1.1. Bengali and English in India

Vowel systems

The vowel inventory of Western Bangla consists of [i, e, æ, a:, ɔ, o, u], and their nasalized counterparts (Rubino & Garry, Reference Rubino and Garry2001). The vowel chart in Figure 1 (from Ghosh, 2012) shows their distribution in the F1×F2 formant space. Note that there is no mid-central vowel category.The vowel system of IE contains the monophthongs [I, i, E, e, æ, ə/ʌ/ɜ, a, ɔ, o, ʊ, u] (Masica, Reference Masica1972; Wells & Wells, Reference Wells and Wells1982). The vowel chart in Figure 2 shows their distribution in the F1×F2 formant space. A single mid-central vowel corresponds to the categories [ʌ,ə,ɜ], which are treated as distinct in many varieties of English (Bansal, Reference Bansal1969; Hickey, Reference Hickey2005; Nihalani, Tongue, Hosali, & Crowther, Reference Nihalani, Tongue, Hosali and Crowther1979; Wells & Wells, Reference Wells and Wells1982). Since the English items used in this study are traditionally transcribed with [ʌ], we use this symbol to indicate the mid-central vowel throughout.

Fig. 1. Vowel chart for oral monophthongs in Bengali, from Ghosh (Reference Ghosh2016)

Fig. 2. Vowel chart for monophthongs in General(ized) Indian English, from Masica (Reference Masica1972)

1.2. What causes transfer and what does it affect?

Existing research has distinguished changes in category representations due to the acquisition of multiple sound systems (e.g., Speech Learning Model – Flege, Reference Flege1995, Reference Flege2007; Perceptual Assimilation Model-L2 – Best & Tyler, Reference Best, Tyler, Bohn and Munro2007), from interaction during accessing, processing, or articulation of these categories (transfer vs. interference (Grosjean, Reference Grosjean2012); competence vs. performance interference (Paradis, Reference Paradis1993)). Given their transient nature, dynamic changes in production during mixed-language use are generally attributed to the latter, e.g., online processing costs (Olson, Reference Olson2013; Šimáčková & Podlipskỳ, Reference Šimáčková and Podlipskỳ2015; Tsui, Tong, & Chan, Reference Tsui, Tong and Chan2019, VOT), language mode (Simonet, Reference Simonet2014, vowel quality), context-awareness (Khattab, Reference Khattab2013, phonological variables). What triggers this interaction? Olson (Reference Olson2016) argues that cross-language phonetic effects could have two potential sources: (i) the local point of switch, and (ii) global co-activation of two languages: bilingual language mode (Grosjean, Reference Grosjean1998).

A number of studies have specifically manipulated language mode, both in the presence and absence of switching. Overall, results suggest that:

1. (i) in the absence of other manipulations, productions in a bilingual language mode show increased cross-language influence compared to a monolingual mode (Simonet, Reference Simonet2014; Simonet & Amengual, Reference Simonet and Amengual2020, vowel quality);

2. (ii) (however, language mode is not the sole source of influence during mixed language use – studies comparing switched and nonswitched tokens produced in the same test block (identical language mode) (Olson, Reference Olson2016; Tsui et al., Reference Tsui, Tong and Chan2019, VOT), or spontaneous conversation (Piccinini & Arvaniti, Reference Piccinini and Arvaniti2015, VOT), have still reported a difference, suggesting that independently of mode, switching between languages triggers a local increase in cross-language transfer;

3. (iii) how the two sources interact to influence the final outcome of transfer is not fully understood: Olson (Reference Olson2016, VOT) found no additive effects of language mode, Olson (Reference Olson2013, VOT) found a balanced language context to inhibit transfer compared to unbalanced contexts. Other studies have not analyzed the two separately, eliciting switched tokens in a bilingual test block and nonswitched tokens in a monolingual test block, separated by a few hours to days (Elias et al., Reference Elias, McKinnon and Milla-Muñoz2017, vowel quality), (Antoniou, Best, Tyler, & Kroos, Reference Antoniou, Best, Tyler and Kroos2011; Bullock & Toribio, Reference Bullock and Toribio2009; Schwartz, Balas, & Rojczyk, Reference Schwartz, Balas and Rojczyk2015; Šimáčková & Podlipskỳ, Reference Šimáčková and Podlipskỳ2015, Reference Šimáčková and Podlipskỳ2018, VOT).

