Multichannel Audio Signals Blind Sepatation Based On Nonnegative Tensor Factorization

Blind signals separation (BSS) is the process of separating the original signals from their mixtures without any knowledge about the mixing system or the signals. Mixed audio signals separation is the original intention of blind signals separation, and also a research emphasis and difficulty of the signal processing field. In recent years, separation of the music signal received great attention, because of its important applications in technologies about melody extraction, music information retrieval, and music coding.Linear instantaneous mixed multichannel music signals blind separation is the research content of this thesis. Nonnegative tensor factorization (NTF) algorithm based on the parallel factor structure is proposed. Three means of nonnegative matrix factorization (NMF) are introduced simply, and the principle of NMF is analyzed. Because of the advantages of NMF, the NMF ideas are applied to tensor factorization. The multichannel music signals model is transferred into parallel factor structure, and then derives the multiplicative update rules of the NTF, which accomplishes the signals blind separation. In order to simplify the separation process and improve the effect of separation, the cluster-NTF algorithm that uses label matrix is brought up, which allows clustering of the components within the decomposition, as opposed to after decomposition.Furthermore, sparsity of audio signals in the time frequency domain is used to make improvement on cluster-NTF algorithm. A simple way is proposed for detection of points in time frequency domain where only single source contributions occur. The hierarchical clustering algorithm is used for estimation of the mixing matrix. Then, the estimated mixing matrix is applied to the cluster-NTF algorithm directly. The estimated matrix shows the mixing proportion of original sources, and reduces randomness of separation results, which makes a better result. At last, comparing the algorithm proposed in this thesis with sparse component analysis (SCA), cluster-NTF algorithm has better separation results.