Solving permutation problem with a cascade combination of phase difference entropy and power spectral correlation

In this paper, we propose a novel method to solve the permutation problem for multi-channel frequency-domain blind source separation problems. For low spectral correlation problem between lower frequencies and higher frequencies, the proposed method utilizes phase difference information between microphones so as to avoid incorrect permutation alignment problems in addition to power spectral information. However, reliability of phase difference information of each frequency is not so high. At first, the proposed method performs a clustering based approach by using power-spectral correlation. The first stage is corresponding with clustering of neighborhood frequencies based on an assumption that power spectral correlation is high between neighborhood frequencies. Next, the proposed method refines permutation alignment problem so as to minimize entropy of phase differences in each cluster. The second stage is based on an assumption that time-difference information which consists of several frequencies is reliable. Experimental results with the 3rd CHiME dataset show that the proposed method is effective.