On Blind Speech Separation Based On Short-Time Synthesis Overlap-add And It's DSP Implement

Blind signal separation (BSS) is applied extensively in speech separation and acoustic recognition, biological signal processing, digital communication and image enhancement etc. Thus it has become a hot spot in signal processing field recently. The mixing style of blind signals includes convolutive mixing and instantaneous mixing. Currently, the research of instantaneous mixing is fully-developed, and separation results is very well. Its main methods include Independent Component Analysis (ICA) and Primary Component Analysis (PCA). However, the speech mixing is convolutive mixing in reverberating or echoing environment. Thus the implementation of BSS is extremely complicated due to three factors: the impulse response of reverberating or echoing environment has a duration of several hundred milliseconds; the term "blind" refers to the fact that very little is known about the source signal or the way that they are mixed together; the system has to be adaptive. So the research ofblind speech separation is a large challenging.People develop a lot methods about blind speech separation. The research of frequency domain is more common. The time-domain convolutive mixing can be transformed to an instantaneous mixing in the frequency domain by Fourier transform. We can use the methods of instantaneous mixing in time-domain, in order to carry out the BSS of convolutive mixing. But the contradition of spectralresolution and permutation alignment will appear in blind speech separation of frequency domain when using ICA. Any side of the contradition will degrade the separation performance. In order to solve the contradition, we introduce the short-time synthesis overlap-add method to BSS, which trades off effectively this contradition and enhances evidently the separation performance. The length of window is equal to the length of FFT transform in Conventional methods. But the length of FFT transform is two times longer than the that of window in short-time synthesis overlap-add method. So the spectral resolution is also twice. We validate the simplicity and feasibility of the method and show very good separation results through the experiments of simulation.Furthermore, a blind speech separation system has been made using TMS320C6701 in developing algorithms and FPGA(FLEX10K10) in designing peripheral circuits and other device needed. So it has been possible to the short-time synthesis overlap-add method been implemented to BSS in real-time.