-native BP speaker (L1 English) recorded the AB stimuli and the X stimuli came from the identification task.

4.3.2.2 ProcedureEach trial began with presentation of a fixation cross on the screen for 1000 ms, followed by auditory presentation of the A, B, and X stimuli with either 250 ms or 1000 ms in between each stimulus. Participants then had 2500ms to indicate whether X sounded more similar to A or B by pressing the corresponding key on a response box. While trials with responses not logged within 2500ms would have been excluded from analysis, all responses were entered in fewer than 2500 ms. Before starting the experimental block, participants completed a 10-trial practice block with feedback and were directed to repeat each incorrect trial a second time before proceeding. No feedback was provided during the experimental blocks.

4.3.2.3 Analysis

14

d’ was calculated by transforming the proportion of hits (correct X=A responses) and false alarms (incorrect X=A responses in X=B trials) to z scores; we converted proportions of 0 and 1 to .01 and .99, respectively, following MacMillan and Creelman (2005). As described in Boley and Lester (2009), false alarms were subtracted from hits and the results were used to determine the corresponding Differencing Model d’ values using Table A5.3 in MacMillan and Creelman (2005, pp. 380-399). The ceiling d’ score was 5.80, and each participant had eight scores (Language (2ENG,BP) x ISI (2250,1000) × Condition (2control, critical)). As with the identification data, RTs from accurate trials were also analyzed after excluding trials with RT shorter than 100ms (nine trials from the English task and six from the BP task). RTs were calculated from the offset of the X stimulus in each trial. Data were again fit to linear mixed models that included by-item (RT only) and by-subject intercepts and slopes across Language, Condition, ISI, Language×Condition, and, in the case of the d’ model, Language×ISI. Pairwise comparisons were Bonferroni corrected and intra-individual Pearson correlations were used to determine the relationships between languages and between ISIs. Alpha was set at .05 for all analyses.

5.0 Results

5.1 Vowel identification task

5.1.1 d’As shown in Table 4, accuracy in the control /a/-/o/ condition is at ceiling with minimal variability. Accuracy approaches ceiling in the /i/ and ø conditions, with a higher rate of variability.

Table 4. Accuracy Rates across Language and Condition

Language /o/-/a/ /i/ ø

M SD M SD M SD

English .98 .06 .94 .10 .91 .16

BP 1.00 .00 .87 .12 .83 .17

The d’ data (Figure 1) illustrates the high rate of variability among the participants in the ø and /i/ conditions, which we discuss in 6.1.

15

Figure 1. d’ sensitivity for /a/-/o/, /i/, and ø conditions in English and BP. Note: Error bars represent 95% confidence intervals.

The model fit to the d’ data yielded a significant Language×Condition interaction (F(2, 28.71) = 3.93); p = .031) and significant main effects for Language (F(1, 14.39) = 6.788; p = .020) and Condition (F(2, 29.21) = 8.95, p = .001). In English, participants’ sensitivity did not differ between any of the conditions (ps ≥ .240). However, in BP, participants’ sensitivity was higher in the /a/-/o/ condition than in the /i/ (p = .001) and ø (p < .001) conditions. There was no difference between languages in the /a/ and /o/ condition (p = .499), while sensitivity was higher in English than in BP in the /i/ (p = .017) and ø (p = .007) conditions. Correlational analyses revealed a significant moderate correlation (r = .569; p = .027) between d’ in BP and English in the ø

condition.

5.1.2 Reaction time

Figure 2. Log transformed RT for the identification task in English and BP. Note: Error bars represent 95% confidence intervals.

In regards to RTs, analysis from the identification data returned a significant interaction of

16

Language×Condition (F(2, 1175.52) = 3.65; p = .026) and main effects of Language (F(1, 15.56) = 12.54; p = .003) and Condition (F(2, 22.91) = 25.06; p < .001).11 In English, reaction times were slower in the ø trials than in the /a/-/o/ trials (p < .001). The /i/ trials logged slower times than the /a/-/o/ trials (p = .001) but no difference was found between the ø and /i/ trials (p = .230). In BP, participants responded more slowly in ø trials than in /a/-/o/ (p < .001) and /i/ (p < .001) trials, and there was no difference between the /a/-/o/ and /i/ trials (p = .513).

5.2 ABX discrimination task

5.2.1 d’ Table 2 illustrates the high rate of accuracy in the Control (VC/a-o/CV – VC/i/CV) and Critical (VC.CV – VC.i.CV) conditions.12 As in the case of the identification task data, accuracy is at ceiling in the Control condition with minimal variability and approaches ceiling in the Critical condition, with higher variability.

Table 5. ABX Accuracy Rates across ISI, Language, and Condition.

Control VC/a-o/CV – VC/i/CV

Critical VCCV – VC/i/CV

Language

ISIM SD M SD

English 250 ms .98 .04 .87 .13

1000 ms .96 .05 .88 .07

BP 250 ms .98 .02 .86 .14

1000 ms .96 .05 .84 .10

A visualization of the d’ scores from the ABX tasks is presented in Figure 3.

11 We limit our discussion of RT to within-language results for two reasons. First, bilinguals are reported to evidence faster RTs in their L1 than L2 (e.g., Fitzpatrick & Izura, 2011); thus, we would expect faster RTs in BP in general. Second, the participants always completed the English task second, which could result in faster RTs in English as a task effect.

