Study On The Algorithm Of Convolutive Blind Source Separation For Speech Signal

The main idea of blind source separation(BSS) is the process of extracting and recovering various source-signals which can not be observed directly from a number of mixed signals. According to the difference of mixed signals, blind source separation can be divided into several different ways, such as linear or nonlinear, convolution or non convolution, positive or undetermined. After 20 years of research, the blind source separation of linear mixtures has been studied very thoroughly, and it has been used in some areas. One of the most commonly methods is independent component analysis(ICA). However, with the development of research, the linear instantaneous blind source separation can't be used in some situations in life, especially in the hybrid transmission of speech signal. The transmission of speech signal includes reverberation and interference, it will have a greater impact on each other for multi-channel mixed signals, and its mathematical model is more close to the convolutive model.At present, more and more attention has been paid to the study of the blind source separation of speech signals all over the world. However, the complexity of speech signal and convolution mixture causes many problems. Such as the effect of blind source separation is not very ideality, low practicability and high computational complexity. And the signal separation between two or more signals is particularly complex, and the stability is not good. It means that the research of this technology is still in the initial stage now. How to realize the convolutive blind source separation of the speech signal with high performance and decompose multiple observational signals efficiently have become the hotspots and difficult problems in the field. This thesis will study these problems, and mainly includes the following aspects:First, the development history and research status of blind source separation and convolutive blind source separation are described in detail. The theory of blind source separation and convolutive blind source separation algorithm is introduced systematically, and the theoretical knowledge of the two algorithms is studied, including the basic model, the constraints of independent component analysis(ICA), the objective function and basic optimization criterias and so on. Several typical blind source separation and convolutive blind source separation algorithms are introduced, and finally the appraisal criterions of convolutive blind source separation algorithms for speech signals are summarized.Second, for the ranking uncertainty of convolutive blind source separation in frequency domain, a method based on multi-band energy sorting algorithm was proposed. Firstly, through the short Fourier transform(STFT) for the mixed signals, an instantanneous mixing model of each point is built in frequency domain and then uses independent component analysis. After that, a multi-band energy sorting algorithm which is based on combining energy-related and the direction of arrival(DOA) sorting methods is proposed to solve the uncertainty of ranking. Then the split speech spectral method is used to solve the uncertainty of magnitude to get the proper sub-signals for each frequency. Finally, the source singals is separated out through the inverse short Fourier tranform. The convergence time, separation efficiency and robustness of the algorithm are simulated by experiment simulation.Third, aiming at the problem of frequency point inversion caused by blind source separation error in frequency domain sorting, a kind of sorting algorithm for inter-frequency correction is proposed based on amplitude correlation sorting algorithm. First of all, according to the analysis of the characteristics of the source signal short time Fourier transform(STFT) and the corresponding frequency signal separation, then find out the similarity of frequency points; secondly, through the establishment of a threshold to find the error frequency points. After that, according to the high similarity principle of adjacent frequency, the ill frequency points are corrected by the method of replacement; Finally, the signals are separated through the inverse transformation of sorted signals; The experimental results show that the separation performance of improved method is better compared with Murata's algorithm, but slightly worse in convergence speed.Fourth, aiming at the bad performance of classical separation algorithms for multi-channel signals, this thesis studies the entropy measurement method of feature alignment, a kind of convolutive blind source separation algorithm based on feature permutation entropy is proposed. The algorithm uses feature permutation entropy measurement to solve the uncertainty of separation. First of all, the complete feature of separated signals is calculated from the first frequency, and then the feature subsets corresponding to each separated signal are calculated in turn. After that, the entropy difference corresponding to each separated signal is compared in turn, and the smallest one is picked out. Finally, the remaining separated signals are sorted in the same way traverse all frequency points. The algorithm's convergence speed is slightly slower than the classical algorithm. However, it can solve convolution mixing problem of three signals successfully, and it has important value to the research of the convolutive blind source separation.This thesis is closely related to the new research direction of the convolutive blind source separation, and the convolutive blind source separation algorithms are analyzed and studied in detail. And to some extent, the convolutive blind source separation algorithm of the speech signal is improved. The theoretical proof and computer simulation are carried out for the new algorithms mentioned above, and different convolutive blind source separation algorithms are compared and analyzed from different perspectives. The three different algorithms proposed in this thesis have different advantages, and the research has a certain practical value and application prospect.