Convolutive blind source separation method based on tensor decomposition

Abstract

A convolutive blind source separation algorithm was proposed based on tensor decomposition framework, to address the estimation of mixed filter matrix and the permutation alignment of frequency bin simultaneously.Firstly, the tensor models at all frequency bins were constructed according to the estimated autocorrelation matrix of the observed signals.Secondly, the factor matrix corresponding to each frequency bin was calculated by tensor decomposition technique as the estimated mixed filter matrix for that bin.Finally, a global optimal permutation strategy with power ratio as the permutation alignment measure was adopted to eliminate the permutation ambiguity in all the frequency bins.Experimental results demonstrate that the proposed method achieves better separation performance than other existing algorithms when dealing with convolutive mixed speech under different simulation conditions.

Keywords

• convolutive blind source separation;tensor decomposition;autocorrelation matrix;permutation ambiguity