A low-complexity permutation alignment method for frequency-domain blind source separation

Abstract

Frequency-domain blind source separation is an effective way to separate the signals from convolutive mixtures. The independence component analysis (ICA) is commonly employed to separate signals in each frequency bin, resulting in the well-known permutation problem. To resolve this problem, we present a low-complexity permutation alignment method based on the inter-frequency dependence of signal power ratio. A bin-wise permutation alignment is first carried out across all the frequency bins by measuring the correlation between the current frequency bin and the previous one, but only the permutation with a high confidence is fixed. The permutation with low confidence is then determined by maximizing the correlation between the current frequency bin and a local centroid, which is calculated from a set of determined frequency bins with high confidence. By so doing, the permutation for most frequency bins is aligned without iterations. Finally, a clustering algorithm with centroids is adopted to achieve the fine global optimization in the fullband with only a few iterations. Experiment results show that the proposed method achieves a comparable performance with the state-of-the-art permutation alignment schemes, but the new method achieves a significant computational saving.