The Research Of Blind Source Separation Algorithm For Speech Signal

Blind source separation is a process of recovering each individual original signal from a number of mixed-signals. Blind speech signal processing is the mind problem of blind source separation. In this paper, three kinds of blind speech signal processing models including normal, underdetermined and over determined are studied as follow:First, a blind separation algorithm PCA-ICA combined by PCA and ICA is used to solve the blind speech signal processing problem. In order to solve the slow convergence problem of ICA based algorithm and high computational cost due to excessive amount data, an blind separation algorithm based on PCA-ICA for speech signal is proposed. PCA is used to remove the second-order correlations among different dimensions of feature from original data. Using similarity coefficient matrix as the separation effect standard, the simulation experiment results show that the proposed method can reduce 90% of iterations and is 3 times faster compared with ICA with the same separation accuracy. Thus the ICA-PCA algorithm effectively solves the slow convergence problem of original ICA method.Second, a two-step sparse component analysis method is used to solve the underdetermined blind speech signal processing problem. At First, in order to get the sparse signal, transform the mixed speech signal to frequency domain by STFT; then obtains the cluster centers by fuzzy C-Means algorithm and estimates the mixing matrix; finally, recovers the source signals using the shortest path decomposition algorithm according to the mixing matrix. The simulation experiment results show that the two-step sparse component analysis can solve the problem of underdetermined blind separation for speech signal.Third, a blind separation algorithm combined by Wigner time-frequency distribution and joint diagonalization algorithm is used to solve the over determined blind speech separation.The traditional blind source separation algorithm of speech requests non-Gaussian, stable and independent speech signals, however, the actual speech signals are not ideal. A blind separation algorithm based on Wigner-Ville is proposed, first pre-white the mix speech signals, and then compute the Wigner-Ville Distribution matrix, finally estimate source speech signals by joint diagonalization algorithm. Simulation results show that the actual non-stable speech signals can be separated well using the proposed algorithm, even in noise environment that the SNR is below 20dB.