Sound source separation based on multichannel non-negative matrix factorization with weighted averaging

Abstract

Herein, we propose a sound source separation method using multi-channel non-negative matrix factorization (MNMF). MNMF uses an iterative update algorithm for decomposing observed signals into sound source components. However, the separation accuracy of MNMF considerably depends on the initial value of the iterative update algorithm. In the proposed method, cluster analysis and multidimensional scaling were conducted using the features of the matrix decomposed by multiple initial values. A plurality of separated signals was obtained using the initial values included in the largest cluster, which were weighted and averaged. The distance between the matrices obtained by the multidimensional scaling method was used as the weight. As a result of the experiment, we found that the separation signal obtained using the proposed method is less dependent on the initial value and that the separation accuracy is improved.