Skip to Main content Skip to Navigation
Theses

Estimation robuste et apprentissage aveugle de modèles pour la séparation de sources sonores

Simon Arberet 1
1 METISS - Speech and sound data modeling and processing
IRISA - Institut de Recherche en Informatique et Systèmes Aléatoires, Inria Rennes – Bretagne Atlantique
Abstract : Blind source separation in the underdetermined case is an ill-posed problem where it is usually assumed that sources are independent and sparse in the time-frequency domain. Separation is then done in two steps : the estimation of the mixture parameters, followed by the estimation of the sources. The assumptions made about the sources are not valid for all the time-frequency points, so that the approaches which naively address all the points identically and independently, are little robust in estimating the mixture parameters and the sources. In this thesis we exploit the local distribution of the mixture in the neighborhood of each time-frequency point, to : - Detect the time-frequency regions where only one source is active and to estimate the direction of the dominant source in these regions ; - Estimate the distribution of the sources in each time-frequency point using the knowledge on the mixture parameters. The proposed local approach is supported by a clustering algorithm called DEMIX, which robustly estimates the mixture parameters in the instantaneous and anechoic cases. On the other hand, the local spatial distribution of the sources can be used to learn Spectral-GMM which until now required a learning step with source examples. We show that this approach improve the source estimation performance of some dB in SDR.
Complete list of metadatas

https://tel.archives-ouvertes.fr/tel-00564052
Contributor : Simon Arberet <>
Submitted on : Wednesday, February 9, 2011 - 2:52:48 PM
Last modification on : Friday, July 10, 2020 - 4:22:54 PM
Document(s) archivé(s) le : Saturday, December 3, 2016 - 8:13:42 PM

Identifiers

  • HAL Id : tel-00564052, version 2

Citation

Simon Arberet. Estimation robuste et apprentissage aveugle de modèles pour la séparation de sources sonores. Traitement du signal et de l'image [eess.SP]. Université Rennes 1, 2008. Français. ⟨tel-00564052v2⟩

Share

Metrics

Record views

445

Files downloads

734