Research Of Mixed Speech Separation Based On Blind Source Separation Algorithm

Blind source separation is a method to recovery the source signals by a transformationwhich is obtained by mixed data vector. The blind separation of speech signals which is anapplication model and important branch of BSS is a research hotspot in the area of signalprocessing, because of its practical value in speech recognition, denoising for mobiletelephone, aid-hearing and other applications. This paper researched on blind speechseparation with linear instantaneous mixture, studied the natural gradient algorithm, blindsource separation algorithm based on wavelet packets and blind source separation algorithmbased on spatial time-frequency distribution. The availability and superiority of thesealgorithms were proved though simulation experiments. The specific research works are asfollows:Firstly, this paper researched on a variable step-size natural gradient algorithm. To solvethe contradiction that fixed step-size natural gradient algorithm brings between convergencerate and steady-state error, the step-size adjusted automatically depended on separation degree.Two parameters were used as the coefficient of separation degree to control separate stabilityand convergence. Simulation results demonstrated that the algorithm had good separationproperty compared with fixed step size algorithm, convergence rate doubled, and steady-stateerror narrowed three times.Secondly, this paper proposed a blind source separation algorithm with combination ofkurtosis and wavelet packets. Wavelet packets transformation was adopted to reduce thesignals’ overlapped degree, that was, the mixed speech signals were decomposed into waveletpackets, and the node that had the highest kurtosis was the optimal wavelet packetsdecomposition node since the kurtosis is the measure of non-Gaussian nature. Thereby, itreduced the signals’ overlapped degree in wavelet packets domain. Then the separation matrixwas calculated by using FastICA algorithm iteratively, and the source signal estimations wereobtained finally. Simulation results demonstrated that the separation performance improvedclearly when compared with FastICA algorithm in time domain and other wavelet packetsFastICA method.Finally, this paper researched on blind source separation algorithm based on SPWVDwith single autoterms selection. To improve the estimation precision of mixed unitary matrix,single autoterms were selected through spatial time-frequency distribution matrix eigenvalues.Those matrices with single autoterms were joint-diagonalized and the least diagonalizationmatrices were eliminated at each iteration. The noise variance was estimated, then noisecompensation was used to reduce the interference of noise. Simulation results demonstratedthat this method had good separation performance, and had good robustness.