Multichannel speech separation using adaptive parameterization of source PDFs

Convolutive and temporally correlated mixtures of speech are tackled with an LP-based temporal pre-whitening stage combined with the natural gradient algorithm (NGA), to essentially perform spatial separation by maximizing entropy at the output of a nonlinear function. In the past, speech sources have been parameterized by the generalized Gaussian density (GGD) model, in which the exponent parameter directly relates to the exponent of the corresponding optimal nonlinear function. In this paper, we present an adaptive, source dependent estimation of this parameter, controlled exclusively by the statistics of the output source estimates. Comparative experimental results illustrate the inherent flexibility of the proposed method, as well as an overall increase in convergence speed and separation performance over existing approaches.