Since participants of this study are in an environment containing mixed-language input and multilingual interlocutors, we expect that phonetic transfer during everyday language use takes place in a bilingual mode (see Grosjean, Reference Grosjean1998). We elicited all utterances in the same bilingual test block. Observed differences in this paradigm would arguably result from interaction during online processing of sounds while switching between languages. In bilingual mode, both language systems are expected to be nearly equally accessible throughout the test block. Thus, a consistent difference between switched and nonswitched tokens in such a paradigm would be possible only if the effects were highly localized – if not, we should expect a gradual convergence over the course of the experiment. This is discussed in the next section.

1.3. Duration of transfer effects

Longitudinal studies and between-subject comparisons of speakers with different durations of L2 exposure suggest that interaction between bilingual sound systems is dynamic (e.g., Bohn & Flege, Reference Bohn and Flege1992). Changes due to transfer are not necessarily unidirectional or irreversible. Sancier and Fowler (Reference Sancier and Fowler1997, VOT) first demonstrated that spending 2-5 months in an L1 or L2 environment causes productions in both languages to “drift” towards the ambient language, showing that cross-language interaction can be triggered in an order of months, and reversed within a similar time range. A recent study by Tobin, Nam, and Fowler (Reference Tobin, Nam and Fowler2017, VOT) reported comparable effects in an even shorter duration (2-4 weeks). Based on a short-term longitudinal study, Chang (Reference Chang2012, VOT) demonstrates that over the course of the first five weeks of learning an L2, there is a gradual convergence of L1 towards L2. For some sounds and features (but not others, see sec.1.5), transfer was additive over time. How long these effects last in the absence of regular L2 input was not tested.

Studies discussed above concern situations where a speaker is operating in any one of their languages. For mixed-language use, the majority of existing studies only analyze the switched (target) token – there are few direct measurements of the duration of short-term transfer effects. One study which measured this in a code-switching paradigm (Bullock & Toribio, Reference Bullock and Toribio2009, VOT) did not find any residual effects on the matrix language following a switch, suggesting that transfer during code-switching is localized. Indirect evidence for this comes from experiments that have elicited switched and nonswitched tokens in the same test block and still reported differences between the two (e.g., Olson, Reference Olson2013; Tsui et al., Reference Tsui, Tong and Chan2019, VOT), suggesting that changes due to transfer are quickly ‘reset’ – in an order of seconds. Note that all the studies of mixed-language production discussed above measure VOT, which is a temporal feature. We have no a priori reason to assume that these durations generalize to vowel quality. However, findings from sub-categorical phonetic shifts triggered by other factors (such as convergence towards an interlocutor) do evidence rapid shifts in vowel quality within comparable time-frames (e.g., Babel, Reference Babel2010, Reference Babel2012; Pardo, Reference Pardo and Morsella2010). In this study, we present switched and nonswitched tokens randomly within the same test block to induce a bilingual mode. We expect the intervening words between two subsequent targets to undo any residual effects of transfer.

1.4. Asymmetries between languages in extent and direction of transfer

A recurring pattern in the literature on short-term phonetic transfer is that the extent and patterns of phonetic shift are not equivalent across the two languages of the bilingual speaker. Studies vary greatly in the exact differences they report. The sources of such asymmetry are of interest because they point towards the factors that mediate cross-language phonetic interaction. In this section, we discuss two proposed sources of such asymmetry that are relevant to the present study:

1.5. Asymmetries between sounds

Asymmetries in the extent and patterns of transfer have not only been observed between languages, but also between different sounds/features of a language, in both long-term and transient interactions. Studies that have examined multiple sound categories have found that interactions between individual sound pairs do not necessarily parallel the overall pattern of global (system-wide) shift (e.g., Chang, Reference Chang2012; Elias et al., Reference Elias, McKinnon and Milla-Muñoz2017, vowel quality), suggesting that ‘extent of transfer’ cannot treated as an atomic measure. In light of the discussion about VOT in section 1.4, this is not surprising – if a general linguistic principle of ‘room for movement’ and contrast constrains transfer, then we should expect it to apply to individual sounds too. Once again, this emphasizes the importance of examining a wider range of sound contrasts.

