Research On Convolution Blind Source Separation Algorithm Of Voice Signal

In the modern communication system,the basic form of signal is often expressed as unknown or only a small amount of prior knowledge;and the communication signals are often reflected as a mixture of multi-channel signals,which has the characteristics of non-stationarity,non-linearity and non-cooperation.Because of this complex signal model,it becomes very difficult to observe,estimate,separate and extract these characteristic values.Therefore,this dissertation studies the blind source separation technology of voice signal based on signal correlation principle,so as to achieve the effective separation of voice signals from multi-channel mixed signals.The preprocessing method of mixed voice is studied and the theoretical basis of blind source separation is analyzed in this dissertation.Because the algorithm that the mixed voice signal is separated in the time domain by convolution processing will be very complex,therefore,the blind source separation algorithm have been studied from the perspective of frequency domain convolution method.Firstly,in the dissertation the time domain convolution blind separation issue is transformed into the instantaneous blind separation process at various frequency points in the frequency domain,and it uses the traditional instantaneous algorithm to reduce the computation of mixed blind signal separation process.Secondly,since the frequency domain algorithm introduces the problem of sorting uncertainty,the dissertation studies the characteristics of the signal separated by the short-time Fourier transform at each frequency point,and analyzes the amplitude correlation of the traditional adjacent frequency points.The sorting method proposes an improved algorithm for the amplitude-dependent sorting of the threshold value judgment using the traditional method to separate the undesired frequency points.The simulation results show that the proposed algorithm based on threshold method can effectively improve the effectiveness of blind source separation.Finally,aiming at the problem of poor robustness of high correlation signals in the sorting process,an improved K-means algorithm based on F-norm distance is studied to realize the re-sorting method of bionic signals in order to improve the robustness of sorting process.