The Research On Blind Signal Separation Algorithm In Time-domain And Its DSP Implementation

Blind signal separation (BSS) is an important topic in the domain of signal processing and has many applications in practice, such as wireless data communication, radar, medicine and earthquake signal processing. The aim of BSS is to recover the original mdependent source signals from their mixed signal vectors, by constructing a certain transform. Here, "Blind" means that the information of sources and the mixing system is unknown. In this thesis, we only consider the situation of audio signals as the input signals.There are two classes of BSS. One is the linear mixture mode, and the other is the nonlinear mixture mode. The linear mixture mode deals with two classes, instantaneous mixing problem and convolution mixing problem. Our interest is mainly on the latter. In this thesis, we mainly research BSS in time-domain.An algorithm of real world separation for convolved non-stationary signals is given. An attractive feature of this method is that only one set of cross-correlation data is used and non-minimum phase systems can be treated. We analyze the model of separation system, separation criteria and give the deviation of adaptive algorithm of blind signal separation. The validity of this method has been confirmed by a computer simulation with the real speech signals.We realized a BSS system (two microphone inputs and two speaker outputs) based on the TMSC6701 floating DSP, using this algorithm. And discuss in brief the realization of our BSS algorithm on DSP. In the real-time processing, it has been proved that our blind signals separation system has obtained a satisfied separationresult.At last, we study the speech signal separation in the noisy environment. In order to avoid t he d eterioration of p erfonnance o f unmixture sy stem, w e u se t he 1 ong-term nonstationarity of speech signals, as well as the long-term stationarity of noise signal to cancel the effect that noise adds to unmixture system. Using the differential approach, the coefficients of the unmixture system can converge to their correct values. Its effectiveness, is shown by means of computer simulation with real speech signals as their mixture inputs.