Training computational models with accents

Accent adaptation research demonstrates that monolingual speakers can quickly adapt to novel foreign accents with brief exposure or training (Baese-Berk et al., 2013). However, variability in language exposure among monolinguals complicates the findings (Castro et al., 2022). Computational modeling provides an avenue to control linguistic diversity and isolate the effects of exposure on accent adaptation (Paszke et al., 2019). Using Large Language Models (LLMs) such as Wav2Vec2 and Whisper that are pretrained for Automated Speech Recognition (ASR), we explore the process of accent adaptation by simulating monolinguals exposed to single or multiple accents. We implement and fine-tune the Wav2Vec2 model using PyTorch and Whisper model with HuggingFace transformer library. These facilitate replicating human experimental designs and building decoders to evaluate model outputs. Findings suggest limited exposure to one or multiple accents fails to significantly enhance adaptation to novel accents (no accent = 69% correct, one accent = 68.6% correct, multiple accents = 69% correct) in Wav2Vec2. However, Whisper model outperforms in any training model (word error ∼5%). By comparing simulation data with human participant studies, we aim to identify the effects of exposure and individual variability on accent adaptation. Our results highlight the challenges of determining sufficient exposure for adaptation and underscore the importance of computational approaches to complement human-based accent perception research.