Research On Speech Signal Underdetermined Blind Source Separation Algorithm Based On Sparse Representation

With the development of information technology,blind source separation has gradually become a crucial research technology in the field of signal processing.It is mainly aimed at separating the source signal based on the observed signal only when the source signal and the mixing parameters are unknown.According to the relationship between the number of source signals and the number of observation signals,blind source separation can be divided into three cases: overdetermined,positively deterministic,and underdetermined.In recent years,firstly,for the number of source signals is more than the number of observation signals under underdetermined conditions,which are most suitable for practical applications,its research and application value is the largest.Secondly,because of the blind source separation technology under overdetermined and positively deterministic conditions is very mature,Conditions for marching towards m-ore difficult underdetermined issues.Therefore,underdetermined blind source separation technology has become the focus of current research.A linear instantaneous mixture model for underdetermined blind source separation is selected.The “two-step method” based on sparse component analysis is used,and the algorithms for mixture matrix estimation and source signal reconstruction in the “two-step method” are studied separately.The improvement mainly includes the following two aspects:(1)A mixing matrix estimation based on WE-FCM mixing clustering is proposed.Aiming at the defects of the traditional fuzzy C-means clustering(FCM)algorithm,which are sensitive to the initial cluster center,easy to fall into local optimum,and vulnerable to noise interference,a mixing matrix estimation algorithm based on WE-FCM mixing clustering is proposed.The proposed algorithm firstly uses an evolutionary programming algorithm to estimate the initial clustering center to avoid being trapped in a local optimum due to the artificially given initial clustering center,and then uses the outlier factor obtained by the local outlier detection algorithm to apply it.The objective function of the FCM algorithm and the fitness function of the evolutionary programming algorithm allow the algorithm to converge quickly in areas with high sample density,but fail to converge in areas with low sample density,thereby avoiding the effect of outlier noise points on the clustering results.The MATLAB simulation results show that the proposed algorithm can effectively improve the estimation accuracy and algorithm stability of the mixing matrix.(2)A source signal reconstruction algorithm based on greedy double sparse dictionary is proposed.First,in the signal sparse representation phase,in order to solve the problems of traditional dictionary algorithms with large calculations,long running time and limited dictionary size,a greedy double sparse dictionary learning algorithm is proposed using a double sparse dictionary structure,which mainly trained samples from the signal domain.Transform to the coefficient domain to avoid a large number of operations and obtain an effective signal sparse representation.Then,a source signal reconstruction model based on compressed sensing is constructed.In view of the disadvantages of the poor stability of the orthogonal matching tracking algorithm and the susceptibility to noise interference,it is effective to filter out low energy points on the accuracy of the algorithm by adding an energy threshold during the algorithm iteration.and at the same time increase the stability of the algorithm.The MATLAB simulation results show that the proposed algorithm can effectively reduce the dictionary training time and signal reconstruction time while ensuring the accuracy of the reconstructed signal.