12 English 250 ms ISI results are only reported for eight participants; data from the remaining seven were excluded from analysis due to a software error during data collection. BP 250 ms ISI results are reported for 14 participants, as participant 1 did not complete the task.

17

Figure 3. d’ sensitivity scores for Critical and Control conditions in the ABX task in BP and English with 250 ms and 1000 ms ISIs. Note: Error bars represent 95% confidence intervals.

The model fit to the d’ data yielded a significant main effect for Condition (F(1, 13.25) = 56.16; p < .001); pairwise comparisons between conditions were significant in both languages for both ISIs. There was a marginally significant relationship (r = .506; p = .055) between English and BP d’ in the Critical condition in the 1000 ms ISI task.

5.2.2 Reaction timeA visualization of the log-transformed RTs from the ABX tasks in English and BP is presented in Figure 2.

Figure 4. Log transformed RT for Critical and Control conditions in the ABX task in English

18

and BP with 250 ms and 1000 ms ISIs. Note: Error bars represent 95% confidence intervals.

The statistical model yielded a significant three-way Language×ISI×Condition interaction (F(1,1031.33) = 11.12, p = .001). In English there is no difference between conditions in the 250 ms ISI task (p = .411). While participants are faster in the control condition than in the critical condition in the 1000 ms ISI task (p = .005), the nonstandardized effect size of this difference is minimal. Moreover, there was no difference between the two tasks in the critical condition (p = .858) and a very strong significant relationship (r = .880; p = .004). In BP, RT is faster in the control condition than in the critical condition (p < .001) in the 250 ms ISI task (again with a small effect) but not in the 1000 ms ISI task (p = .422). When comparing ISIs in BP, participants were faster in both conditions in the 250 ms ISI task than in the 1000 ms ISI task (control p < .001, critical p = .017) although the nonstandardized effect size is nominal and there was a strong significant relationship between the ISIs in the critical condition (r = .782; p = .001).

6.0 Discussion In the following sections, we first review the results of the identification task and then the ABX tasks. We then proceed with a discussion of their implications for L2 acquisition of phonotactic structure and L1 restructuring, and the consideration of future steps in this line of inquiry. 6.1 Discussion of identification task results

6.1.1 SummaryTo summarize the results, the participants’ accuracy approached ceiling in all three conditions in English. The d’ analysis revealed no difference across conditions in English, although RTs were slower in the ø condition than the other conditions. Similarly to the L2 English data, accuracy approached ceiling in all three conditions in BP, although participants were more sensitive to vowel presence in the /a/-/o/ condition than in either the ø or /i/ conditions, and reaction time data patterned with the d’ data.

6.1.2 L2 perceptionIn light of the explicit nature of the task, our findings suggest that the learners as a group are able to perceive coda stops when relying on metalinguistic strategies. This conclusion gains some traction when we compare our results with monolingual BP identification task data.13 The BP monolinguals in the similar vowel identification task in Dupoux et al. (2011) (also reported in Parlato-Oliveira et al., 2010) were accurate in only 34% of trials in an analogous ø condition, and a similar population in Cardoso (2011) was accurate in 37% of ø trials in a vowel identification task examining word-final stops. In a two-choice identification task, eight BP monolinguals in Azevedo et al. (2017) correctly identified a ø stimulus approximately 60% of the time when the stimulus contained /p/ as coda consonant. Even among advanced L2 English learners tested in Brazil, Cardoso notes that accuracy is still well below target (59%). The accuracy in these studies is substantially lower than the 83% accuracy in the ø trials in our identification experiment. Thus, it appears probable that extensive experience with English in the L2 environment has resulted in the ability to accurately perceive a coda stop in English. This is clear for the 10 participants’ whose d’ scores were at ceiling, while the remaining five have scores

13 Our comparisons are limited to rate of accuracy, as d’ data in these studies were either not provided or were calculated differently.

19

between 1.45 and 2.47. However, even a d’ score of 1.45 is equivalent to an accuracy rate of 60%, which is still much higher than the 34% documented for BP monolinguals in the task after which we pattern ours here (Dupoux et al., 2011). It is of note that, for the bilinguals not at ceiling in English in the identification task, evidence of illusory vowel perception is limited to [i], indicative of L1 BP influence in L2 English perception. Thus, even if there were an illusory vowel in English (which we would predict to be [ə]), there is no evidence of English illusory vowel effects in the participants’ L2 perception (or their L1 perception, for that matter). Considering that the only attested L2 effects on L1 perception in late sequential bilinguals comes from a similar identification task, it is reasonable to assume that, if we were to expect L2 illusory vowel effects, we would have found them in this identification task data.

We recognize that additional processing steps might be required for accurate perception in this task, as indicated by the between-condition difference in RT. The /i/ and ø conditions pattern together in terms of response latency, which leads us to consider the role of metalinguistic awareness. That is, if the participant is metalinguistically aware that illicit codas yield an illusory vowel in BP, they might question whether the /i/ in a VC/i/CV stimulus is part of the signal or if it is illusory. While to our knowledge there is no study of conscious phonotactic awareness in L1 BP/L2 English learners, studies such as Venkatagiri and Levis (2007) reveal a correlation between proficiency and metaphonological awareness. It is therefore logical to follow the assumption that the speakers in the current study have some level of metaphonological awareness, especially when we consider the proficiency of our sample. To better gauge the role of metaphonological knowledge in this result, future studies would benefit from the addition from methods used to tap this type of knowledge, such as debriefings and think-aloud protocols.

