Voice Mixed Signal Blind Source Separation Research

Blind source separation of speech signals is a process of recovering source signals only by the observed mixing signals and the statistical independence hypothesis of source signals under the conditions of not knowning source signals’ distribution and the mixing channel parameters. Blind separation of speech signals has become one of the attractive trend in the field of signal processing in recent years, owing that it has a wide application prospect in speech recognition system, computer hearing, mobile speech data communication and television and telephone conference. Blind speech source separation technology of instantaneous mixing is more mature, however, because of the time-delay and reverberation, the mixture of speech signals in fact is not instantaneous mixing, but convolutive mixing, it is difficult to separate them, so many researchers at home and abroad are making some studies for this issue.The fundamental theories of blind source separation are introduced and the main algorithms of instantaneous mixed blind source separation model and convolutive mixed blind source separation model are investigated in this thesis. Particularly, the independent vector analysis model is used for convolutive mixed blind source separation in the frequency domain, and derive the independent vector analysis natural gradient algorithm.For the problem of slo5tvw convergence of the independent vector analysis natural gradient algorithm,firstly according to the relationship between the iterative step size and the change of estimated cost function, an adaptive step size independent vector analysis algorithm based on steepest step size gradient is proposed; Then according to the relationship between the iterative step size and the change of estimated separated matrix, an adaptive step size independent vector analysis algorithm based on the estimate function is also proposed.The separation of artificial mixed speech signals and the separation of real world speech signals prove that the effectiveness of the above algorithms.The results show that two proposed adaptive step size independent vector analysis algorithms have better performance and faster convergence compared to the independent vector analysis natural gradient algorithm.