In the present study, we focus on two vowel categories in Indian English: the mid-central vowel [ʌ] and the low front vowel [æ]. There are two plausible sources of asymmetry between these:

1. 1. Position in the IE vowel space: compared to [æ], [ʌ] exists in a part of the vowel space that has lower vowel density. This affords a greater latitude for movement without risking the loss of a contrast, particularly in the vowel height (F1) dimension. Thus, considering purely phonological constraints on IE, we expect that a greater degree of shift is possible in [ʌ] compared to [æ]. Any movement in [æ] is expected to be primarily in the backness (F2) dimension.

2. 2. Target of transfer: existing research on cross-language phonetic interaction suggests that changes in production during mixed-language use are not random, but rather targeted with respect to categories in the other language. Thus, another source of asymmetry is the fact that the category [ʌ] is absent in Bengali, whereas [æ] is a common category across both languages (see sec. 1.1).

Flege's Speech Learning Model (SLM) (Reference Flege1995, Reference Flege2007) posits that sound categories which are common across languages influence each other because they share a common acoustic-phonetic space. Thus, if Bengali and English differ in their canonical realizations of [æ], then we should expect the English [æ] to shift towards the corresponding Bengali category in the mixed condition.

There is no obvious competing L1 category during the production of [ʌ]. However, it is unlikely that this should altogether preclude a shift in [ʌ], since at least one existing study has reported transfer effects on a non-common vowel category in a comparable paradigm (Simonet, Reference Simonet2014). Here, participants’ Catalan [ɔ] shifted towards an acoustically close category [o], which is common to Catalan and Spanish, suggesting that movement might be towards a set of phonetic norms rather than a specific corresponding category.

We expect the vowel [ʌ] to shift towards an acoustically/perceptually close category in IE – a probable candidate being the low vowel [a:], as anecdotal observations of Bengali-accented English suggest that the sounds are perceived as close by Bengali speakers.

Our research questions in this study concern whether L2 vowels are produced differently in nonswitch vs switch conditions. Existing literature suggests that such shifts are directional and targeted. Although our primary concern is whether productions in the two conditions differ, we wanted to understand how the direction of shift, if any, relates to corresponding Bengali vowel categories. To estimate average baseline differences in vowel quality between the relevant English and Bengali categories, we processed and used existing speech data from a freely available Bengali corpus, SHRUTI (Das, Mandal, & Mitra, Reference Das, Mandal and Mitra2011).

We extracted formant values from syllable-initial productions of [æ] and [a:] at 55% into the vowel, and compared these to our participants’ productions in the nonswitched English condition. This is visualized in Figure 3. The F1×F2 plot indicates that the Bengali category [æ] is more fronted, and [a:] lower and more fronted, compared to the English categories [æ] and [ʌ] respectively.

Figure 3: Participants’ unilingual English vowels (shaded ellipses) compared to Bengali categories from the SHRUTI corpus (Das et al., Reference Das, Mandal and Mitra2011)

