The Research Of Speech Enhancement Based On Wavelet Transformation

Speech is the most important means by which people transmission information eachother, but speech is affectec inecitably by the noise of the surrounding enxironment andtransmission mediun bringing in the voice communication process. In addition, the speechsignal, used in the process of speech application that is speech recognition, speechsynthesis and so on, will be subjected by the external noise interference. So it is necessaryfor the noisy speech signal denoising. Therefore, speech enhancement method of waveletthreshold is currently more popular denoising methods. In this paper, on the basis ofspeech signal and noise characteristics, we focus on wavelet threshold denoising to do thefollowing research:First of alll, we study the wavelet basic functions and level of wavelet decompsition,which are teo more elements of the wavelet threshold denoising. According to analyse thewavelet basic fuanctions of different wavelet families, and reserch speech signalcharacteristic and the analysis of the decomposition level, we determine the wavelet basicfunction and the decomposition level which are more well to denoising the noisy speechsignal. The experiments prove the correctness of the analysis.Secondly, in view of th shortage of the tranditional thresholding alogrithm, this paperstudy a new thresholding alogrithm. The improved thresholding alogrithm iscomparatively good at continuity and smoothness.This paper proves the alogrithm intheory to overcome the disadvantage that is shake, because of when the waveletcoefficient absolute value is less than the threshold the wavelet coefficient is set to zero inthe tranditional thresholding alogrithm. Simulation results show that the improvedthresholding alogrithm is better than the thresholding alogrithm in denoising effect, and isbetter than other thresholding alogrithm in SNR ratio index.Once more, this paper studies and reaserchs four different threshold selection. On thebasis of this, we give a new threshold selection method based on the spectral subtractionthought. The new method, firstly gains the pure noise segments of the noisy speech signal,then processes the pure noise segments and noisy speech by the same method, finally determiane the threshold according to that this paper studies the threshold selectionmethod. Simulation experiments prove that the new threshold selection method iseffective.Finally, in view of the noisy speech of the strong background noise, the effect ofsimply using the wavelet threshold denoising is not very good. The paper studies a speechenhancement method of wavelet threshold based on the ICA. It is, firstly the noisy speechsignal for ICA pretreatment, and then the use of wavelet threshold denoising. And thesimulation experiments prove that the new method is effective.