Blind Source Separation Using Independent Low-Rank Matrix Analysis with Spectrogram-Consistency Regularization

Block permutation is a problem that occurs in the separated signals obtained by independent low-rank matrix analysis. To deal with this problem, a new method that induces a low-rank property on the projection of the separated spectrogram to the so-called consistent spectrogram for short-time Fourier transform has recently been proposed. In contrast to conventional methods, which disregard how far the separated spectrogram is from the projected spectrogram, we design a regularization term that induces the separated spectrogram itself to be consistent. For the newly designed optimization problem, we derive the update rule based on the majorization-minimization algorithm and the vectorwise coordinate descent method. We verify the effectiveness of the proposed method experimentally.