Separating Stereo Audio Mixture Having No Phase Difference by Convex Clustering and Disjointness Map
Atsushi Hiruma, Kohei Yatabe, Yasuhiro Oikawa · 2018
In this paper, a method of constructing binary masks for stereo source separation is proposed. The proposed method consists of two main factors: (1) Disjointness map, and (2) convex directional clustering. The disjointness map quantifies the degree of mixing at each time-frequency bin based on instantaneous frequencies. This map enables to effectively utilize phase information which is usually omitted in stereo music separation. Then, the convex clustering utilizes flexible definitions of adjacency of the time-frequency bins for incorporating more information into directional clustering. Experimental results indicate that the proposed method can obtain a mask closer to the ideal one than the conventional directional clustering.