1.6 A note on paradigms: connected speech or not?

This section motivates a secondary aim of this study: to verify if the extent and patterns of phonetic interaction in bilingual participants are independently mediated by the choice of experimental paradigm. Studies of phonetic transfer during mixed-language use have largely employed spontaneous or scripted code-switching (CS) to elicit mixed productions. CS is defined as the use of multiple languages within a single utterance, i.e., in connected speech (Myers-Scotton, Reference Myers-Scotton1993). Other paradigms have included cued picture-naming, delayed repetition, interpreting across languages, reading word-lists etc. Olson (Reference Olson2013) argues that code-switching introduces discursive factors such as discourse planning and pragmatic considerations, which ultimately reflect patterns of language practice, rather than cognitive behavior. This finds support from studies that report transfer effects before the switch point (Bullock & Toribio, Reference Bullock and Toribio2009), or other evidence for planning such as hyperarticulation (Muldner et al., Reference Muldner, Hoiting, Sanger, Blumenfeld and Toivonen2019) and interlocutor-awareness in spontaneous CS (Khattab, Reference Khattab2013). To avoid such confounds, Olson (Reference Olson2013) endorses cued picture-naming (where pictures are named in rapid succession in either language, based on the given language cue) as an alternative paradigm to isolate purely phonetic effects in controlled setups. This observation has important implications for how the existing literature on short-term phonetic interaction is interpreted: if the discursive properties of a CS paradigm independently affect the phonetic outcome of mixed-language use, this calls to question the comparability of results across studies using these different paradigms. Because the reported outcomes of short-term phonetic interaction vary so greatly across studies (see section 1.4), generalizations about the mechanisms that underlie these patterns often rely crucially on the assumption of such comparability. Therefore, identifying paradigm-specific differences is particularly relevant for this line of research.

Note that while there is evidence for planning in CS, precisely how this affects phonetic interaction cannot be predicted from the existing literature: while Bullock and Toribio (Reference Bullock and Toribio2009) found tokens immediately preceding the switch point to be phonetically distinct from the rest, one participant group converged towards the switch language (explained in terms of proficiency) while the other showed divergence (explained in terms of extra-linguistic factors). The participants in Muldner et al.'s (2019) study hyperarticulated switched tokens (indicating some degree of planning), and did not evidence a significant shift during mixed-language use. Moreover, recall that the findings of Olson (Reference Olson2016) suggest that there is a limit on the extent of transfer – multiple factors do not necessarily result in additive effects. Thus, the presence of additional factors in CS does not necessitate that this should alter the extent of observable transfer compared to any other paradigm. It is possible that the outcomes in the studies discussed above are better attributed to some other aspect of the experimental setup, rather than the act of code-switching in particular.

The precise contribution of various (linguistic and extra-linguistic) factors and their interactions to the phonetics mixed-language use merits detailed investigation in its own right. In the present study, however, our focus is limited to verifying whether, in an experimental setting, the postulated discursive elements of a CS paradigm alter the extent of observable transfer compared to a paradigm that arguably lacks these: cued picture-naming. We are less concerned with the exact nature of differences than with their comparability. While researchers have successfully employed each of these paradigms to induce phonetic interaction, comparison across studies is difficult since they also differ in other factors. We provide a starting point for such comparison by eliciting productions of the same tokens from the same set of participants, in both paradigms. A difference in results will suggest that the planning involved in connected speech and code-switching affects transfer above and beyond the phonetic outcome of switching itself, which has implications that future work should consider.

Research questions and hypotheses

RQ1 – L2 vowels: Are L2 vowel categories of proficient bilingual speakers affected by dynamic phonetic transfer during mixed-language production?

H1 – L2 vowels: We expect to see differences in the spectral properties of L2 vowels in switched vs nonswitched tokens

RQ2 – Asymmetry: Do patterns of transfer differ between the two vowel categories?

H2 – Asymmetry: Since existing literature demonstrates targeted and phonologically-constrained shifts, we expect transfer effects to be category-specific. Based on the position of IE vowels compared to those of Bengali vowels from Das et al. (Reference Das, Mandal and Mitra2011), we expect [ʌ] to be lowered and fronted in the mixed condition. [æ] is expected to be fronted. We hypothesize the extent of shift in [æ] to be lesser than [ʌ], as it is in a part of the IE vowel space that has greater vowel density.

