Predicting non-native speech perception using the Perceptual Assimilation Model and state-of-the-art acoustic models

Proceedings of the 25th Conference on Computational Natural Language Learning (CoNLL), pages 661–673 November 10–11, 2021. ©2021 Association for Computational Linguistics 661 Predicting non-native speech perception using the Perceptual Assimilation Model and state-of-the-art acoustic models Juliette Millet CoML, ENS/CNRS/EHESS/INRIA/PSL LLF, University of Paris, CNRS CRI, FAN, IIFR, University of Paris Paris, France Ioana Chitoran Department of Linguistics, University of Paris Paris, France Ewan Dunbar University of Toronto, Toronto, Canada CoML, ENS/CNRS/EHESS/INRIA/PSL, Paris, France juliette.millet@cri-paris.org ioana.chitoran@univ-paris-diderot.fr ewan.dunbar@utoronto.ca Abstract Our native language influences the way we perceive speech sounds, affecting our ability to discriminate non-native sounds. We com- pare two ideas about the influence of the na- tive language on speech perception: the Per- ceptual Assimilation Model, which appeals to a mental classification of sounds into native phoneme categories, versus the idea that rich, fine-grained phonetic representations tuned to the statistics of the native language, are suffi- cient. We operationalize this idea using rep- resentations from two state-of-the-art speech models, a Dirichlet process Gaussian mixture model and the more recent wav2vec 2.0 model. We present a new, open dataset of French- and English-speaking participants’ speech percep- tion behaviour for 61 vowel sounds from six languages. We show that phoneme assimila- tion is a better predictor than fine-grained pho- netic modelling, both for the discrimination be- haviour as a whole, and for predicting differ- ences in discriminability associated with dif- ferences in native language background. We also show that wav2vec 2.0, while not good at capturing the effects of native language on speech perception, is complementary to in- formation about native phoneme assimilation, and provides a good model of low-level pho- netic representations, supporting the idea that both categorical and fine-grained perception are used during speech perception. 1 Introduction Our native language has an impact on our speech perception. Importantly, it influences the way we perceive unfamiliar sounds—for example, from other languages. For instance, Japanese native speakers have difficulties in telling apart the En- glish sounds [r] and [l] (Yamada and Tohkura, 1992), which are not in their native phoneme in- ventory, and confuse the English words ‘right’ and ‘light’ (Yamada and Tohkura, 1990), while this con- trast is highly distinct for English native speakers. The influence of native language on speech per- ception is not yet fully understood. A simple ap- proach might assert that two unfamiliar sounds are only discriminable if they would be classified as distinct phonemes by the listener. However, some unfamiliar sound contrasts are easy. For example, English-speaking listeners have no difficulty with Zulu click contrasts (Best et al., 1988) or the Thai [W]–[7] vowel contrast (Tyler et al., 2014). The Perceptual Assimilation Model (PAM: Best and Tyler 2007) proposes that non-native speech perception is governed by the complete profile of listeners’ assimilations to their native phoneme cat- egories, which may be indeterminate. Thus, even if English has neither [W] nor [7], English-speaking listeners do not assimilate [W] to any specific En- glish vowel, while they categorise [7] as English [2]. Since one of the two unfamiliar phones …

Content truncated to 500 words