Revisiting rapid accent adaptation with computational modeling

Abstract

Studies on accent adaptation report that monolinguals can rapidly adapt to a novel foreign accent with short exposure or training (Baese-Berk et al., 2013; Clarke & Garrett, 2004; Maye et al., 2008). While most of these studies rely on self-reported monolinguals, there is great variability in how much language diversity monolinguals encounter regularly (Castro et al., 2022). We return to this question of rapid accent adaptation in monolinguals through computational modeling where linguistic diversity is directly controlled. We simulate monolingual speech perception using PyTorch’s Wav2Vec2 model (Paszke et al., 2019), pre-trained on the Librispeech corpus with American-accented English (Panayotov et al., 2015). We replicate the experimental design of Baese-Berk et al. (2013) by fine-tuning this model on a single accent (n = 150 sentences across 5 talkers) and multiple accents (n = 30 sentences for each of 5 talkers). Preliminary results find that exposure to a single accent (68.6% correct) or multiple accents (69% correct) does not induce accent adaptation to a novel accent (no accent training = 69% correct). We predict that exposure to a single or multiple accents will increase accuracy but requires many additional hours of exposure. We discuss implications of our models against accent adaptation studies.