HDR Defense, Anthony Larcher

Title : Acoustic modelling for speaker recognition.

Jury :

• Claude BARRAS, Lecturer HDR, LIMSI, Paris
• Guillaume GRAVIER, Research Director, l’IRISA/CNRS, Rennes
• Jean-François BONASTRE, Professor, Université d’Avignon et des Pays de Vaucluse
• Denis JOUVET, Research Director, CNRS, LORIA, Nancy
• Sylvain MEIGNIER, Professor, Université du Mans

Abstract :

Since the 1990s, advances in speaker recognition enable the deployment of automatic systems for applications that do not require a critical level of performance. Systems’ robustness to noise, transmission channel and lack of data has greatly improved.

For 15 years, I have carried out my research within the framework of the most generic speaker verification (text-independent), but mostly, of its constrained version for which the user must pronounce a predefined text: the text-dependent speaker verification. The ergonomic constraint imposed on the user is justified by the performance of current technologies. When collected voice samples are of short duration (a few seconds), constraining the spoken text greatly improves performance by reducing the variability between the reference sample, refered to as training sample, and the sample to be compared, refered to as test sample. My work at the Laboratoire d’Informatique d’Avignon (LIA), the Institute for Infocomm Research (I2R, A*STAR) and the Laboratoire d’Informatique de l’Université du Mans (LIUM) focuses on improvement and optimization of acoustic modelling for speaker recognition. It is also related to language recognition, speaker-specific speech recognition as well as speech activity detection and speaker diarization.