RELAXED DISJOINTNESS BASED CLUSTERING FOR JOINT BLIND SOURCE SEPARATION AND DEREVERBERATION

We propose a novel clustering technique based on a relaxed disjointness assumption for joint blind source separation (BSS) and dereverberation. A disjointness assumption in conventional clustering techniques for BSS is that, at each time-frequency point, observed mixtures consist of a single source only. However, this is not the case in reverberant environments, which causes the performance of the conventional techniques to degrade. To deal with reverberant environments, we introduce a relaxed disjointness assumption: at each time-frequency point, dereverberated mixtures consist of a single source only. Under this assumption, the proposed algorithm alternates dereverberation and clustering-based source separation iteratively, where clustering is performed on dereverberated mixtures. This algorithm is derived based on maximum a posteriori (MAP) fitting of a probabilistic generative model to observed reverberant mixtures. In experiments, the proposed method outperformed a state-of-the-art clustering technique in terms of a signal-to-interference ratio (SIR) by 0.6-4 dB.