6.1.3 L2 influence on L1 perceptionWe posit that the participants’ ability in the L2 to perceive coda stops in this task type is reflected in L1 perception, based on a combination of factors. First, while sensitivity in the critical ø condition was higher in English than in BP, the positive correlation between English and BP d’ indicates that individuals that had higher sensitivity in English, had higher sensitivity in BP. In other words, while there was individual variation, those that more accurately perceived coda stops in English, perceived them similarly in BP. Second, BP accuracy in the ø condition is substantially higher than in any previously published studies. These findings and conclusion align with the correlation between L2 English proficiency and coda perception in a similar explicit task in Alcorn and Smiljanic (2017a). When considering individual data, 10 participants are at ceiling in English, five of whom are also at ceiling in BP. The other five’s BP scores range from 1.45 to 2.47, which suggests a smaller degree of L2 influence while still evidencing BP data that deviates from the monolingual norm. The remaining five participants have scores below ceiling in both languages, ranging from 1.45-2.47 in English and .79-2.58 in BP. Of these five, only the data from the participant with the .79 d’ score (35% accuracy) lines up with the monolinguals in the vowel identification task data in Dupoux et al. (2011). Thus, while we see sizable variation between participants in degree of deviation from the monolingual norm, 14 of 15 demonstrate sensitivity in BP that is potentially reflective of L2 influence.

20

6.2 Discussion of ABX results

6.2.1 SummaryAs in the identification task, L2 English accuracy approaches ceiling in both ABX tasks although learners show higher sensitivity in the control condition (here, a VC/a-o/CV–VC/i/CV pair) than in the critical condition (a VC.CV–V.C/i/.CV pair) in the 1000 ms ISI task. While there is no evidence of additional processing steps needed to perceive coda stops in what is posited to be auditory perception (i.e., in the 250 ms ISI task), the difference in between-condition RTs in the longer ISI could be suggestive of an increased processing burden in what is thought to be phonological perception. However, no RT difference was found between the two ISIs in the critical condition, there was a strong correlation between the two ISIs, and the nonstandardized effect size is minimal across all RT comparisons (see Figure 4), which leads us to question the meaningfulness of the differences in these RT data.14 Thus, this suggestion is tentative.

Sensitivity in both BP tasks aligns with the L2 English data, and this finding is bolstered by the positive marginally significant correlation in the 1000 ms ISI task. The RT data, however, are less clear. The finding that RTs are faster overall in the 250 ms ISI task than in the 1000 ms ISI task could indicate that the processing burden is higher when engaging phonological representation. This would be a logical consequence in the case of simultaneous activation of the L1 and L2 phonological grammars, but it could also be a result of the faster pace of the 250 ms ISI task. However, we are unable to interpret the between-condition difference in the 250 ms ISI task compared with the lack of difference in the 1000 ms ISI task and the between-ISI correlation, which does not follow a logical pattern and again leads us to question the meaningfulness of the RT differences in the ABX data set.

6.2.2 L2 perception Returning to our predictions from section 3.2, we begin with a comparison of the 250 ms ISI accuracy with the monolingual BP accuracy (58%) in the analogous VC.CV–V.C/i/CV condition in an 120 ms ISI ABX task in Dupoux et al. (2011). Based on the assumption that auditory memory is estimated to last 200-300 ms (e.g., Gerrits & Schouten, 2004), we assume that these two ABX tasks with an ISI of < 300 ms should be comparable for the purpose of this study. What we find is that the participants’ accuracy in the current study sharply contrasts with that of the participants in Dupoux et al. Instead, our data align with the 82% accuracy in Matthews and Brown’s (2004) L1 Japanese/L2 English acoustic perception task as predicted, but without the predicted additional processing cost that they found. This difference could be due to the fact that our learners were tested in the L2 environment and arguably had higher proficiency and a more highly activated L2, although we acknowledge the possible task effect given that the English tasks were always completed after the BP tasks. Next, we compare the results of the 1000 ms ISI task with BP monolingual data in Parlato-Oliveira et al. (2010) from a sequence recall task. While we recognize that a sequence recall task should not be directly comparable with the 1000 14 We limit our discussion of effect size for pairwise comparisons to nonstandardized effect sizes, as is not possible to calculate a standardized effect size for a pairwise comparison in an LMM.

21

ms ABX task, we use it as a point of comparison here given that it is BP monolingual data from a task thought to tap phonological processing. We find that the 1000 ms ISI accuracy contrasts with the BP monolingual data in Parlato-Oliveira et al. (2010); participants were accurate in VC.CV-V.C/i/CV trials approximately 24% of the time in a sequence recall task said to reflect phonological processing. Our participants’ accuracy also exceeds the 68% accuracy in the phonological processing task in Matthews and Brown (2004) (an AX task with a 1500 ms ISI), and confirms the prediction that our data would more closely match those of Archibald and Yousefi (2017) from their ABX task with an 800 ms ISI, whose learners were also tested in the L2 environment.15 At an individual level, we see a range in scores in the critical condition in the 250 ms and 1000 ms ISIs (2.31-5.80 and 2.05-5.80, respectively, see Appendix B), with fewer participants at ceiling in both ISIs than in the identification task. That said, even the lower end of the accuracy range in both tasks (69% and 88%, respectively) is greater than the monolingual BP means from the ABX task in Dupoux et al. (2011) and the sequence recall task in Parlato-Oliveira (2010) that serve as a point of comparison. Therefore, we conclude that the learners have developed a level of sensitivity to obstruent codas in phonological perception16 unlike what has been reported for BP monolinguals, and which appears to be on par with existing L2 English data from tasks with comparable ISIs. We take this as confirmation that any lack of L2 influence on L1 perception should not be attributed to the possibility that the learners’ L2 perception is still L1 like.

