UC Berkeley

In phonetics research, language alternation–including code switching (speaker-initiated) and cued switching (researcher-prompted)–can be used as a tool to investigate various aspects of speech production and perception in bilingual or multilingual speakers (Bullock & Toribio, 2009a). Studies on the production of voice onset time (VOT) during language alternation have demonstrated that bilingual speakers—for example, Spanish-English bilinguals—have a convergence effect, in which VOT of a given language near switch sites becomes more similar to the VOT norms of the other language (e.g., Toribio et al., 2005; Bullock et al., 2006; Balukas & Koops, 2015; Piccinini & Arvaniti, 2015; Olson, 2016). As Olson (in press) summarizes, some general patterns have been identified in the literature: (1) “unidirectional interference has been the most common finding in the literature. When unidirectional interference is reported, the language with long-lag VOTs shifts in the direction of the short-lag language”, and (2) “when bidirectional interference has been found, the magnitude of the switch differs between the two languages. The degree of shift is generally larger for the long-lag language than the short-lag language” (Olson, in press, p. 7). However, regardless of these general patterns, it is evident that a great number of distinct outcomes has been reported, namely, convergence presence and directionality (no convergence, unidirectional convergence, or bidirectional convergence), the language(s) more prone to convergence (only long-lag language or both), and the magnitude of the effect (small or large shifts). The wide array of different task types and methodologies used to study this phenomenon makes it all the more difficult to pinpoint the cause of the aforementioned inconsistent directionality, propensity, and magnitude of the convergence effect. For example, some studies relied on word list reading tasks (e.g., Olson, 2013) or passage reading tasks (e.g., Toribio et al., 2005; Bullock et al., 2006), and others used speech produced during sociolinguistic interviews (e.g., Balukas & Koops, 2015), inter-subject conversations (e.g., Balukas & Koops, 2015; Piccinini & Arvaniti, 2015), or puzzle tasks (e.g., Piccinini & Arvaniti, 2015).

In order to shine light on the possible effect of task type on the directionality and magnitude of convergence effects in VOT bilingualism research, the present study analyzes VOT productions across four of the most popular research tasks (i.e., word list reading, passage reading, puzzle–spot the difference–and casual interview tasks) from a single group of Spanish-English bilingual speakers to obtain VOT measurements. In addition to the aspect of research task, the present study incorporates linguistic (i.e., language, place of articulation, and speech rate) and social (i.e., language history, language proficiency, language usage, and language attitudes) factors that could help predict VOT production patterns in language alternation. A total of 60 Spanish-English bilingual subjects participated in the four aforementioned tasks for this study, which yielded nearly 65 hours of recorded code-switched speech, in addition to a standardized demographic questionnaire: the Bilingual Language Profile (Birdsong et al., 2012). Data collection took place in a sound booth at the Berkeley PhonLab. The audio for the word list and passage reading tasks were annotated by hand. The audio for the spot-the-difference puzzle and the interview were segmented into Spanish and English speech. A Python script generated automatic transcriptions for each language using OpenAI’s Whisper language model for automatic speech recognition (Radford et al., 2022). The data for all four tasks were then forced aligned using the Montreal Forced Aligner (McAuliffe et al., 2017). Through these aligned annotations, VOT measurements were obtained for word-initial voiceless stops /p t k/ in non-cognate words in both languages using AutoVOT software (Keshet et al., 2014).

Two statistical models were performed for this study–one for the methodological and linguistic factors and another for social factors. The results from the first model indicate that there is an effect of research task, with passage reading and interview tasks displaying the highest degree of convergence–in English, specifically–compared with the word list and puzzle tasks. In addition, there were significant results of language (English yielded longer VOT), place of articulation (for English, /p/ was shorter than /t k/; for Spanish, /p t/ were shorter than /k/), and speech rate (slower speech rate led to longer VOT). Finally, the second statistical model reports relevant factors of language proficiency (participants who report higher ability scores for reading and writing–regardless of language–display less convergence than participants who report higher scores for speaking and comprehension abilities), language history (participants who learned Spanish later in life display more convergence), language usage (participants who use more English with friends and in the workplace display less convergence), and language attitudes (participants who more closely identify with a Spanish-speaking culture display more convergence than participants who identify with an English-speaking culture). The social model also indicated that participants who had higher exposure to Spanish spoken with an American English accent display less convergence than participants with exposure to any other type of accented speech.

The results from this study are discussed in relation to previous empirical studies and predictions made by relevant theoretical frameworks for (Spanish-English) bilingualism and language alternation. In particular, this study elaborates on (a) the potential language processing mechanism at play across the different types of research tasks during language alternation, (b) the significance of divergent VOT patterns of /t/ between the two languages, and (c) the effects of exposure to accented speech on a speaker’s own speech patterns. All in all, this research study provides (1) a thorough analysis and comparison of the research methodologies typically used in code switching studies in order to uncover task effects in production studies and (2) a better understanding of the language processing mechanisms that are engaged during language alternation behaviors.