Research On Separation Of Singing Voice And Music Accompaniment

In recent years, due to the rapid development of digital music, demands for singer identification and lyrics alignment is increasing. and voice separation as an important part of these techniques attracts more and more attention. Accurate voice separation can effectively improve singer identification and audio retrieval. Thus, the study of the voice separation technology for audio signal processing is important.In this thesis, under the theoretical framework of the separation matrix decomposition of the voice, we carry out a detailed study based on non-negative matrix factorization, robust principal component decomposition, low-rank decomposition vocal separation. On this basis, focusing on the disadvantage of the robust principal component decomposition technique,we proposed a two-stage separation matrix factorization algorithm. Finally, for the well expression of vocal and background music,we use the deep neural network algorithm, it can effectively improve the accuracy of voice separation. The main research work and innovation of this thesis is as follows:1. Study on voice separation based on matrix decomposition algorithm1) First, we study the voice separation algorithm based on matrix decomposition, summed up the basic framework matrix decomposition algorithm. For the typical matrix factorization algorithms, such as non-negative matrix factorization algorithm, robust principal component analysis, and low-rank decomposition algorithms, we do some researches and simulations. The results showed that the voice separation algorithm based on robust principal component analysis is more thorough, and it is robust.2) For the incomplete problem of voice separation based on robust principal component analysis, this thesis proposes a two-stage matrix decomposition algorithm, the first step to use robust principal component analysis algorithm do voice separation, then use the harmonic source separation algorithm to separate the isolated incomplete vocals, and finally regroup the results of two separate, to obtain the final separation of the accompaniment parts and vocal parts. Experimental results show that the two-stage matrix factorization algorithm is not only ensure the robustness of the original premise, but also effectively improve the separation performance.2. Study on voice separation based on deep learning algorithmFor the poor expression of Matrix factorization algorithm for accompaniment, this thesis use the theory of deep learning,we proposes a voice separation algorithm based on deep neural network. Due to the layer-wise and nonlinear activation function, it can well express the vocal and background music. At the same time, the deep recurrent neural network can use the context of music, we study the voice separation based on the deep recurrent neural network. Experimental results show that it can be more completely isolated from the voice with accompaniment and high stability.