Underdetermined blind separation of convolutive mixtures of speech with directivity pattern based mask and ICA

Abstract

We propose a method for separating N speech signals with M sensors where N > M. Some existing methods employ binary masks to extract the signals, and therefore, the extracted signals contain loud musical noise. To overcome this problem, we propose using a directivity pattern based continuous mask, which masks N - M sources in the observations, and independent component analysis (ICA) to separate the remaining mixtures. We conducted experiments for N = 3 with M = 2 and N = 4 with M = 2, and obtained separated signals with little distortion. © Springer-Verlag 2004.