Blind Separation of Convolutive Speech Mixtures Based on Local Sparsity and K-means

Abstract

In this paper, an accurate and efficient blind source separation method based on local sparsity and K-means (LSK-BSS) is proposed. Specifically, the proposed LSK-BSS approach exploits the local sparsity of speech sources in the transformed domain to obtain closed-form solution for per-frequency mixing system estimation. On this basis, through designing superior initial points of clustering, the well-established K-means algorithm is employed to achieve accurate permutation alignment. Simulations with real reverberant speech sources show that the LSK-BSS approach yields competitive efficiency, robustness and effectiveness, in comparison with the state-of-the-arts methods.