Research On Real-time Blind Speech Source Separation And Localization

The area of blind signal processing has long yielded great interest in signal processing community. Blind source separation (BSS) is a part of blind signal processing. It has been applied extensively in speech recognition, biological signal processing, digital communication and image enhancement and etc. The mixing styles of blind signals include convolutive mixing and instantaneous mixing. Currently, the research of instantaneous mixing is fully-developed, and separation results are very well. However, the speech mixing is convolutive mixing in echoing and reverberating environments. So the research of blind speech separation is a large challenge. The research in this thesis is mainly focused on adaptive blind speech separation in convolutive mixing and the problem of speech localization.In the first part, we propose a new separating structure that combines the beamforming and a decorrelation algorithm for BSS. In the stage of beamforming, the speech sources are localized by the cross-power spectrum phase method. Furthermore, a decorrelation algorithm is used to separate signals in the frequency domain. It is shown that this algorithm performs better than the conventional decorrelation algorithm in the time domain.In the second part, two algorithms based on high order cumulants are studied. The basic idea of the two algorithms is described by the characteristics of high order cumulants. The performance and computational complexity of the two...