Reducing reverberation effects in convolutive blind source separation

In this paper, we propose a new method for reducing the reverberation effects in convolutive blind source separation which lead to reduced intelligibility of the separated sources (e.g., speech signals). The existing methods mainly try to maximize the separating performance without paying much attention to the linear distortion in the separated signals. We propose a modification to the existing algorithms that reduces the distortions introduced by the demixing filters. In particular we investigate the possibilities of modifying the frequency responses of the demixing filters to have no spectral peaks, which leads to near allpass character of the overall system. The good performance of the modified algorithm will be demonstrated on real-world data.