Multichannel Audio Signals Blind Separation Based On Low-rank Nonnegative Tensor Factorization

Now there are more and more demand for voice applications,along with the popularity of smart phones,such as speech recognition,music recommendation,Shazam Encore and so on.Therefore,in recent years,the research on audio signals has become more and more popular.One of the main aspects is mixed audio signals blind separation,it can be used in many aspects,such as speech recognition,speech enhancement,music classification and clustering,speaker recognition and automatic music etc.In this paper,we mainly research the separation of linear instantaneous mixed multi-channel underdetermined music signals,and propose nonnegative tensor decomposition algorithm based on low rank approximation algorithm.Firstly,the basic principles of NMF(Nonnegative Matrix Factorization,NMF)are briefly introduced.Then the low rank approximation algorithm is applied to the NMF algorithm,and its iterative updating formula is deduced,which effectively improves the computing speed of the NMF algorithm.On this basis,the basic principles of NTF(Nonnegative Tensor Factorization,NTF)are expounded,and two kinds of nonnegative tensor decomposition forms are introduced:nonnegative standard decomposition method(Canonical Polyadic)and nonnegative Tucker decomposition.The low rank approximation algorithm is extended to NTF algorithm,and its iterative updating formula is derived.Nonnegative tensor decomposition algorithm based on low rank approximation algorithm is proposed.Then we use the accelerated proximal gradient algorithm to solve the NTF problem,and improve the nonnegative tensor factorization algorithm,reduce the complexity of the algorithm and improve the convergence speed.Combining the low rank approximation algorithm and the accelerated proximal gradient algorithm,the NTF algorithm is optimized,and the algorithm speed is further improved.Finally,the accelerated NTF algorithm based on the low rank approximation algorithm is applied to the blind separation of audio signals.Good separation effect was achieved.