Blind source separation using correlation at neighboring frequencies

We propose blind source separation using correlation at neighboring frequencies based on DUET. DUET assume that sound sources meet the W-disjoint orthogonal and each source exists in the different position. Using these conditions, the amplitude and phase parameters are estimated to create amplitude-phase histogram from two mixtures. Each peak in the histogram corresponds to each source. However, because there exists acoustic echo in the real environment, the histogram distributes over wide range and we can not classify each source. In the frequency domain, acoustic signals have high correlations at neighboring frequencies. We incorporate this idea to the DUET to separate the sources more exactly. Through the simulation of both DUET and our new method, we compare the separation performance of our new method with that of DUET and confirm the validity of our method.