Using computational models to study accent adaptation: A tutorial

Abstract

One of the limitations in accent perception research is the difficulty of quantifying how much exposure is needed to observe adaptation to a novel accent. While numerous studies have investigated this issue (Baese-Berk et al., 2013; Clarke & Garrett, 2004), it has been impossible to control for individual-level differences and solely focus on linguistic differences with human participants. However, computational models can be intentionally set in terms of these features to better understand the process of adaptation to novel accents. Here we discuss how machine learning models can be implemented using PyTorch, a framework in Python made by Meta. PyTorch allows for the easy creation of language learning models that can be used in speech-recognition experiments (Paszke et al., 2019). Here we use the Wav2Vec2 model created by Meta, pre-trained on 960 hours of the Librispeech ASR corpus (Panayotov et al., 2015). We discuss importing the model and fine-tuning it using samples of different accents. Next, we demonstrate building a decoder using a greedy algorithm to transform the model's classification into a transcript. Our goal is to compare simulations against data from human participants and tease apart the direct effects of individual variability on accent perception.