Music Signal Separation Based on Supervised Nonnegative Matrix Factorization with Orthogonality and Maximum-Divergence Penalties
In this paper, we propose a new stereo signal separation scheme based on multi-divergence supervised nonnegative matrix factorization (SNMF). In previous studies, a hybrid method, which concatenates superresolutionbased SNMF after directional clustering, has been proposed for multichannel signal separation. However, the optimal divergence in SNMF temporally fluctuates because the separation and extrapolation abilities depend on spatial conditions of sources in music tunes. To solve this problem, we propose a new scheme for multi-divergence, where optimal divergence can be automatically changed in each time frame according to the local spatial conditions. Experimental results show the proposed method efficacy.