Mixed Voice Blind Signal Separation System Based On Independent Component Analysis

Blind signal separation (BSS) is a new powerful technique in modern signal processing. Nowadays, it has been wildly applied in diverse fields such as speech signal processing, image processing, multi-user communication, array processing and biomedical signal processing. As main technology of solving BSS, independent component analysis (ICA) is used to extract the mutual statistical independent sources. In this paper, we focus our attention on the application of ICA in instantaneous BSS system and convolutional BSS system, and FastICA algorithm and FastICA BSS system are proposed accordingly. Experiments with real recordings have been carried out, showing the effectiveness of the algorithm for real-world signals.Detailedly, in this article we firstly stress our research on instantaneous BSS system, and related simulations are carried out with speech signals. As for the time-delay and convolutional mixture BSS issues, we forward our research on the convolutional BSS system in time domain. Finally, FastICA in time domain is derived based on instantaneous BSS system, and convolutional mixture BSS algorithm in frequency domain is obtained based on convolutional mixture BSS in time domain. Then we combine the two algorithms together and present the complex-value FastICA algorithm in frequency domain, which is a new BSS technique. We merge preprocessing in time domain and correlation coefficient solution in time domain into this new algorithm and present Time-Frequency domain FastICA BSS System. We also apply this system into experiments with real recordings. Compared with the convolutional BSS system in time domain, this system effectively solve the problems on iterative duration and real-time property. Experiments with real recordings show that the system introduced here significantly improves the systematic performance on accuracy and convergence speed.Creative ideas in this thesis are listed as follow:For the issues on BSS in real environment, we present complex-value FastICA algorithm in frequency domain, which is a new technique based on Independent Component Analysis(ICA) theory and its applications of 'FastICA...