RQ3 – Paradigm: Is there an observable difference between the extent of transfer in a cued picture-naming vs. a code-switching paradigm?

H3 – Paradigm: While researchers have discussed code-switching as a source of discursive factors likely to affect transfer behavior, this has not been systematically studied. Existing literature does not allow us to predict precisely what form this effect might take (see discussion in Section 1.6). Thus, this question is exploratory, and meant to serve as a starting point for more within-study comparisons.

2.1. Participants

Participants were recruited through an open call at the EFL University campus in Hyderabad. Volunteers completed a brief language background questionnaire, which was used to screen participants.

The university campus has a linguistically diverse population with varying L1s and varieties of English. Our aim was to minimize between-participant random variability, while maintaining ecological validity considering the actual day-to-day settings in which transfer behaviors likely take place. We chose two metrics for this: (i) where the participant grew up; and (ii) Medium of Instruction (MoI) in school. The first metric ensures that the variety of Bengali that individuals were exposed to is similar across participants, resulting in similar target representations for Bengali vowel categories. This was especially important because our study does not elicit unilingual Bengali vowels, instead using existing corpus data to estimate the position of Bengali vowels in the formant space. Since we expect L2 vowels to shift towards related L1 categories during mixed language utterances, we wanted to ensure that any between-participant asymmetries reflected individual differences in transfer, rather than different acoustic targets. Another reason to control for location was that it is a fairly reliable predictor for language(s) of education (see supplementary materials for details). The second metric reflects sustained formal education in a language, which predicts LSRW (listening, speaking, reading, writing) skills. Since this study uses written stimuli, we wanted to ensure that all participants were sufficiently comfortable with reading and speaking both Bengali and English, to avoid variability due to task effort.

Given these considerations, we recruited participants who (i) grew up and lived in the state of West Bengal for the majority of their lives; (ii) had Bengali-speaking parents; (iii) received at least 5 years of formal education in Bengali (all respondents had 5+ years of education in English and were pursuing degrees taught in English); and (iv) self-reported LSRW proficiency in Bengali.

28 bilingual speakers of Bengali and English (18 female) responded to the recruitment call. Using the inclusion criteria reported above, 10 volunteers (5 female, 5 male, age range 19–28) participated in the study, and were compensated for their time. We followed up with a detailed language background survey at a later date to learn more about the language profiles of the participants. This survey was a modified and consolidated version of three popular bilingualism profiling tools: the Bilingual Language Profile (BLP) (Birdsong, Gertken, & Amengual, Reference Birdsong, Gertken and Amengual2012), the Language Experience and Proficiency Questionnaire (LEAP-Q) (Marian, Blumenfeld, & Kaushanskaya, Reference Marian, Blumenfeld and Kaushanskaya2007), and the Bilingual Switching Questionnaire (BSWQ) (Rodriguez-Fornells, Kramer, Lorenzo-Seva, Festman, & Münte, Reference Rodriguez-Fornells, Kramer, Lorenzo-Seva, Festman and Münte2012). Together, these incorporate questions about language history, usage, self-assessed proficiency, attitudes, and switching experience. Table 1 summarizes some key biographical data from the survey. The complete questionnaire, detailed discussion of the linguistic situation in India, and language profiles of participants are included as Appendix S1 in the OSF repository for this study.

Table 1. Key biographical information summarized across participants. Scales: age 0(since birth)—5(since age 5) etc.; likert scale: 0(not proficient at all)—6(highly proficient)

Although existing studies in the field have used comparable participant numbers, we acknowledge that this is a small sample size. As detailed above, even after a recruitment call that already controlled for language background and education level, we were able to include only 10 out of 28 volunteers (35.7%) in the final study. The criteria used here establish only a minimal level of between-participant similarity in exposure, dominance, habits, and proficiency, such as is extremely common in most comparable studies. This demonstrates the degree of linguistic variability in the present population. Norms for experimental research on bilingualism have largely been established in a Western context. In settings where linguistic variability at the population level is much greater than these canonically studied groups, establishing comparable levels of experimental control requires significantly greater resources. Given that small sample sizes are likely to have low statistical power, replication with larger sample sizes are important to confirm the observed effects, and provide much-needed data diversity. A discussion of power analyses in bilingualism research is included as Appendix S2 in the OSF repository.

