Research Seminars

Extracting several audio sources from stereophonic recordings: some approaches based on sparsity

Dr. Rémi Gribonval, IRISA/INRIA, Rennes, France.

Wednesday 7 December 2005, 4:00pm, Room 105

Abstract

Most musical recordings are a mixture of several instruments on two channels. Extracting the sound of each instrument in such recordings is a degenerate blind source separation problem which solution requires some prior knowledge on the instruments. The DUET algorithm has popularized separation techniques based on time-frequency masking, under the assumption that sources have sparse, disjoint Short Time Fourier Transforms. The goal of this talk is to give an overview of some recent contributions to this family of approaches, based on multichannel sparse decompositions of the recordings in a redundant multiscale time-frequency dictionary. We will discuss the estimation of both the number of present sources and the mixing matrix, as well as practical and theoretical aspects of the 'inversion' of the mixture model with Matching Pursuit. If time permits we will discuss some alternative approaches based on Gaussian Mixture models used in single channel separation.