Research On Blind Signal Separation Of Audio Signals

Blind Signal Separation (BSS) is mentioned more and more in recent decade. The research can be divided into two classes: 1) the separation of instant mixed signals, 2) the separation of dynamic mixed or mixed signals obtained after convolution of original signals, is called as blind deconvolution or blind identification. Because the blind processing techniques do not need a prior information of the original source signals and the channel, they can be widely applied in wireless communication, telephone conference system, medical analysis, geological detection, image enhancement and image identification.In this thesis, the theories and algorithms of BSS are discussed in details for the case of instant mixed signals and the case of mixed signals after convolution. The main contributions of this thesis are as follows:(1) For the separation of the instant mixed signals, we give a new algorithm based on second-order statistics. The new algorithm looses the assumption of existed algorithms, which require the source signals must be mutual independent (i.e. spatial independent). Furthermore, the computational complexity of the new algorithm is low. The experiments demonstrate that the audio signal can be extract efficiently by the new algorithm when the background noise is correlated .(2) For the separation of the instant mixed signals, we give a new adaptive BSS algorithm based on natural gradient. Compared with the existed Equivariant Algorithm of Signal Identification (EASI), of which the main assumption is that the auto-correlation matrix of source signals is an identical matrix, the new algorithm only requires the auto-correlation matrix is a diagonal matrix. Therefore, the new algorithm can be applied more widely. The experiments demonstrate the efficiency of the new algorithm.(3) For the separation of the mixed signals after convolution, we combine the matrix analysis, stochastic natural gradient descend method and high-order statistics to get a novel blind deconvolution algorithm of one-input one-...