A real-time blind source separation scheme and its application to reverberant and noisy acoustic environments

In this paper, we present an efficient real-time implementation of a broadband algorithm for blind source separation (BSS) of convolutive mixtures. A recently introduced generic BSS framework based on a matrix formulation allows simultaneous exploitation of nonwhiteness and nonstationarity of the source signals using second-order statistics. We demonstrate here that this general scheme leads to highly efficient real-time algorithms based on block-online adaptation suitable for ordinary PC platforms. Moreover, we investigate the problem of incorporating noncausal delays which are necessary with certain geometric constellations. Furthermore, the robustness against diffuse background noise, eg., in a car environment is examined and a stepsize control is proposed which further enhances the robustness in real-world environments and leads to an improvement in separation performance. The algorithms were investigated in a reverberant office room and in noisy car environments verifying that the proposed method ensures high separation performance in realistic scenarios.