Blind separation of convolutive mixtures: a Gauss-Newton algorithm

Abstract

This paper addresses the blind separation of convolutive mixtures of independent and non-Gaussian sources. We present a block-based Gauss-Newton algorithm which is able to obtain a separation solution using only a specific set of output cross-cumulants and the hypothesis of soft mixtures. The order of the cross-cumulants is chosen to obtain a particular form of the Jacobian matrix that ensures convergence and reduces computational burden. The method can be seen as an extension and improvement of the Van-Gerven's symmetric adaptive decorrelation (SAD) method. Moreover the convergence analysis presented in the paper provides a theoretical background to derive an improved version of the Nguyen-Jutten (1995) algorithm.