While there may not be consensus over the nature of the knowledge that is tapped in these different tasks, we cannot overlook that listeners are likely left to rely on abstract representation in the 1000 ms ISI ABX task. That is, we posit that experience with English has led the group to perceive the stop coda in the L2 even after the acoustic detail is no longer available in the input. Recognizing that the data point to residual variation, we posit that, as a group, the learners are approaching convergence on the L2 target. One way of modeling the change to the L2 grammar is by returning to the stochastic model we introduced in section 2.3 in which phonological perception of illusory vowels depends on the interaction of the *[burst]/C(.)/ and *[ ]/i/ cue constraints and *CODA & *[-cont] structural constraint. In this case, we would hypothesize that changes in low-level phonetic processing have driven gradual adjustment over time of the probability distributions of these constraints. Such changes would lead to a target-like phonological grammar in L2 English, although variable outputs could be attested in the case that critical constraints occupy ranges that overlap. If selection points are chosen in the range of overlap during evaluation of a candidate set, an illusory /i/ will be perceived. On the other hand, selection points made outside of the overlap will yield an English-like output. While there is virtually no overlap for those participants whose sensitivity was at ceiling, we posit that others’ 15 Matthews and Brown (2004) and Archibald and Yousefi (2017) do not report sensitivity scores; our only point of comparison is rate of accuracy.16 Our data do not capture the between-ISI difference documented in studies such as Matthews and Brown (2004), which could mean that a) the learners have surpassed an intermediate stage at which auditory perception is L2-like, or b) the two tasks engage the same (phonological) processing strategies regardless of ISI, and phonological perception subsumes auditory perception.

22

grammars likely evidence substantial overlap but are distinct from a BP grammar. An alternative account, which we attribute to an anonymous reviewer, is that these changes in low-level phonetic processing could also be what drives change in metalinguistic judgment. The hypothesis that changes to phonetic processing give way to changes in phonological processing and metalinguistic judgment is an empirical question which we intend to examine longitudinally going forward.

6.2.3 L2 influence on L1 perceptionWhile we are unable to make any conclusions based on the RT data, there is a clear distinction between the L1 BP sensitivity data and the BP monolingual data and this difference is found at both ISIs. Individually, there is a wide d’ range in the critical condition in the 250 ms and 1000 ms ISIs (1.31-5.80 and 2.05-5.80, respectively, see Appendix B), with fewer participants at ceiling in both ISIs than in the identification task. That said, the rate of accuracy even towards the lower end of the range in both tasks is still descriptively higher than the monolingual means that serve as a point of comparison. Between languages, individual data reflect the positive English-BP correlation, with only one case of potential resistance to L2 influence. This participant has a 3.18 d’ in English in the 1000 ms ISI task and .88 in BP. Taken together, the English and BP data reveal that, although there is substantial individual variation, overall, the participants perceive coda obstruents in this task differently than a BP monolingual would be predicted to, and that this difference largely coincides with what we observe in the participants’ L2 interlanguage. If we follow Freeman et al. (2016), one possible explanation for the parallel performance between the participants’ L1 and L2 perception is that these listeners activate their L2 phonotactic constraints (i.e., the L2 ranking of critical cue and structural constraints) during L1 processing in perception, which could explain the parallel between languages exhibited in our data.17 Of note is that the L1-L2 relationship in this data set is suggestive of L2 influence in the use of a range of processing strategies, and is not limited to auditory processing. With that said, the significant positive L1-L2 correlation in the ID task in the critical ø condition, compared with a lack of a significant correlation in the ABX 250 ms ISI task and only a marginally significant correlation in the 1000 ms ISI task suggests that the degree of influence might be greater in metalinguistic judgments than in phonetic and phonological processing. This distinction parallels the late L2 data in Parlato-Oliveira et al. (2010), who reported L2 influence on the L1 in an explicit task but not an implicit task.

6.3 Future directions

A next step is to explicitly test the hypothesis of parallel activation. In addition to a task similar to the phonological priming paradigm implemented in Freeman et al. (2016), we plan to use 17 An anonymous reviewer proposes an alternative hypothesis that (at least a subset of) participants were better able to accurately perceive illicit codas than others at the onset of exposure to L2 English, which would in turn facilitate acquisition of the L2 phonotactic structure. Longitudinal testing starting at the onset of L2 acquisition will allow us to adjudicate between our hypothesis and this alternative hypothesis.

