Study On Blind Adaptive Speech Enhancement Algorithm Based On H∞ Filter

IntroductionBecause the source of noise is numerous and the application situation is different,their characteristic is also not same. So it is hard to find out a general speech enhancementalgorithm, which can be use to an environment of various noises. And speech enhancementnot only has something to do with the speech signal processing technique but also involvethe person hearing, feeling and phonetics. So we must aim at the different noise, adopt thedifferent speech enhancement algorithm. More than 30 years, people has done a lot ofworks on speech enhancement techniques, and proposed various methods for differentcircumstance. This paper studies a blind adaptive algorithm based on H∞filtering forspeech enhancement.H∞FilterIn recent years, H∞Optimal Design has become a valid method in feedback systemdesign. One property is that the exterior disturbances no longer suppose to be fixed, butbounded. Another one is that the model of controlled plant is not supposed to be certain.H∞filter algorithm is to solve these two problems. In general, the model should belong to agiven known set. And the exterior signals including disturbed signal, noise of sensor andthe command signal are not known exactly, but the given known signal set. Base on thistheory, an H∞filter for the estimation of the original speech from the polluted speech and isstudied in this paper. It can promote SNR in certain degree.Blind adaptive filteringBlind adaptive filtering problem is concerned with filtering when neither speech nornoise is separately accessible nor the prior knowledge of speech and noise parameters. Thefilter consists of three subsystem: noise identification, feed-forward de-noising subsystem,and virtual feedback de-noising subsystem.Noise identification The various types of phonemes are mixed in a fashion that mayappear to an outsider as random jumps in frequency at time intervals of random duration tocreate human speech. So, speech signal is non-stationary over a relatively long timeinterval. But normally, noise signal is stationary. According to the features of speech andnoise, the approximated version of the estimation procedure for the noise power spectrumis obtained.Feed-forward de-noising subsystem Consider a set of weights that multiply the...