On separation performance enhancement in convolutive blind source separation

Abstract

In this paper we present a method for estimating the bin-wise separation performance in convolutive blind source separation. A common way to separate convolutive mixtures is the transformation to the time-frequency domain and the separation of the single bins using an instantaneous ICA algorithm. This approach reduces the complexity but leads to the so-called permutation problem which has been widely studied. Another problem arises when the single bins are only poorly separable or even not separable at all. These bins can significantly reduce the overall performance. In this paper we propose a method for detecting such bins based on properties of the unmixing matrices.