A Research Of Blind Separation Of Speech Signals Based On A Convolution Model

Blind Source Separation (BSS)is a technique for estimating the original source signals from the observed mixtures at sensors. According to the different forms of signal mixings, blind source separation involves instantaneous mixing and convolutive mixing. While the early method of instantaneous blind signal separation is applied to an ideal environment, the convolution mixture model is closer to the actual environment. The convolution model typically uses a method which converts time-domain convolution into the frequency-domain mixed mode by using Fourier Transformation, then carries out an independent component analysis. However, a lack of prior knowledge on the original signal and signal path in the blind separation process results in ambiguities of amplitude and permutation.Therefore, this dissertation is focused on the research of frequency-domain blind separation of speech signals, which can be outlined as follows:First, this dissertation elaborates on the background and significance of research on BSS, research status and some important areas of applications, then the thesis analyzes the convolutive mixtures model, blind deconvolution problems and second-order statistics of convolutive mixtures in frequency domain.Second, we have introduced time-domain correction method, the complex domain ICA based on polar coordinate for the complex domain instantaneous separation algorithm. Then this thesis focuses on the Joint Approximate Diagonalization of Eigen-matrices (JADE) theory and algorithm process, setting forth the observed point dynamics model and the solution to complex domain ambiguities, i.e. ,the ambiguities of amplitude and permutation.Finally this dissertation researches and explores the blind source separation of synthetic mixing speech signals and speech signals recorded in real world by conducting simulation experiments. With simulation experiments, we arrive at some conclusions and inspirations which are useful for further research in this area.