Spatial separation of sound sources

Abstract : Blind source separation is a promising technique for the identification, localization, and ranking of sound sources. The aim of this dissertation is to offer methods for separating incoherent sound sources which may overlap in both the space and frequency domains by exploiting spatial information. This is found of interest in acoustical applications involving the identification and ranking of sound sources stemming from different physical origins. The fundamental principle of all proposed methods proceeds in two steps, the first one being reminiscent to source reconstruction (e.g. as in near-field acoustical holography) and the second one to blind source separation. Specifically, the source mixture is first expanded into a linear combination of spatial basis functions whose coefficients are set by backpropagating the pressures measured by an array of microphones to the source domain. This leads to a formulation similar, but no identical, to blind source separation. In the second step, these coefficients are blindly separated into uncorrelated latent variables, assigned to incoherent “virtual sources”. These are shown to be defined up to an arbitrary rotation. A unique set of sound sources is finally recovered by searching for that rotation (conjugate gradient descent in the Stiefel manifold of unitary matrices) which minimizes some spatial criteria, such as spatial variance, spatial entropy, or spatial orthogonality. This results in the proposal of three separation criteria coined “least spatial variance”, “least spatial entropy”, and “spatial decorrelation”, respectively. Meanwhile, the condition under which classical decorrelation (principal component analysis) can solve the problem is deduced in a rigorous way. The same concept of spatial entropy, which is central to the dissertation, is also exploited in defining a new criterion, the entropic L-curve, dedicated to determining the number of active sound sources on the source domain of interest. The idea consists in considering the number of sources that achieves the best compromise between a low spatial entropy (as expected from compact sources) and a low statistical entropy (as expected from a low residual error).
Complete list of metadatas

Cited literature [153 references]  Display  Hide  Download
Contributor : Abes Star <>
Submitted on : Friday, July 10, 2015 - 3:31:06 PM
Last modification on : Saturday, May 18, 2019 - 3:34:09 AM
Long-term archiving on : Monday, October 12, 2015 - 11:26:44 AM


Version validated by the jury (STAR)


  • HAL Id : tel-01175498, version 1


Bin Dong. Spatial separation of sound sources. Acoustics [physics.class-ph]. INSA de Lyon, 2014. English. ⟨NNT : 2014ISAL0040⟩. ⟨tel-01175498⟩



Record views


Files downloads