Séminaire inter-équipe S2S-LSD

Vendredi 12 septembre 2025 de 12h à 14h
Laboratoire Parole et Langage, salle de conférences B011

Sheng-Fu Wang

Modeling durational variability in Taiwanese with surprisal from neural language models

Résumé :
Using variability in syllable duration in Taiwanese spontaneous speech as a case study, this study aims to demonstrate that probabilistic measurements from neural language models can be used not only to improve the modeling of phonetic variability but also configured in different ways to test hypotheses about speech planning. The focus on probability (surprisal) and phonetic variability arises from decades of research showing that acoustic properties of speech sounds have been shown to correlate with their contextual probability — the likelihood of occurring in specific linguistic contexts (Seyfarth, 2014; Van Son & Van Santen, 2005; Jurafsky et al., 2001), which is often interpreted as listener-oriented speech planning that aims to balance information distribution within and across utterances (e.g., Aylett & Turk, 2004). Phonetic studies in this area often have the implicit assumption that it is local contexts that are relevant, based on their employment of simple n-gram (and mostly bigram) models, despite evidence of nonlocal speech planning (e.g., Fuchs et al., 2013; Krivokapic, 2012). Neural language models, with their ability to easily take longer contexts into account when estimating probability, open up opportunities to test this locality assumption that was potentially a byproduct of methodological restriction. Employing bidirectional Transformer-based masked language models (RoBERTa, Liu et al., 2019), this study seeks to answer questions on the relevance of longer contexts and the potential role of linguistic chunking.

Page Web : https://sheng-fu.github.io/