The Underdetermined Blind Source Separation Of Speech Signals

As an important direction in the speech and acoustical signal processing area,speech separation plays a high role in speech recognition and speech enhancement.In this dissertation,we analyze, summarize the previous work of BSS and make research on blind speech separation in underdetermined case to solve the problems of the traditional algorithm: the good sparsity requirement of signals, unsatisfactory results in the noise case . Here are the achievements:1.To solve the problems of the traditional algorithms based on sparse analysis: the low precision of the recoverd matrix and results, the method based on linear membership function is proposed to estimate the mixing matrix..To solve the problem, we propose a linear membership function based on the difference of vector angles to classify the data ,which has a good ability to capture the lines data cluster to.By obtaining the extreme value of the function,we can get the lines. And then we can get the mixing matrix by the lines. Experiment results show that compared with potential-function-based method, mixed matrix error decreased by 30%, and the SNR of the recovered signals increased by 5db using the proposed method.2.To solve BSS in the noise condition, a method based on noise compact factor is proposed. Given the impact of noise on data, we strength function ability of anti-noise by setting weights to put more emphasis on the more reliable data. Compared with other methods, this method can do better in the noise case. Experiment results show that compared with potential-function-based method, mixed matrix error decreased by 47%, and the SNR of the recovered signals increased by 4db using the proposed method.3.To relax the sparseness condition, the method based on K-SCA is proposed. Compared with SCA, K-SCA requires less sparseness of source. Based on this assumption, we can get the mixed matrix by find the the hyperplanes data cluster to. To solve the problem, we define the hyperplane membership function based on the difference of vector angles to classify the data which has a good ability to capture the hyperplane data cluster to. By obtaining the extreme value of the function, we can get the hyperplanes. Experiment results show that compared with other algorithms, the larger SNR of the recovered signals can be achieved by using the proposed algorithm.