Paula Arkhangorodsky, Ewan Dunbar

Classification Vectors as a Tool for Modelling Human Speech Perception

Introduction

Classification vectors as a tool for modelling human speech perception . Explore human speech perception using classification vectors. Studies show computational models like wav2vec 2.0 simulate identification and discrimination better than acoustic models, solidifying categorical perception's role.

Abstract

Humans perceive speech through the lens of their native phonemic categories, which fundamentally shape both the identification and discrimination of speech sounds. This connection between how we identify and discriminate phonemes is widely recognized as the phenomenon of Categorical Perception. We present two studies that quantitatively explore the role of categorical perception in human speech perception, focusing on how native phonemic categories influence the identification and discrimination of speech sounds when comparing an approach which operationalizes categorical perception compared to one that does not. In Study 1, we apply classification vectors to model English-listener identification experiments, using Mel-Frequency Cepstral Coefficients (MFCCs) and other input representations such as wav2vec 2.0 and DeepSpeech2. Study 2 extends this approach to English-listener discrimination experiments. Our findings indicate that computational models, particularly those using wav2vec 2.0, offer more precise simulations of human speech perception surpassing acoustic-only models. By demonstrating that these classification vectors can effectively model both identification and discrimination tasks, this research provides a quantitative framework that solidifies the importance of categorical perception in human speech perception. 

Full Text

Comments

You need to be logged in to post a comment.