Location – Dodd 146

New perspectives on speech variability from large-scale studies

I present two studies which aim to understand the structure and sources of variability in speech production, enabled by novel quantitative methods. I also discuss several open-source tools for automatic analysis of speech which enable such large-scale studies by speeding up or replacing manual processing.

The first study asks how much “the same” phonetic effects vary across languages, and what their distributions can tell us. It is well known that vowel F0 tends to be higher for high vowels compared to low vowels (VF0) and F0 following voiceless obstruents tends to be higher than F0 following voiced obstruents (CF0). These “intrinsic F0” effects have been extensively investigated, but largely in lab speech in individual languages, making it difficult to settle the long-standing questions of to what extent they are “automatic” versus “controlled”, and whether cross-linguistic variation can be systematically explained (e.g. smaller VF0/CF0 in tone languages). We map out the distribution of intrinsic F0 effects across 20 languages, using large corpora of read speech. We find that both VF0 and CF0 effects are robust across languages in terms of effect direction, but that languages vary greatly in VF0 and CF0 effect size. Most of this variability appears to be unexplained by phonological properties of the language (e.g. tone). Finally, we find that the CF0 effect is larger than the VF0 effect in almost every language, and the CF0 effect is more variable across languages. These patterns suggest a possible explanation for the cross-linguistic tendency for CF0 effects to lead to sound change (“tonogenesis”) much more often than VF0 effects. (Joint work with Connie Ting & Meghan Clayards)

A second study is part of the SPeech Across Dialects of English (SPADE) project, which examined existing datasets of Old World (British) and New World (North American) English spanning 100+ years, to gain an improved understanding of stability and variation in a spoken language across space and time. We consider variability in English sibilants in 5k speakers from 27 geo-social-ethnic regions. We analyse spectral peak measures using a Bayesian statistical model which explicitly models token, speaker, and region- level variability. Following previous phonetic and sociolinguistic research we expected English /s/ to be more variable than /sh/, and for both sounds to look broadly similar across dialects. The results, however, differ according to the level at which we consider variability — for example, /s/ is more variable within-speaker than /sh/, but not more variable across  regions. Our results likely reflect physical and social-indexical sources of variation, and connect to basic issues in sociolinguistics and sound change. (Joint work with Jane Stuart-Smith, Jeff Mielke)