Charlotte Pouw

University of Amsterdam

Homepage

Talk: Interpreting models for speech generation and understanding using methods from psycholinguistics

Abstract

Alongside text-based LLMs, models for speech generation and understanding are rapidly improving. To evaluate the extent to which these models exhibit human-like patterns in speech perception and production, we can treat them as psycholinguistic subjects and analyze their responses to carefully controlled stimuli. In this talk, I will present case studies that examine how well these models capture linguistic structures, focusing in particular on phonology and syntax.

Bio

Charlotte Pouw is a final-year PhD student at the University of Amsterdam (Institute for Logic, Language and Computation). She is part of the InDeep consortium, a Dutch research initiative dedicated to developing analysis methods for deep neural networks. Charlotte’s research focuses on analyzing the inner workings of deep learning models trained on speech data, taking inspiration from psycholinguistic research to guide her approach. Her research has been published in *ACL, Interspeech, and Computational Linguistics.