The Study Of Blind Source Separation Of Speech Signals

Blind signal separation is a new technique of signal processing aimed at retrieving a system's unknown information from its output only, which can be used in array processing and data analysis. There are many potential applications, especially in wirless communication, medical signal processing, speech and image processing and radar signal processing. In recent years, blind signal separation has been an attractive trend in the area of academia. In this thesis, the theory and applications of speech signal blind source separation has been studied. The thesis consists of following parts:First, the mathematical description of BSS is given after a relatively thorough introduction to the historical perspective of BSS, including the mathematical modle of BSS, the assumptions made about BSS problems and the mathematical theory and methods commonly used in BSS.BSS includes linear and non-linear mixed signals. And the linear models include instantaneous model and convolution model. In the practical environment, many kinds of signals such as speech signals exist as convolutive mixed signals. Many convolutive BSS algorithms are based on the instantaneous BSS algorithms. Thus the thesis firstly descripes two the instantaneous BSS algorithms and prove its effectiveness in experimentThen the thesis focuses on the convolutive blind separation of acoustic signals. Nowadays most blind signal separation algorithms are based on the High Order Statistics, because it is only high Order Statistics that can make the outputs are independent with each other.The Second Order Statistics can only make the outputs decorrelations with each other. But Second Order Statistics can be used to separate the non-stationary signals and colored signals. Make use of the unique traits of acoustic signals that they are both non-stationary and colored signals, we try combining the traditional non-stationary signals separation algorithm and colored signals separation. At last we proposed a blind acoustic signal separation algorithm based on Second Order Statistics and demonstrate its effectiveness in experiment.At last, according to the character of this algorithm, a detailed analysis is maked and then compares it with some mature algorithms; prove that it has advantage in speech blind separation.The work and experience in this thesis has an universal meaning in speeh signal process.