↓ Aller au contenu

Séminaires d’équipe

The Binary-Attribute Likelihood Ratio estimation (BA-LR) approach

Intervenant·e
Jean-François Bonastre (LIA, Avignon)
Date
Lieu
Salle 306, bâtiment IMAG + visio

This seminar presents BA-LR, a new approach in automatic voice comparison developed in the framework of the LIAvignon chair and Imen Benamor’s thesis work. This approach is dedicated to explicabilty and interpretability aspects, contrarily to classical approaches more focused on “accuracy”. These aspects take a particular importance in the field of voice comparison because this field raises strong societal and ethical issues, mainly related to forensic applications.

The first part of the seminar defines automatic voice comparison and proposes a brief introduction on what means explicability and interpretability in AI. Voice comparison is used here as a a case study. The second part of the seminar is devoted to the BA-LR approach.While classical voice comparison approaches produce a single value, the LR, without further explanation, our approach proposes to see the output of the system as the composition of a set of factors, related to the voice attributes. Moreover, the behavior of each attribute is described by three easily understood parameters. Finally, in addition to building a voice comparison system that is more explicit and understandable than standard solutions, our approach allows us to stay within the main paradigm of voice comparison, the likelihood ratio, which is known to be the “correct and logical” paradigm for forensic applications. Next the limits of the current implementation and the remaining steps to be taken for a fully transparent approach, where each piece of information and decision is defined precisely and understandable by all, is discussed. Finally, a conclusion is proposed, keeping in mind foransic application of voice comparison.