Underdetermined blind source seperation of non-disjoint nonstationary sources in the time-frequency domain

Abstract

This paper considers the blind separation of nonstationary sources in the underdetermined case, in which we have more sources than sensors. A recently proposed algorithm applied time-frequency distributions (TFDs) to this problem and gave good separation performances in the case where sources were disjoint in the time-frequency (TF) plane. However, in the non-disjoint case, the method simply relied on the interpolation at the intersection TF points implicitly performed by a TF synthesis algorithm, instead of directly treating these points. In this paper, we propose a new algorithm that combines the abovementioned method with subspace projection in order to explicitly treat non-disjoint sources. Another contribution of this paper is the estimation of the mixing matrix in the underdetermined case.