Convolutive blind separation of speech mixtures using the natural gradient

Convolutive blind separation of speech, also known as the "cocktail party problem", is a challenging task for which few successful algorithms have been developed. In this paper, we explore two novel algorithms for separating mixtures of multiple speech signals as measured by multiple microphones in a room environment. Both algorithms are modifications of an existing approach for density-based multichannel blind deconvolution (MBD) using natural gradient adaptation. The first approach employs non-holonomic constraints on the multichannel separation system to effectively avoid the partial deconvolution of the extracted speech signals within the separation system's outputs. The second approach employs linear predictors within the coefficient updates and produces separated speech signals whose autocorrelation properties can be arbitrarily specified. Unlike MBD methods, the proposed techniques maintain the spectral content of the original speech signals in the extracted outputs. Performance comparisons of the proposed methods with existing techniques show their usefulness in separating real-world speech signal mixtures. (C) 2002 Published by Elsevier Science B.V.