Mixed Speech Signal Blind Source Separation Algorithm

During the past decade, Blind Source Separation (BSS) has been a focus in the research area of signal processing. Ranging from wireless communication to medical signals processing, to image enhancement and to audio mixtures separation, BSS is a powerful tool to tackle those problems because it can reconstruct the original signals from the observed signals without prior knowledge of the mixing system and source signals.BSS for audio mixtures is original intention of BSS technique, and also is a challenging problem in signal processing realm. In this thesis, the fundamental theories of BSS and main methods for audio separation are exploited and investigated.First of all, blind separation algorithms of linearly mixed source signals are studied. To deal with the difficulty for the classic BSS algorithms to choose a pretty step size, we propose the algorithm of blind source separation with an adaptive step size. Simulation results show that the algorithm with adaptive step size can effectively improve the performance of the blind source separation algorithm. Then an algorithm based on time-frequency distribution is proposed. In the algorithm, a fast joint diagonalization algorithm is adopted to improve the performance of the algorithm. This algorithm does not require pre-whitening of the data. Simulation results show that the new algorithm can separate sources correctly without usual pre-whitening of the data. Finally a natural gradient method is presented to separate convolved speech signals. An estimate of the score functions based on the generalized Gaussian distribution is adopted in this algorithm. The key advantage of the direct estimation of the score functions lies in the fact that it enables the algorithms to separate hybrid mixtures of sources that contain both super-Gaussian and sub-Gaussian signals successfully.