Séparation aveugle de source : de l’instantané au convolutif

Abstract : The problem of audio source separation consists in estimating the signals of several sound sources recorded simultaneously by several microphones. This problem is classically known as the "cocktail party problem". Several difficulties appear in this problem. In the most general case, we only know the recordings of the various sources, without even knowing "how", they were recorded: it is the blind case. In addition, an echo effect is added naturally, which must be taken into account in the most realistic cases: we are dealing with a problem of deconvolution. Finally, depending on the number of microphones and the number of sources recorded, one can be in a case said over-determined, where there are more observations than sources, or conversely in an under-determined case. If the problem of the blind separation of instantaneous mixtures, that is to say without echo, if over-determined can be treated in the framework of the independent component analysis (ICA), the case under-determined and convolutional mixtures still pose serious challenges. For audio signals, it is natural to use the notion of parsimony in the time-frequency domain. This is the purpose of the Fourier transform in the short term. With this representation, we can approach the convolutive mixture of several instant-determined problems in each frequency band, and exploit techniques ICA plus the permutation problem of the sources estimated. In the under-determined framework, the separation is generally done in two steps: a step of estimating the mixing system, and a source separation step proper. Here again, these two steps can benefit from the use of parsimony in the time-frequency domain. However, when the mixture is reverberant, the estimation of the mixing system remains a major issue. Especially since, even when this one is known, the estimation of the sources remains a poorly posed inverse problem. The parsimony-based approaches allow here to obtain reference results in terms of objective performance measurements, such as signal-to-distortion ratios, or interference signal.
Complete list of metadatas

Cited literature [113 references]  Display  Hide  Download

Contributor : Fangchen Feng <>
Submitted on : Friday, April 6, 2018 - 10:30:57 PM
Last modification on : Thursday, September 13, 2018 - 12:08:15 PM


Files produced by the author(s)


  • HAL Id : tel-01760974, version 1


Fangchen Feng. Séparation aveugle de source : de l’instantané au convolutif. Sciences de l'ingénieur [physics]. Universite paris sud, 2017. Français. ⟨tel-01760974v1⟩



Record views


Files downloads