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

Blind source separation refers to the process of extracting and isolating the signal of interest from the observed multiple mixed signals under the circumstance of the source signal and the mixed matrix prior knowledge are unknown.As a powerful signal processing method,it has a pivotal position in many fields no matter in the mentioned research or in the referred application.With the rapid development of blind source separation technology,it has very important research value and practical significance in the other scientific fields include of signal processing and neural network.Underdetermined blind source separation technology is an extensive and challenging signal processing technology.Under the condition that the number of observed signals is less than the number of source signals,the processing technology will face a bigger challenge,and the separation technology remains to be further developed.The underdetermined blind source separation technology under the linear instantaneous mixture model is mainly researched.The main problem is to study the underdetermined blind source separation problem through the sparse component analysis algorithm of “two-step method”,that is,the estimation of the mixing matrix is performed first,then the over-complete dictionary is trained according to the prior knowledge of the source signal,and the sparse representation of the source signal is obtained.And the compressive sensing signal reconstruction algorithm is introduced into the underdetermined blind source separation to realize the blind separation of the speech signal.The main research work of this article includes the following two aspects:(1)The accuracy of the mixed matrix estimation directly affects the separation effect of the source signal.For the fuzzy C-means clustering algorithm relies too much on the initial clustering center,which has the disadvantages of low accuracy and poor robustness in the mixed matrix estimation.In this paper,genetic algorithm and simulated annealing algorithm are combined to complement each other and a hybrid matrix estimation algorithm based on genetic simulated annealing optimized FCM(GASA-FCM)hybrid clustering and Hough transform is proposed.Firstly,the algorithm combines the advantages of global search and high precision of the simulated annealing algorithm and the powerful spatial search ability of the genetic algorithm.The cluster center point obtained by the genetic simulated annealing algorithm should be assigned to FCM,which can avoid the randomness of initial value selection.At the same time,in order to improve the precision of the mixed matrix estimation,the Hough transform is introduced to the center of each type of data obtained by clustering.The experimental results show that the proposed algorithm improves the precision of the estimation of the mixed matrix and the robustness of the algorithm.The estimation of the mixed matrix is feasible and effective.(2)The method of underdetermined blind source separation for speech signal based on the dictionary learning is studied.Firstly,a redundant dictionary is trained by adopting double sparse KSVD(DSKSVD)dictionary training algorithm,which can be of sparse representation and in which the sparse decomposition of the observed signal is carried out.Then it analyzed the equivalence property of underdetermined blind source separation and compressed sensing equivalence problems,built a model of underdetermined blind source separation based on compressed sensing,and also applied the orthogonal matching pursuit algorithm to reconstruct the signal to achieve speech signal underdetermined blind source separation.In the premise of guaranteeing the separation accuracy,DSKSVD algorithm reduces the computational complexity of dictionary construction,improves the validity of signal sparse representation and reduces the running time of reconstruction algorithm.A large number of experiments show that the algorithm is better than KSVD algorithm and online dictionary learning algorithm,which greatly improves the computational efficiency.