Separation of stereo speech signals based on a sparse dictionary algorithm

Abstract

We address the problem of source separation in echoic and anechoic environments, with a new algorithm which adaptively learns a set of sparse stereo dictionary elements, which are then clustered to identify the original sources. The atom pairs learned by the algorithm are found to capture information about the direction of arrival of the source signals, which allows to determine the clusters. A similar approach is also used here to extend the dictionary learning K singular value decomposition (K-SVD) algorithm, to address the source separation problem, and results from the two methods are compared. Computer simulations indicate that the proposed adaptive sparse stereo dictionary (ASSD) algorithm yields good performance in both anechoic and echoic environments.