23

event-related potentials (ERPs) to gain further insight into the underlying mechanisms used in the management of two phonological systems (see Wu & Thierry, 2010, for an ERP paradigm that found effects of phonological co-activation). Moreover, we will explore the production of epenthetic vowels in these bilinguals and the relationship between their perception and production. The BiPhon model employs a single grammar to account for listener and speaker behavior, which predicts symmetry in perception and production. While this symmetry has been found in Alcorn and Smiljanic (2017a), asymmetry has been documented for epenthetic vowels in the L2 in Shin and Iverson (2011), which would suggest that the two modes might be governed by different grammars. Our impressionistic observation during data collection in a syllable concatenation task patterns with Shin and Iverson (2011). That is, the majority appear to produce an epenthetic vowel in English and BP. Confirmation of this initial observation would require separate models of perception and production, as discussed in research such as Frazier (2009). It is our hope that these projects will shed further light on the architecture of late sequential bilingual phonological systems.

7.0 ConclusionThis study has examined the perception of word-medial coda stops by L1 BP/L2 English adults in their L1 and L2. Given that these coda stops are illicit in BP and their presence in the auditory input generates an illusory /i/ in perception, we asked whether this illusion can be overcome in the L2, and in turn whether this perception extends to the learner’s L1. To understand the conditions under which the learners are able to perceive coda stops, we implemented a vowel identification task and two ABX tasks. Results from the three tasks indicate that it is possible to overcome the perception of epenthetic /i/ that is generated as a result of illicit phonotactic structure in the L2 English auditory input. Furthermore, this ability, which we posit to be a result of L2 English experience, extends to the L1 such that perception in the L2 correlates with perception in the L1. Data from the range of tasks suggest that, while not fully L2 target-like, this group of bilinguals is better able to perceive coda stops than what previous research would lead us to predict for monolingual BP speakers in both of their languages. This is the case whether they employ what we assume here to be metalinguistic strategies, lower-level encoding strategies, or higher-level encoding strategies in perception. These findings contradict Parlato-Oliveira et al. (2010), who concluded that a non-dominant language can influence L1 metalinguistic strategies, but not phonological encoding.

These results are indicative of the maintenance of plasticity of the speech perception system and phonological grammars in adulthood, in line with previous research in L2 perception (e.g., Darcy et al., 2012) and L1 perceptual restructuring (e.g., Celata & Cancila, 2010, but cf. Parlato-Oliveira et al., 2010). Unlike the outcome in Celata and Cancila (2010), however, the outcome is a gain and not a loss; the current results instead point to a bilingual advantage for these listeners that appear to have (at least partially) acquired this less restrictive phonotactic

24

structure in their L2. Crucially, our data parallel the findings from Carlson et al. (2016), Freeman et al. (2016), and Lee-Ellis (2012) for early bilinguals’ perception, in spite of the fact that the participants’ L2 has been acquired in adulthood, is not their dominant language, and is not target-like for all participants. That is, L2 perception does not need to be fully target-like in order to see these effects. This outcome is the first evidence to our knowledge to demonstrate influence of adult-acquired L2 phonotactics in L1 phonological perception.

1

Appendix A

Duration (ms) of the VC.CV and V.C/i/.CV stimuli from the ID and ABX tasks [mean (SD)]

English StimuliV1 C1

closureC1 VOT /i/ C2

closureC2 VOT V2

[-voi] [+voi] [-voi] [+voi]VC.CV(ID/ABX ‘AB’)

128(25)

71 (46)

14.75 (10)

-87 (13)

N/A 48 (34)

39 (16)

-21 (9)

151 (9)

VC.CV(ABX ‘X’)

118 (19)

44.5 (29.71)

18.5 (12)

-67 (7)

N/A 42 (31)

26 (5)

-65 (3)

113 (17)

V.C/i/.CV(ID/ABX ‘AB’)

120 (19)

42 (37)

36 (22)

-52 (4)

98 (27)

48 (33)

40 (23)

-46 (28)

154 (23)

V.C/i/.CV(ABX ‘X’)

106 (15)

31 (26)

31 (19)

-48 (7)

63 (26)

39 (36)

44 (25)

-45 (13)

109 (22)

BP StimuliV1 C1

closureC1 VOT /i/ C2

closureC2 VOT V2

[-voi] [+voi] [-voi] [+voi]VC.CV(ID/ABX ‘AB’)

152 (34)

41 (39)

13 (7)

-64 (13)

N/A 58 (46)

29 (26)

-63 (9)

101 (29)

VC.CV(ABX ‘X’)

128 (17)

46 (42)

21 (17)

-74 (16)

N/A 48 (42)

29 (21)

-65 (13)

87 (17)

V.C/i/.CV(ID/ABX ‘AB’)

116 (20)

49 (40)

29 (19)

-66 (17)

79 (20)

35 (51)

37 (8)

-68 (24)

105 (14)

V.C/i/.CV(ABX ‘X’)

144 (19)

65 (48)

44 (33)

-78 (18)

80 (12)

80 (40)

33 (18)

-32 (54)

109 (8)

2

Appendix B

Bilingual individual identification task data (d’ sensitivity and raw RT)

Participant BP English

d’ RT (ms) d’ RT (ms)

natural(VC.CV)

/i/(VCi.VC)

natural(VC.CV)

/i/(VC.i.CV)

natural(VC.CV)

/i/(VCi.VC)

natural(VC.CV)

/i/(VCi.CV)

