Research On Fast Blind Source Separation Algorithm Based On Frequency Domain Independent Component Analysis

As the most common and direct expression of information,speech signal is a very important part of daily life.However,most of the speech signals obtained from the outside world are mixture of different speech signals,from which effective information cannot be obtained directly.Therefore,there is an urgent need for a technology to separate mixed speech signals and independently select the sounds they are interested in.Blind Source Separation(BSS)algorithm is also created in this environment.The blind source separation algorithm was originally proposed to solve the cocktail party problem.With the continuous research of scholars,the algorithm has been widely used in speech signal processing,biomedical signal processing,mobile communication and other aspects.This paper mainly focuses on the application of blind source separation algorithm in binaural hearing system.Due to the high computational complexity of the algorithm,it is in contradiction with the low delay and low power consumption of the hearing system.Therefore,to apply BSS to hearing AIDS,the most important problem is to reduce the computational complexity of the algorithm.Aiming at this problem,two fast blind source separation algorithms are proposed to reduce the computational complexity and maintain the separation performance.The main work can be divided into the following parts:1.A fast BSS algorithm using DOA to estimate the initial value of the separation matrix is proposed to solve the problem of slow convergence of the algorithm.First,an independent component analysis(ICA)iteration is performed on the frequency bins in the frequency domain where spatial aliasing does not occur,to obtain a separation matrix and to estimate the DOA of the source signal.Then,the determinant of the mixed signal covariance matrix is used to select frequency bin selection in the entire frequency domain.If the initial separation matrix is the unit matrix,it is difficult for the algorithm to converge to the global optimum during the iterative process.Therefore,DOA of the source signal is used to initialize the separation matrix of each selected frequency bin,and ICA iteration is performed to obtain the separation matrix.Secondly,because the primary frequency bin selection may select the frequency bins with poor separation performance,this paper introduces the second stage frequency bin selection based on outlier detection to ensure the accuracy of DOA,and the removed outliers are classified into the set of unselected frequency bins.Next,the average value of DOA obtained from the final selected frequency bins is used to construct the separation matrix of the unselected frequency bins and solve the sorting ambiguity problem.Finally,the uncertain amplitude problem is solved for the separation matrix of all frequency bins,and the separation of mixed signals is completed.2.A fast BSS algorithm for frequency selection by frequency band is proposed to solve the problem of high computational complexity of the algorithm.The characteristic of the algorithm is to measure the characteristics of frequency bins from two aspects of energy and independence.Different frequency regions use different frequency point selection criteria,make full use of the characteristics of each frequency region,and get the best separation performance.In the low frequency region,the determinant of the mixed signal covariance matrix is taken as the selection criterion.In the high frequency region,MSC of mixed signals is taken as the selection criteria.Another characteristic is to use the normalized attenuation and delay parameters extracted from the separation matrix to solve the problem of inaccurate delay parameters in the high frequency region.Next,this paper introduces the second-stage frequency bin selection method based on normal distribution,and the frequency bins with poor separation performance are further discarded to ensure the accuracy of the normalized attenuation delay parameters.Finally,the normalized attenuation delay matrix is used to construct the separation matrix of unselected frequency bins,and the separation matrix of all frequency bins is solved according to the principle of minimum distortion.For the final frequency bin with better separation performance,the separation is completed.For the unselected frequency bins,a wiener filtering post-processing method is introduced to re-separate the signal.3.MATLAB was used to build the reverberation room used in the simulation experiment,and the traditional algorithm and the two fast algorithms proposed in this paper were compared under the binaural space.The experimental results under the condition of no reverberation and no reverberation show that the two algorithms presented in this paper not only have low computational complexity,but also have certain improvement in separation performance,which can better promote the application of blind source separation algorithm in hearing AIDS.