A three-stage permutation approach for frequency-domain blind source separation

Abstract

The permutation ambiguity is a critical issue for the convolutive blind speech separation in the frequency domain. To solve this problem, a novel three-stage approach is proposed in this paper. The method of choosing benchmark frequency bins is employed to ensure the performance of the second stage. Meanwhile, the correlation of the adjacent frequency bins is combined with the DOA method to bring out the best in each other. Experimental results in simulated and real environment verify the effectiveness and robustness of our method.