Séparation aveugle de sources sonores par factorization en matrices positives avec pénalité sur le volume du dictionnaire

Abstract

Audio source separation concerns techniques used to extract unknown signals called sources from a mixed signal. In this paper, we assume that the audio signal is recorded with a single microphone. Considering a mixed signal composed of various audio sources, the blind audio source separation consists in isolating and extracting each of the sources on the basis of a single recording. Usually, the only known information is the number of estimated sources present in the mixed signal. Based on a time-frequency representation of the signal, classical source separation techniques integrate algorithms such as nonnegative matrix factorization (NMF). Optimization problems in blind audio source separation are based on the minimization of criteria such as the Kullback-Leibler and Itakura-Saito divergences, both divergences belonging to the family of β-divergences. In this paper, we present a new model of separation based on the minimization of the Kullback-Leibler includinga penalty term promoting the columns of the dictionary matrix to have small volume. In order to solve this problem, the global cost function is replaced by a convex and separable auxiliary function that will be minimized. We will show that we obtain more interpretable results in the case where the factorization rank (that is, the number of sources present into the mixed signal) is overestimated.

More information

Identifiers

ORCID for Man Shun Ang: orcid.org/0000-0002-8330-758X

Catalogue record

Export record