Chargement Évènements

« Tous les Évènements

  • Cet évènement est passé

Séminaire du D4 “TALC” le 22 octobre

octobre 22, 2013 @ 14:00 - 15:00

Le mardi 22 octobre à 1’h en salle C005, aura lieu le séminaire du département 4 “traitement des langues et des connaissances”. Antoine Deleforge, de l’équipe PERCEPTION d’Inria Grenoble fera un exposé ayant pour titre “Acoustic Space Mapping for Sound Source Separation and Localization”.

Abstract:

The cocktail party effect describes the ability to focus one’s listening attention on a single sound source among a mixture of conversations, music and/or background noises, ignoring other interfering sources. While human listeners solve this task routinely and effortlessly, it is still a current research problem to address it from the perspective of “machine hearing”. Two key aspects of the cocktail party problem are the localization and the separation of several sound sources. In this talk, I will present a new framework for addressing these tasks, referred to as “Acoustic Space Mapping”. Traditional methods in sound source localization are usually based on simplifying geometrical assumptions of sound propagation in the system. Instead, we propose to automatically learn the relationship between sound source positions and associated auditory cues recorded with two microphones. I will first present a setup and some experimental protocols specifically designed to gather auditory-cue-to-position associations for training. Then, I will show how the manifold structure of acoustic data can be exploited using a probabilistic locally-linear mapping model and an Expectation-Maximization (EM) inference technique. We’ll finally see how this probabilistic model and technique can be progressively extended to separate and localize multiple natural sound sources emitting at the same time in a real world environment with very high accuracy. The talk will be illustrated with videos and auditory results.

Abstract:

The cocktail party effect describes the ability to focus one’s listening attention on a single sound source among a mixture of conversations, music and/or background noises, ignoring other interfering sources. While human listeners solve this task routinely and effortlessly, it is still a current research problem to address it from the perspective of “machine hearing”. Two key aspects of the cocktail party problem are the localization and the separation of several sound sources. In this talk, I will present a new framework for addressing these tasks, referred to as “Acoustic Space Mapping”. Traditional methods in sound source localization are usually based on simplifying geometrical assumptions of sound propagation in the system. Instead, we propose to automatically learn the relationship between sound source positions and associated auditory cues recorded with two microphones. I will first present a setup and some experimental protocols specifically designed to gather auditory-cue-to-position associations for training. Then, I will show how the manifold structure of acoustic data can be exploited using a probabilistic locally-linear mapping model and an Expectation-Maximization (EM) inference technique. We’ll finally see how this probabilistic model and technique can be progressively extended to separate and localize multiple natural sound sources emitting at the same time in a real world environment with very high accuracy. The talk will be illustrated with videos and auditory results.

Share on FacebookShare on Google+Tweet about this on TwitterShare on LinkedIn

Détails

Date :
octobre 22, 2013
Heure :
14:00 - 15:00
Catégorie d’Évènement:

Lieu

C005

En ce moment

Colloquium Loria 2017
Présentations et vidéos

Logo du CNRS
Logo Inria
Logo Université de Lorraine