Research On Blind Separation Of Convolutive Mixture Of Speech Signals

In recent years, blind signal processing has received considerable attention in intelligent signal processing fields, among which blind source separation (BSS) is one of the most important topics. Blind source separation aims to estimate the source signals and/or the mixture system based only on the observed mixtures signals, it has many potential applications in communication, biomedical engineering, speech enhancing, image reconstruction, and finance data analysis domains, just name a few. Along with the development of research work on blind signal separation, the problem of blind separation of convolutive mixture of speech signals has been one of the most challenging tasks in blind signal processing area. The thesis focuses on this problem based on some of the former research results.The problem of blind separation of speech signals is introduced including the fundamental model and some typical approaches. The convolutive mixture model is a more practical model considering the time delay and reflection effects during the propagation of speech signal and filter effects of microphone arrays. The thesis discusses the problem of speech signal blind separation by exploiting the temporal characteristics of speech signal.The conventional methods solve the problem of blind separation of convolutive mixtures in the frequency domain, but it may lead to the problem of frequency domain permutation indetermination, which was due to the inherent indetermination of permutation of separated signal, and the noise introduced during the process of inverse transformation. In the thesis, this problem is tacked with from the temporal domain perspective, and two methods are proposed. The first method defines the temporal cost function directly, computes the separation matrix by the joint diagonalizing spatio-temporal relation matrices consisted of observation data and an iterative optimization procedures to realize the separation of signal, it makes use of the nonstationarity property of speech signal. The second method firstly transforms the convolutive BSS into instantaneous BSS, then generalizes the joint approximate diagonalization algorithm to achieve blind separation of speech signals. The thesis also proposes a global optimization whitening algorithm by exploiting the short-time stationarity of speech signal, which improves the separation performance of convolutive mixture of speech signals under strong noise circumstances.