The Denoise Research Of Speech Based On The Wavelet

This thesis researches the denoise algorithm of speech based on wavelet. Wavelet transform which can auto adjust the wide of analysis's window depend on the frequency's change has swiftly and violently developed since the middle term of 1980s. And it had been applied widely in many fields which include the signal process, computer seeing, image process, speech denoise and compress.In fact the wavelet transforming is the array of filter which include one low pass filter and a serial of belt pass filters (high pass filter). By them the wavelet transform decompose the signal into detail part and approximate part (low frequency part) and in the same time the characteristic of the signal in time field is not changed.After the wavelet transforming the main information of speech whose value of wavelet's coefficients are high collect in the low frequency area while the white noise which value of wavelet's coefficients are low focus in the high frequency area. So we may process the subwaves by properly threshold algorithm which set the coefficients lower than the threshold (produced by noise) to zero and keep the coefficients greater than the threshold (produced by signal). It will restrain effectively the noise. At last the reverse wavelet transform is executed to reconstruct the signal. This is the wavelet filter.The denoised algorithm based on the wavelet transforming run into several steps as followed:1. Original signal transform by wavelet which relate to select the wavelet base. The thesis make the compare between several wavelet bases: Haar wavelet has only one vanishing moment. But it has shortest support and it is orthogonal and (antisymmetrical) symmetrical; DB wavelet has high vanishing moment and orthogonal, but it is not (antisymmetrical) symmetrical; although biorthogonal wavelet is not orthogonal, it is (antisymmetrical) symmetrical and analysis wavelet and synthesis wavelet have high vanishing moment. When the vanishing moment is fixedness DB wavelet is the shortest support in the orthogonal wavelet. So DB wavelet is the best in this meaning. And it can show the signal's detail characteristic preferably. So DB wavelet is more popular.2. Threshold estimator. This thesis compared several popular thresholdestimator: universal threshold estimator, include Donoho's universal threshold and t = cσ,c∈[3.0,4.0] threshold based on "3σrule" ofNormal Distribution; several unbiased risk estimators; Generalized cross-validation (GCV) estimator. The final conclusion is that universal threshold estimator is best and "3σrule" is not only simple but also agelity. It's performance is good also.3. Threshold function. Balance three threshold functions: hard-threshold function is simple while it lose the high frequency detail signal easy; soft-threshold function can ensure the continuity of signal which are processed. But a great deal of wavelet coefficients shrinked will result in high warp; semisoft threshold function combine their excellence above two functions but need to estimate two thresholds that increase the complexity. From above discussed the conclusion are drawed that the improved semisoft threshold function are used which only estimate one threshold but it's performance is almost as same as other's.4. Threshold selecting strategy include unite threshold and independence threshold. By experiment it is proved that independence threshold distortion is more than unite threshold.5. evaluate standard is pay more attention to subjective than objective. So more the high frequency detail of speech are reserved.The thesis by studying and carrying out denoise by wavelet transform technology deeply, considering the speech signal characteristic, objective evaluate and subjective evaluate is same important, present enhanced denoise algorithm of speech based on wavelet transform: decomposing speech signal by wavelet packet, adjust the threshold estimator into the dynamic coefficient is multiplied on original threshold estimator to improve the estimator's adaptability so that more detail of speech are reserved. And adding the procedure removing the wild dot in wavelet coefficients, enhance the wavelet denoise effect. A lot of experiment results proved that the algorithm this thesis presented is better in subjective evaluate than others in past.