P01 2.58 1.97 1529 769 1.97 1.97 656 997

P02 4.01 4.01 1233 520 4.01 4.01 1029 533

P03 1.97 1.97 1472 968 1.55 2.47 954 1602

P04 4.01 4.01 633 831 4.01 4.01 410 342

P05 1.97 1.97 868 777 2.47 4.01 666 681

P06 1.97 1.97 1446 876 4.01 4.01 823 633

P07 1.97 4.01 540 562 4.01 1.97 933 1029

P08 4.01 2.58 933 689 4.01 4.01 508 305

P09 0.79 4.01 2336 552 1.45 4.01 1555 944

P10 4.01 1.97 863 408 4.01 4.01 527 432

P11 1.97 4.01 605 534 4.01 4.01 452 379

P12 1.97 1.97 1102 938 4.01 4.01 677 576

P13 2.58 4.01 508 548 4.01 4.01 372 399

P14 4.01 1.97 870 1016 4.01 1.97 557 534

P15 1.97 2.58 640 484 2.47 4.01 490 680

P16a 4.01 -1.59 -- 480 0.35 4.01 1739 1236

Note: d’ ceiling is 4.01. aParticipant excluded from analysis based on negative d’ score

3

Bilingual individual ABX Data (critical condition d’ sensitivity and raw RT in ms)

Participant BP English

ABX 250 ABX 1000 ABX 250 ABX 1000

d’ RT d’ RT d’ RT d’ RT

P01 -- -- 2.44 857 -- -- 2.50 1041

P02 2.21 697 4.77 676 4.77 652 5.80 513

P03 1.31 840 0.88 841 -- -- 3.18 768

P04 3.83 544 3.39 553 4.77 514 5.80 556

P05 2.05 533 1.83 813 -- -- 2.89 546

P06 2.89 637 2.50 557 2.72 561 3.11 541

P07 2.31 437 2.72 481 -- -- 2.05 501

P08 4.77 513 3.83 606 5.80 651 4.77 538

P09 3.39 534 3.39 860 -- -- 2.89 889

P10 2.99 626 3.11 554 3.11 579 5.80 613

P11 4.77 317 4.77 332 4.29 340 3.39 338

P12 4.29 728 2.70 737 -- -- 2.72 578

P13 4.77 335 2.99 409 -- -- 3.39 454

P14 4.29 785 2.50 778 2.31 983 3.11 768

P15 5.80 436 2.72 422 4.29 525 2.70 409

P16a 0.64 515 0.64 881 -- -- 2.51 488

Note: d’ ceiling is 5.80. aParticipant excluded from analysis based on negative d’ score in identification task.

4

Appendix C

L1 English data18

L1 English identification task

Accuracy across conditions/o/-/a/ /i/ ø

M SD M SD M SD

1.00 0.00 0.97 0.08 0.89 0.21

The descriptively lower accuracy and d’ sensitivity in the ø condition is due to one participant (P21; accuracy = .38, d’ = .79), who selected a response of <u> in four trials and <o> in one trial. Linear mixed models with a random by-subject intercept, and with a by-subject intercept and slope across Condition did not converge. A linear model with the fixed effect Condition and the target variable of d’ did not yield a significant main effect (F(2, 24) = 2.049, p = .151).

L1 English individual identification task data (d’ sensitivity and raw RT)d’ RT (ms)

natural(VC.CV)

/i/(VCi.VC)

natural(VC.CV)

/i/(VCi.CV)

P17a 1.97 2.58 1246 959

P18 4.01 4.01 1514 1474

P19 4.01 1.97 851 447

P20 1.97 4.01 642 829

P21 0.79 4.01 1333 1029

P22 2.58 4.01 942 701

P23 4.01 4.01 417 175

P24 4.01 4.01 966 462

P25 4.01 4.01 843 1199

P26 4.01 4.01 896 1017

18Given the outcome of the bilingual reaction time data, we limit our statistical analysis of the L1 English control data to d’ sensitivity.

5

Note. d’ ceiling is 4.01. aParticipant excluded from group-level analysis based on accuracy in control condition (< 80%). L1 English ABX data

Accuracy across conditionsControl VC/a-o/CV – VC/i/CV

Critical VCCV – VC/i/CV

ISIM SD M SD

250 ms 0.98 0.05 0.98 0.03

1000 ms 0.94 0.06 0.98 0.02

A linear mixed model with the fixed effects ISI and Condition, random by-subject slopes across ISI and Condition, and the target variable of d’ did not yield any significant main effects (ISI F(1, 7.99) = .19, p = .676; Condition F(1, 9.27) = 1.41, p = .265. The ISI×Condition condition was significant F(1, 8.69) = 7.45; p = .024), although Bonferroni corrected pairwise comparisons reveal that the interaction is likely due to the difference between conditions in the 1000 ms ISI task: Accuracy was higher in the critical condition than in the control condition (p = .030). Given the direction and magnitude of the difference, we do not consider this difference to be meaningful.

L1 English individual ABX Data (critical condition d’ sensitivity and raw RT in ms)Participant English

ABX 250 ABX 1000

d’ RT d’ RT

P17 5.80 552 5.80 688

P18a 3.83 1101 4.26 947

P19 5.80 427 5.80 464

P20 3.96 248 5.80 293

P21 5.80 507 4.78 711

P22 3.83 682 4.71 618

P23 5.80 283 5.80 357

P24 5.80 307 5.80 472

6

P25 3.83 519 5.80 689

P26 5.80 214 5.80 266

Note: d’ ceiling is 5.80. aParticipant removed from group-level analysis based on accuracy in control condition (< .80).

