Speech source separation using a generalized mean shift algorithm

Abstract: Speech source separation in the time–frequency domain is a modern approach that exploits the sparsity of speech when it is represented in such domain. Several methods based on this approach exist, DUET being the most remarkable of these. In this work we propose a novel time–frequency domain algorithm for sound source separation, based on a generalization of the mean shift clustering method. The proposed algorithm can be applied to separate an undetermined number of sources from two mixtures. The new method is compared to the DUET algorithm, as well as with a modification of DUET based on k-means, for different types of mixtures: linear speech mixtures, binaural speech mixtures, linear speech and noise mixtures and linear speech and music mixtures. From the results we note that the use of the proposed algorithm based on mean shift for speech separation shows a significantly better performance than the DUET algorithm. [Copyright &y& Elsevier]