Improving the Robustness of the Correlation Approach for Solving the Permutation Problem in the Convolutive Blind Source Separation

In this paper, we propose a modification to the correlation approach in convolutive blind source separation to achieve an improved robustness. An often used approach for separation of convolutive mixtures is the transformation to the time-frequency domain. This allows for the use of an instantaneous ICA algorithm independently in each frequency bin, which greatly reduces complexity. The drawback of this approach are the so-called permutation and scaling problems. Here, we modify the well known correlation approach for making it more robust. We propose to incorporate a confidence function based on estimated SIR which allows for detection of frequency bins with high probability of wrong permutations. The results of the new algorithm will be shown on an real-world example.