Research On Blind Separation Of Speech Signal With Muti-channel

In recent years there has been an increasing focus on the research of Independent Component Analysis (ICA): a new method of Blind Source Separation(BSS), which is a promising approach for source estimation without any prior knowledge except the observed signals. The application of ICA is concerned with many disciplines, such as wireless communication, medical signals processing, image enhancement and speech separation.Independent component analysis (ICA) was introduced for the first time in early 1980s. In mid-1990s, some new theory and algorithms concerned with ICA were proposed in succession, which has attracted broad attention from most researchers. One of the important applications of ICA is the blind separation of multi channel speech signals.The thesis focuses on the ICA theory and algorithm for the blind separation of muti-channel mixed speech signals. The main works in the thesis are as follows:(1) The ICA approaches based on kurtosis maximization and Infomax algorithm are studied. Detailed theoretical analysis and experimental simulation are given; and the validity of algorithm has been tested.(2) Instantaneous mixed speech signals are separated by using the algorithm based on kurtosis maximization and Infomax. The convolutive mixed speech signals are separated in time-domain by making use of the algorithm which is extended from Infomax. The simulation indicates that the separation system in time domain is hard to get the ideal result for the sake of limitation of the algorithm itself.(3) The technique of speech blind separation in the frequency domain is studied. After that, we put forward an ideal of blind source separation for each frequency point using extended Infomax algorithm. The difficulties caused by the selection of optimal frame size and indeterminacies of ICA(permutation and scaling) have been solved, and the new speech blind separation algorithm in the frequency domain is described in detail. The experiment results show that the proposed method has good performance of blind separation for real world data.