2.2. Stimuli

Twenty monosyllabic English words, ten each containing the vowels [ʌ] and [æ], were selected as target words. To minimize coarticulatory effects on the vowel, pre-vocalic consonants were all voiced plosives ([b] or [d]). We avoided English words that are already lexicalized loanwords in Bengali. 10 monosyllabic English words that do not contain the target vowels served as filler items. The complete list of target words can be found in the “Materials” component of the OSF repository for this project.

2.3. Procedure

Participants were recorded individually in a sound attenuated room. Productions were recorded in Praat (Boersma & Weenink, Reference Boersma and Weenink2016) at a 44.1k sampling rate. Participants completed four iterations each of two tasks – cued picture-naming and code-switching. Utterances in the two conditions (nonswitch and switch/mixed) were randomly interspersed within each iteration. In the nonswitch condition, the target word was preceded by another English word in the utterance. In the mixed condition, the target was preceded by a Bengali word and thus produced as a switch from Bengali. Every target word appeared twice during a test block; once in a nonswitch and once in a switch condition. Participants alternated between the two tasks, and were allowed to take a short break after each test block. The tasks are described below:

Cued picture-naming: Participants were presented with slides in the following sequence (see fig. 4a,4b)

Figure 4: (a) English (nonswitch) trial (b) Bengali-English (switch/mixed) trial Picture-naming task sample sequence: language cue (not read out), picture (named as per language cue), target (read out), distractor

The language cue was used to manipulate the word preceding the target, giving target words in nonswitched and switched (mixed) utterances.

Code-switching: Each slide contained one English target word embedded in either an English (nonswitch) or a Bengali (mixed; target produced as a switch from Bengali) carrier2. sentence which was read out by the participant. The carrier sentences are given below:

• • Unilingual sentence: That is a yellow [target word].

• • Mixed-language sentence: o-ʈa æk-ʈa kalo [target word].

Gloss and translation for Bengali sentence:

• o-ʈa æk-ʈa kalo [target word]

• that-CLF one-CLF black [target word]

• “that is a black [target word]

Carrier sentences were such that the sound preceding the target word was uniform across all sentences (the final sound in “yellow” is [o] in Indian English), to minimize differences in coarticulatory effects on the target vowel. We chose a specifier-noun construction, which has identical word-order in Bengali and English. This gave a uniform prosodic context for the target word across conditions. While presenting the mixed-language utterances, we used mixed orthography.

2.4. Acoustic analysis

The first iteration of each task was treated as a trial, and excluded from further analyses. An additional 67 tokens (2.79%) were excluded due to errors in presentation/production, giving 2333 target word tokens ([ʌ]: 1137, [æ]: 1196) for analysis. Recordings were manually segmented and annotated in Praat (Boersma & Weenink, Reference Boersma and Weenink2016), and the first and second formant frequencies (F1 and F2) of the target vowels [ʌ, æ] were measured at five points: in 10% increments starting at 5% into the vowel. This gave measures of F1 and F2 at 5%, 15%, 25%, 35% and 45% into the vowel, allowing us to observe any dynamic changes in the extent of transfer over the course of producing the vowel. Individual vocal tract differences affect formant frequencies across speakers. To reduce between-speaker differences, we used z-score normalization (using the PhonR package (McCloy, Reference McCloy2016) in R (R Core Team, 2013), representing each speaker's formant frequencies as the scaled distance from their own mean).

Plotting these vowels shows a difference between male and female speakers, such that male speakers have overall lower formant frequencies (Figure S1a). Since gender-related differences in formant frequencies are not of interest in this study, we normalized individual values by gender (representing each value as a scaled distance from the group mean; Figure S1b).Footnote ¹ Statistical analyses were performed on these normalized values.