Blind separation of non stationary sources using joint block diagonalization

Recovering independent source signals from their convolutive mixtures without any a priori knowledge on their structure represents a great challenge in signal processing. We present an efficient solution that is based on the joint block-diagonalization of positive spatio-temporal covariance matrices. In the case of instantaneous mixtures, robust solutions have been proposed previously. Taking advantage of possible non-stationarity of the sources, this new technique uses only second order statistics. The new approach has been successfully applied to the separation of speech signals.