7

Reference list1. Alcorn, S., & Smiljanic, R. (2017a, January). The role of L2 phonetic experience in L1

phonological restructuring in Portuguese-English bilinguals. Paper presented at Bilingualism in the Hispanic and Lusophone World, Tallahassee, Florida.

2. Alcorn, S., & Smiljanic, R. (2017b, October). The production of epenthetic vowels by late Brazilian Portuguese English bilinguals. Paper presented at the Hispanic Linguistics Symposium, Lubbock, Texas.

3. Archibald, J. & M. Yousefi (2018, March). The redeployment of marked L1 Persian codas in the acquisition of marked L2 English onsets: Redeployment as a transition theory. Paper presented at ConCALL 3. University of Indiana.

4. Azevedo, R., Matzenauer, C., & Alves, U. (2017). Formalização fonético-fonológica da interação de restrições na percepção da epêntese vocálica no português brasileiro. ReVel, 15, 289-313.

5. Birdsong, D., & Gertken, L. M. (2013). In faint praise of folly: A critical review of native/non-native speaker comparisons, with examples from native and bilingual processing of French complex syntax. Language, Interaction and Acquisition, 4(2), 107-133.

6. Birdsong, D., Gertken, L. M., & Amengual, M. (2012). Bilingual Language Profile: An easy-to-use instrument to assess bilingualism. COERLL: University of Texas at Austin.

7. Bley-Vroman, R. (1989). What is the logical problem of foreign language learning? In S. Gass & E. Schachter (Eds.), Linguistic perspectives on second language acquisition (pp. 41-68). Cambridge: Cambridge University Press.

8. Boersma, P. (2007). Some listener-oriented accounts of h-aspiré in french. Lingua, 117, 1989-2054.

9. Boersma, P., & Hamann, S. (2009). Loanword adaptation as first-language phonological perception. In A. Calabrese & W. L. Wetzels (Eds.), Loan phonology (pp. 11-58). Amsterdam: John Benjamins.

10. Boersma, P., & Hayes, B. (2001). Empirical tests of the gradual learning algorithm. Linguistic Inquiry, 32, 45-86.

11. Brien, C., & Sabourin, L. (2014). Age of L2 acquisition has a greater influence on L1 metalinguistic awareness than proficiency. Paper presented at the meeting of the Canadian Linguistic Association, St. Catharines, Ontario.

12. Broselow, E., & Kang, Y. (2013). Second language phonology and speech. In J. Herschensohn & M. Young-Scholten (Eds.), The Cambridge handbook of second language acquisition (pp. 529-554). Cambridge: Cambridge University Press.

13. Cabrelli Amaro, J. (2017). Testing the Phonological Permeability Hypothesis: L3 phonological effects on L1 versus L2 systems. International Journal of Bilingualism, 21, 698-717.

14. Cardoso, W. (2011). The development of coda perception in second language phonology: A variationist perspective. Second Language Research, 27, 433-465.

8

15. Carlson, M. T. (2018). Making room for second language phonotactics: Effects of L2 learning and environment on first language speech perception. Advance online publication. Language and Speech. doi:10.1177/0023830918767208

16. Carlson, M. T., Goldrick, M., Blasingame, M., & Fink, A. (2016). Navigating conflicting phonotactic constraints in bilingual speech perception. Bilingualism: Language and Cognition, 19, 939-954.

17. Celata, C., & Cancila, J. (2010). Phonological attrition and the perception of geminate consonants in the Lucchese community of San Francisco (CA). International Journal of Bilingualism, 14, 1-25.

18. Chang, C. (2012). Rapid and multifaceted effects of second-language learning on first-language speech production. Journal of Phonetics, 40, 249-268.

19. Cuetos, F., Hallé, P. A., Dominguez, A., & Segui, J. (2011). Perception of prothetic /e/ in #sC utterances: Gating data. In Proceedings of the 17th International Congress of Phonetic Sciences (Vol. 17-21, pp. 540-543). Hong Kong, China: 17th ICPhS.

20. Darcy, I., Dekydtspotter, L., Sprouse, R., Glover, J., Kaden, C., McGuire, M., & Scott, J. (2012). Direct mapping of acoustics to phonology: On the lexical encoding of front rounded vowels in L1 English–L2 French acquisition. Second Language Research, 28, 5-40.

21. de Jong, K., & Park, H. (2012). Vowel epenthesis and segment identity in Korean learners of English. Studies in Second Language Acquisition, 34, 127-155.

22. de Leeuw, E., Mennen, I., & Scobbie, J. (2013). Dynamic systems, maturational constraints and L1 phonetic attrition. International Journal of Bilingualism, 17, 683-700.

23. de Leeuw, E., Mennen, I., & Scobbie, J. (2012). Singing a different tune in your native language: First language attrition of prosody. International Journal of Bilingualism, 16, 101-116.

24. de Leeuw, E., Tusha, A., & Schmid, M. (2017). Individual phonological attrition in Albanian–English late bilinguals. Bilingualism: Language and Cognition, 1-18.

25. Dupoux, E., Kakehi, K., Hirose, Y., Pallier, C., & Mehler, J. (1999). Epenthetic vowels in Japanese: A perceptual illusion? Journal of Experimental Psychology, 25, 1568-1578.

26. Dupoux, E., Parlato, E., Frota, S., Hirose, Y., & Peperkamp, S. (2011). Where do illusory vowels come from? Journal of Memory and Language, 64, 199-210.

