Study On Methods Of Digital Speech Signal Enhancement

Two kinds of methods about single channel speech enhancement based on linear prediction:delayed linear prediction and no delayed linear prediction method were investigated in this paper.For linear prediction method, how to improve pitch detection accuracy in the noisy environment isthe key problem. Recursive formula of average magnitude difference function on a sample basiswas used in this paper for pitch estimation. By doing that the problems of conventional methodsuch as discontinuity between frames, unsuitable processing of the unvoiced sound and transitionregion from voiced to unvoiced sound or vice versa can be solved, so pitch detection accuracy isimproved. It demonstrates more robust for pitch estimation than conventional AMDF in a noisyenvironment. For no delayed linear prediction method, linear prediction filter with long taplengths, a cascaded circuit connection is presented. In a colored noise environment, the noisesuppression effect is enhanced by adding a method of controlling the coefficients in the algorithmused for linear prediction. The principle of the two method are described in the paper, the noisesuppression effect for the noisy speech which is contaminated by white noise and colored noise isconfirmed by simulation using the MATLAB software. The comparison of the two methods ispresented in the last of the paper.