École Nationale Supérieure de
Techniques Avancées

Abstract

The process of how infants acquire speech skills remains an open question. One of the key challenge is understanding the mechanisms behind speaker normalization within the human brain. This study aims to address this question by tackling a vowel clustering task. First we will implement a supervised classifier and compare its performance to the analytical normalization methods found in the literature. Following this, we will investigate how contrastive loss can be used within a reinforcement learning architecture to cluster vowels within some representation space. This report emphasizes the importance of exploring the early establishment of the relationship between abstract phonetic representations and articulatory spaces, shedding light on the complex mechanisms involved in speech acquisition during infancy