The Research Of Speech Signal Blind Source Separation Based On LPFT

As the main medium of communication between people, speech signal is widely used incommunication, control, speech recognition, military surveillance and other fields. With thedevelopment of electronic technology, there is more interference in voice signal acquisitionand transfer, because of the increasingly complex electromagnetic environment, therefore thequality of speech signal greatly improved. While traditional voice signal enhancement, noisereduction technology in multiple mixed speech signal processing are not effective, thereforeneed to find a better algorithm to separate the signal which people interested in from complexenvironments.This paper researched the LPFT theory and the spectrogram of the LPFT, discussed theinfluence of the cross-term, which produced by LPFT. By compared whit STFT and WVD,this article proved LPFT has high time-frequency resolution and less cross-term interferencein time-frequency analysis of speech signal, and more conducive to estimate the speech signalparameters (instantaneous frequency, instantaneous phase and amplitude) based on sine model.This paper proposed the speech signal blind source separation algorithm based on LPFT, byjoint diagonalize the time-frequency distribution matrix, separate the speech signal whichaffected by noise, improved the SNR of speech signal. And then separated the three originalsignal form three observed signals, proved that this algorithm has better noise suppressionability and better signal reduction ability compared with the classical FastICA algorithm.And this article also proposed an algorithm that BSS in frequency domain then sinesynthesis speech signal in time domain to solve the convolution mixing condition. In that part,deeply researched the theory of sine synthesis, and separate the frequency matrix to extractsinusoidal parameter by frame the time-frequency distribution matrix, then rebuild the speechsignal by those parameter. This algorithm does not need to estimate the hybrid filter, just needto get the sinusoidal parameters of the original signal to rebuild the speech signal. What’smore, there are a lot of signals can be represented as a set of combination form of sine wave,so the sine signal model can be extended to general, and the algorithm can be applied to more areas. In a word, this algorithm has a good application prospect.