The role of passage length on acoustic voice variability in bilingual speech

Abstract

An individual's voice is determined in part by the limitations of their anatomy and physiology, in addition to language-specific phonological and phonetic structure. When a bilingual switches between languages, how much do they change their voice? Previous work using a corpus of spontaneous speech from early Cantonese-English bilinguals found surprisingly little variability across individuals' languages [Johnson et al., Proc. of Interspeech (2020)] compared to earlier research on across-talker acoustic voice variability [Lee et al., JASA (2019)]. A crucial difference between these two studies, however, is passage length. A longer passage (e.g., 30 min) potentially allows for a more stable structure to emerge in a principal components analysis, while a shorter sample (e.g., 2 min or less) may instead be subject to ephemeral variation, and potentially misrepresent the overall variability of a voice. Building on Johnson et al. (2020), the present study asks: to what extent does passage length impact the results of principal components and canonical redundancy analyses designed to elucidate within-talker (across languages) and across-talker (within language) idiosyncratic variation? These results are important for theories of talker recognition, identification, and discrimination, in addition to improving understanding of talker-specific acoustic-phonetic variation.