The Algorithms Research On Blind Separation Of Speech Signals

Blind source separation (BSS) is a process of recovering original signals from the observed signals without knowing the knowledge of the mixing system and source signals. It finds broad application owing not to require the priori knowledge of original signals. The blind separation of speech signals which is an application model and important branch of BSS is a research hotspot in the area of signal processing, because of its practical value in speech recognition, denoising for mobile telephone, aid-hearing and other applications.This thesis is focused on the research of blind speech signals separation. The main research works are as follows:First, this thesis introduces the background and significance of research on BSS, and expounds its research status and some important areas of applications. And the thesis analyzes the fundamental theories of BSS, including three main mathematical models of BSS, independent component analysis, the preprocessing methods of data, the performance criteria for separation, the fundamental properties of speech signal, and so on.Then this thesis analyzes and researches BSS algorithms for different mixing models, and it specially introduces three typical BSS algorithms, i.e., jointly approximate diagonalisation of eigenmarices(JADE), an sparse BSS algorithm for pure-delay mixing model(we name it Sparse-delay) and joint block diagonalization(JBD) which is a frequency-domain algorithm for convolutive mixing model.Finally this thesis researches and explores the problem of blind speech source separation by simulation experiments. JADE algorithm, Sparse-delay algorithm and JBD algorithm have been used in simulation experiments by which we investigate their performance for different kinds of synthetic mixing speech signals and for speech signals recorded in real world. When separating synthetic mixing speech signals, we respectively investigate the performance of these three algorithms for separating instantaneous mixing speech signals, pure-delay mixing speech signals, convolutive mixing speech signals and time-domain correlated sources mixing speech signals. When separating speech signals recorded in real world, we investigate the effects of the distance between speaker and microphone and the way to speak on separation results. By simulation experiments, we arrive at some conclusions and inspirations which are useful for further research in future.