Modeling the concurrent development of speech perception and production in a Bayesian framework

Abstract : It is widely accepted that both motor and auditory processes interact in the brain during speech perception, but little is known about the functional role played by motor processes. To address this question we consider a Bayesian model of speech communication based on three sets of variables: motor representations M, sensory representations S and objects O (e.g. phonological units such as phonemes). The model comprises two internal branches. Firstly, an auditory identification sub-system connects S and O. Secondly, a motor production sub-system connecting M and O and a sensory-motor sub-system connecting M and S can be combined to provide “motor identification” of sounds S, from S to M and from M to O, in an analysis-by-synthesis process. The auditory identification sub-system, the motor production sub-system and the sensory-motor sub-system are learned in a supervised learning scenario, in which a master agent provides sensory signals s and their respective object o. Learning the auditory identification system is straightforward using experimental < s; o > pairs. On the other hand, learning the motor sub-system is more complicated. The learning agent infers motor gestures in an “accomodation process”: the learning agent tries to reproduce the input sensory signal s by selecting a motor gesture m. Performing m yields s’, the resulting sensory output. Triplets < m; s’; o > are used to update the parameters of the motor identification system. We show that the direct inference process involved in auditory identification provides rapid and efficient learning but generalizes poorly. By contrast, the more complex inference process required in motor identification learns more slowly and performs less accurately. However, this system happens to have captured more variable situations during learning, and generalizes better (e.g. in noise). This could provide the basis for a complementarity between auditory and motor identification systems in the human brain.
Type de document :
Poster
Workshop: Probabilistic Inference and the Brain, Sep 2015, Paris, France. 2015
Liste complète des métadonnées

https://hal.archives-ouvertes.fr/hal-01202420
Contributeur : Marie-Lou Barnaud <>
Soumis le : mercredi 23 septembre 2015 - 14:42:11
Dernière modification le : vendredi 31 août 2018 - 09:13:02
Document(s) archivé(s) le : mardi 29 décembre 2015 - 08:53:52

Fichier

Poster.pdf
Fichiers produits par l'(les) auteur(s)

Identifiants

  • HAL Id : hal-01202420, version 1

Citation

Marie-Lou Barnaud, Julien Diard, Pierre Bessière, Jean-Luc Schwartz. Modeling the concurrent development of speech perception and production in a Bayesian framework. Workshop: Probabilistic Inference and the Brain, Sep 2015, Paris, France. 2015. 〈hal-01202420〉

Partager

Métriques

Consultations de la notice

698

Téléchargements de fichiers

73