27. Fitzpatrick, T., & Izura, C. (2011). Word association in L1 and L2: An exploratory study of response types, response times, and interlingual mediation. Studies in Second Language Acquisition, 33, 373-398.

28. Flege, J. (1987). A critical period for learning to pronounce foreign languages? Applied Linguistics, 8, 162-177.

29. Frazier, M. (2009). The production and perception of pitch and glottalization in Yucatec Maya (Unpublished doctoral dissertation). The University of North Carolina at Chapel Hill, North Carolina.

30. Freeman, M. R., Blumenfeld, H., & Marian, V. (2016). Phonotactic constraints are activated across languages in bilinguals. Frontiers in Psychology, 7.

9

31. Gerrits, E. (2001). The categorisation of speech sounds by adults and children:A study of the categorical perception hypothesis and the development weighting of acoustic speech cues Unpublished Doctoral Dissertation). Utrecht University, Utrecht, Netherlands.

32. Gerrits, E., & Schouten, M. (2004). Categorical perception depends on the discrimination task. Perception & Psychophysics, 66, 363-376.

33. Hopp, H., & Schmid, M. S. (2013). Perceived foreign accent in first language attrition and second language acquisition: The impact of age of acquisition and bilingualism. Applied Psycholinguistics, 34(2), 361-394.

34. John, P., & Cardoso, W. (2017). Estrutura silábica e variação fonológica. Ilha do Desterro, 70, 169-184.

35. Kabak, B., & Idsardi, W. (2007). Perceptual distortions in the adaptation of English consonant clusters: Syllable structure or consonantal contact constraints? Language and Speech, 50, 23–52.

36. Kang, Y. (2011). Loanword phonology. In M. van Oostendorp, C. Ewen, E. Hume, & K. Rice (Eds.), The Blackwell companion to phonology (pp. 2258-2281). Hoboken: Wiley-Blackwell.

37. Koerich, R. (2006). Perception and production of vowel paragoge by Brazilian EFL students. In B. Baptista & M. Watkins (Eds.), English with a Latin beat: Studies in Portuguese/Spanish–English interphonology (pp. 91-104). Amsterdam: John Benjamins.

38. Lee-Ellis, S. (2012). Looking into bilingualism through the heritage speaker's mind. (Unpublished doctoral dissertation). University of Maryland, College Park, Maryland.

39. Lord, G. (2008). Second language acquisition and first language phonological modification. In J. Bruhn de Garavito & E. Valenzuela (Eds.), Proceedings of the 10th Hispanic Linguistics Symposium (pp. 184-193). Somerville, MA: Cascadilla Proceedings Project.

40. Macmillan, N., & Creelman, C. (2005). Detection theory: A userʼs guide. Mahwah, NJ: Lawrence Earlbaum.

41. Major, R. (1992). Losing English as a first language. The Modern Language Journal, 76, 190-208.

42. Major, R. (2010). First language attrition in foreign accent perception. International Journal of Bilingualism, 14(2), 163-183.

43. Matthews, J., & Brown, C. (2004). When intake exceeds input: Language specific perceptual illusions induced by L1 prosodic constraints. International Journal of Bilingualism, 8, 5-27.

44. Mayr, R., Price, S., & Mennen, I. (2012). First language attrition in the speech of Dutch–English bilinguals: The case of monozygotic twin sisters. Bilingualism: Language and Cognition, 15, 687-700.

45. Mennen, I. (2004). Bi-directional interference in the intonation of Dutch speakers of Greek. Journal of Phonetics, 32, 543-563.

46. Parlato-Oliveira, E., Christophe, A., Hirose, Y., & Dupoux, E. (2010). Plasticity of illusory vowel perception in Brazilian-Japanese bilinguals. Journal of the Acoustical

10

Society of America, 127, 3738-3748. 47. Prince, A., & Smolensky, P. (1993) [2004]. Optimality Theory: Constraint interaction in

generative grammar. Malden, MA: Blackwell.48. Schmid, M. (2016). First language attrition. Language Teaching, 49, 186-212. 49. Schwartz, B., & Sprouse, R. (1996). L2 cognitive states and the full transfer/full access

model. Second Language Research, 12, 40-72. 50. Selkirk, E. (1982). Syllables. In H. van der Hulst & N. Smith (Eds.), The structure of

phonological representations (Vol. 2, pp. 337-383). Dordrecht: Foris.51. Shin, D.-J., & Iverson, P. (2010). Individual differences in vowel epenthesis among

korean learners of english. Journal of the Acoustical Society of America, 128, 2488. 52. Silveira, R. (2011). Pronunciation instruction and syllabic-pattern discrimination.

DELTA, 27, 21-36. 53. Venkatagiri, H., & Levis, J. (2007). Phonological awareness and speech

comprehensibility: An exploratory study. Language Awareness, 16, 263-277. 54. Wayland, R., & Guion, S. (2003). Perceptual discrimination of thai tones by naïve and

experienced learners of Thai Applied Psycholinguistics, 24, 113-129. 55. Wu, Y., & Thierry, G. (2010). Chinese–English bilinguals reading English hear

Chinese. Journal of Neuroscience, 30, 7646-7651.56. Yang, J., & Fox, R. A. (2017). L1–L2 interactions of vowel systems in young bilingual

Mandarin-English children. Journal of Phonetics, 65, 60-76.