Algorithm Research Of Voice Singal's Blind Source Separation

Blind Source Separationis a hot research topic in the field of signal processing,this technology has been applied to many domains,such as radar,image enhancement,medical signal processing,etc.Speech signal is the main carrier of information transmission,it has a great significanceto recover the mixed speech signal by Blind Source Separation.This articleaim atresearch the speech signalBlindSeparationalgorithms which based on instantaneous mixture and convolutive mixtures,The main work is as follows:After expound basic theory of Blind Source Separation,through a lot of experimentscompare the separation performance of NM-ICA,NMI-ICA,ML-ICA,notice when under the similar separation performance,NM-ICA algorithm has faster convergence speed.Research THINICA algorithm and EFICA algorithm which repose on independent component tanalysis,at the same time,WASOBI algorithm which based on joint diagonalization is discussed.The algorithmsare applied to the blind separation of simulation data and artificial speech signals,the results shows that the separation performance of THINICA algorithm is better than EFICA algorithm and WASOBI algorithm.When analysis EFICA algorithm and WASOBI algorithm's assumption for source signal statistics characteristics,a joint independent component analysis algorithm called EW-ICA algorithm was proposed.Using simulation data and artificial mixed speech signal separation to test algorithms,the result shows EW-ICA algorithm's separation performance is better than EFICA algorithm,WASOBI algorithm and THINICA algorithm.Through a large number of simulation experiments,clearNon Gauss property of source signals and Noise intensity in mixed signal's effect on the separation performanceof EW-ICA algorithm and THINICA algorithm.The result shows when the non-gaussian of the source signal is stronger,the noise intensity in the mixed signal is smaller,the algorithm's separation effect is better.For convolutive blind source separation,study an algorithm named TFBSS algorithm which based on time domain.An independent vector analysis mode are proposed,on the basis of this model,FASTIVA algorithm was discussed.Research a blind separation algorithm for mixed audio signals,called TCDBSS algorithm.Using the mixed voicerecorded from realenvironment with noise and no noise to test the above mentioned deconvolution algorithms,the outcome indicateTCDBSS algorithm has better effect on separation performance,also has faster convergence rate and stronger noise immunity.