Research Of Blind Source Separation Algorithm For Speech Signal Based On Variational Bayesian

The Algorithm of Blind Signal Separation(BSS)is given more attention by many researchers,due to its wide application in research and engineering.In the field of signal processing in recent years,BSS algorithm has been widely applied and adopted for its favorable result.However,when separating speech signals,Independent Component Analysis(ICA)takes no consideration to the noise's interference to the mixed system,makes few uses of signal observed and prior information known,and neglects the intrinsic structural features of voice signal.Aiming at refining the above method,the thesis analyzed separation system of a mixed source by Independent Component Analysis of variational Bayesian,so as to separate the speech in much accordance with actual situation and to make it more applicable.Autoregressive modeling was used concerning the time structural feature of speech signal.Then,the thesis concluded with a variational Bayesian Independent Component Analysis basing on the autoregressive model.Finally,the effectiveness of the conclusion was demonstrated through simulation and assessment criteria system.The thesis can be divided into three parts:Firstly,a general idea of Blind Signal Separation(BSS)theory was introduced,including related principles and models,objective function and optimization algorithm of Independent Component Analysis(ICA),and the pretreated method of algorithm.The two classic ICA algorithms were reasoned and analyzed as well.Secondly,the variational Bayesian independent component analysis was then integrated into the separation work of mixed speech system with noise.And its prior information was fully used after Bayesian network and Bayesian inference was introduced.Then the whole ICA principle of variational Bayesian was deducted by variation approximation as to solve the complex probabilities calculation.Finally,the algorithm of the thesis concluded was proved to be more effective through simulation and assessment on the two classic ICA algorithms.Finally,catering to the time feature of speech signal,the thesis was proposed an algorithm with Variational Bayesian ICA basing on generalized autoregressive modeling.This algorithm can study time structure and system noise as a whole by modeling the time structure of voice signal through generalized autoregressive modeling.And the entire derivation of the theory was demonstrated.At last,speech signal mixed with noise was separated by variation Bayesian.Simulation comparison between the two results was made to prove that the modified algorithm has better separation result.