Seminar from Félix Saget, PhD student at LIUM

Date: 10/03/2025
Time: 11h00
Place: IC2, Boardroom
Speaker: Félix Saget
 
 

Advances in measuring the interpretability of speaker representation spaces

Features extracted by speaker representation models have proven to be potent and versatile, yielding respectable performance in various speaker-related tasks. However, a human user is incapable of explaining what any of these features correspond to in human-legible terms. Various approaches have claimed to produce “interpretable” neural representations, each one bringing their own definition of interpretability. When it comes to comparing these methods, we need a common definition, as well as a robust way of evaluating the representation with regard to this definition.

As part of my PhD thesis, we have been investigating both unsupervised and supervised metrics, seeking correlations between their behaviors. Conjointly, we have worked on defining a protocol aiming to evaluate interpretability with the help of humans, through a project which is being carried out by Masters students. In this talk, I will report on our most recent progress and perspectives on defining and measuring speaker representation interpretability.