Classifying NMF components based on vector similarity for speech and music separation

This paper presents a nonnegative matrix factorization (NMF) components classification algorithm for single-channel speech and music separation. Music only and music-speech mixture segments are firstly classified from the audio stream via audio segmentation technique. Then NMF is applied for signal decomposition. The basis matrix of the NMF output of music only segments provides the prior knowledge of music component in the mixture signal. NMF components, i.e. basis and gain vectors of the mixture signal are classified into speech and music based on the vector similarity between each basis vector and the priori music basis matrix. A set of SNR-dependent thresholding coefficients are empirically determined for the classification. The separated speech and music signals are reconstructed from the respectively classified NMF components. Experimental results show the effectiveness of the proposed method for speech and music separation, and its superior performance over the traditional NMF-based separation methods.