Speech Enhancement Based On Wavelet Transform And Karhunen-Loeve Transform

Speech signal is the most convenient and shortcut way of intercommunion. Presently, speech communication gains more and more widely application in variety of fields, and the speech signal can be interfered by the noise at varying degrees, meanwhile the research of speech enhancement technology plays very important role in improving the quality of speech communication.Through the in-depth research of speech enhencement,, we recognize that: because noise signal is random, and it is impossible to get pure signal from the noisy signal. So the target of speech enhancement is to acquire the original signal as far as possible from the noisy signal, and suppress the background noise. Noise can be classified period noise, impulse noise and wideband noise (also called white noise). The period noise can be filtered though the adaptive filters, the impulse noise can be filtered though none linearity filtered. It is very difficult to filter the white noise, because the white noise has very broad frequency spectrum which almost overlay the frequency spectrum of speech signal, furthermore, there are no rules, especially, the surd of speech almost the same of noise. At present, methods being adopted are spectral subtraction algorithm, wiener filter ect.. In recent years development of the wavelet analysis to eliminate such noise has a very good effect.Firstly this thesis introduces the different characteristics between speech, hearing and noise, discuss various of speech enhancement methods. Recommend spectral subtraction algorithm and its modified, wiener filtering , MMSE and based on the model of speech production speech enhancement method. There is the"noise music"after the spectral subtraction denoising, then, in the papers , speech enhancement based on wavelet and KLT (Karhunen-Loeve Transform) has been approved. The paper focuses on some characteristics and algorithm of the Dyadic wavelet transform to achieve the threshold selection, as well as its voice in the use of noise cancellation. The main idea of wavelet threshold lies in that when noise signal transforms from time domain to wavelet domain, the signal's wavelet coefficients will concentrate in a few area of wavelet domain, while the wavelet coefficients of noise will spread to all area of wavelet domain. Although the energy of noise is bigger than that of the signal, its wavelet coefficients are smaller than that of signal. So we can use threshold function to cut off the coefficients of noise and use the rest of coefficients to reconstruct the denoising signal. There are no requirements of the matrix inversion, but of the processing of decision about the noisy voiced or unvoiced, and the non-stationary wideband noise is easily controlled. By comparison in term of spectral distortion, the proposed approach is very useful at speech enhancement in the situation of non-stationary wideband noise. Using wavelet denoising a signal can be well preserved useful signal peak and mutation of the threshold value in the application of hardware for speech enhanced, non-voiced contains many of the high-frequency noise similar to that used wavelet decomposition elimination of such elements will inevitably serious damage to reconstruction, a wavelet noise reduction, often fail to achieve the best speech enhancement effect, it is necessary to smooth the wavelet denoising further on the basis of dealing with the noise down. So here based on KLT speech enhancement is a good improvement.This paper presents a simple model of KLT, by the wavelet transform and a cascade KLT denoising. Paper also discusses the basic principles of the KLT transform, and gives a simplification. The method in accordance with the characteristics of each vector signal adhesive depended on the signal energy will transform eigenvector KLT sort. In the main feature vectors, we can provide an accurate estimate the capacity of speech signal, it will not happen for the negative valuation of the phenomenon. When the estimates of other feature vectors are nagtive, the errors would not cause any major voice signal distortion.In this paper, the speech enhancement algorithms based on wavelet transform and the simple KLT, and based on the adoption of wavelet space and the vector space decomposition KLT noisy voice signal, without the white noise and the size of pre-calculated SNR achieved speech enhancement, and KLT save in a non-orthogonal apace, reducing the amount of computation. Wavelet denoising methods and KLT denoising combination can be well preserved the peak and the mutation of the useful signals, and enhance the ability of high-frequency noise denoising. In different voice enhancement system through experiments SNR subjective evaluation and analysis show that the algorithm method overcomes the inherent defects in the past, it plays a good role on the voice enhancement and noise suppression plastic in the broadband noise. Through MATLAB simulation, as well as enhanced voice subjective evaluation method of comprehensive and objective evaluation showed that the effectiveness of the method.