SPARSENESS-BASED MULTICHANNEL NONNEGATIVE MATRIX FACTORIZATION FOR BLIND SOURCE SEPARATION

This paper deals with the problem of audio source separation using multichannel observation. Utilizing the sparseness of sound signals in the time-frequency domain is a successful approach to source separation that enables us to perform separation based on spatial features obtained from a microphone array. On the other hand, nonnegative matrix factorization (NMF) is also a promising approach for audio source separation, which performs separation based on spectral features. This paper incorporates the idea of NMF into sparseness-based source separation and proposes a novel approach to multichannel source separation based on both spatial and spectral features. Experimental results reveal that our proposed method improves the signal-to-distortion ratio (SDR) by 0.26 dB and the signal-to-interference ratio (SIR) by 1.96 dB compared with a conventional sparseness-based approach. In addition, our proposed model eliminates the need for a number of matrix inversions thanks to the sparseness assumption, and thereby requires a much lower computational cost than a previously-proposed multichannel NMF approach, which also utilizes spectral and spatial features.