Multi-source separation based on non-negative matrix factorization and source distribution

Abstract

In the scene where multiple sound sources simultaneously occur in a reverberant room, it is difficult to separate each sound source effectivity. In this paper, a multiple sound sources separation method based on non-negative matrix factorization (NMF) and angle distribution is proposed to solve this problem. The proposed method can be divided into two parts. The first part uses the NMF to remove the reverberation components in the signals with multiple sources sound simultaneously. The second part uses a method, which regards the sparse components as the primary processing parts while the non-sparse zone is used as the auxiliary parts to perform multiple source separation. Specifically, after the dereverberation step, the recorded signals are divided into sparse and non-sparse zones in the time-frequency domain. The separation step is processed according to the different characteristics of these two kinds of zones. As for the sparse zones, which consist of the components from a single source, the angular distribution probability statistics are used to carry out the separation. For the non-sparse zones, the regression prediction is introduced to separate components from different sources in the recorded signals. Finally, the signals belonging to different sources in the two zones are properly matched and post-processed. In the multiple source separation method proposed in this paper, the microphone recording signal is processed in layers, which reduces the computational complexity while improving the quality of the separated sound source signals. Experiments on both simulated data and data recorded in an actual acoustic chamber have proved that the proposed method could achieve good results, especially in the actual scene with large reverberation, its performance is stable.