Study Of Threshold De-noising Algorithm Of Speech Signal Based On Wavelet Transformation

In practical application, such as speech telecommunication, speech synthesis, speaker recognition, the degradation of speech by all kinds of noise is inevitable and it affects the performance of the speech equipments very badly. So, it is necessary to denoise for the degraded speech. De-noising of speech signal has become an important branch of digital signal processing, its main function is reducing noise of noised speech signal, and enhancing voice definition.There is various de-noising algorithm of speech now, among them, wavelet has good localizing quality at time domain and frequency domain simultaneously and the characteristic of multi-resolution ratio analysis, it implements multiscale analysis to the signal by the translation and dilation of the mother wavelet, it can effectively extract information from signal. So it can fulfill all kinds of wave-filtering needs such as low-pass, high-pass, sink wave, random noise denoising. Compare with traditional wave-filtering methods, wavelet has incomparable advantage. Wavelet has become an effective means of signal analysis.There are three main methods of wavelet de-noising: modulus maxima de-noising method, correlation de-noising method and thresholding de-noising method. Among them wavelet thresholding de-noising gets extensive application because of its simple calculation.This thesis firstly introduces the characteristics of speech and noise. It discusses various of speech de-noising methods. Correlation theory of wavelet transformation application to speech de-noising is analyzed. And speech de-noising method based on wavelet thresholding is researched. Then it focuses on reasonable selection of some important parameter of wavelet thresholding de-noising, that are mother wavelet, wavelet decomposition level, thresholding and thresholding function.Thresholding function has relation to the continuity and precision of reconstructed signal, and it has great influence to wavelet de-noising. Donoho's hard thresholding and soft thresholding algorithm is frequently used to de-noising and has obtained a good effect now. Discontinuity of hard thresholding function results signal after de-noising has obviously noise. Soft thresholding function has good continuilty but it has a constant deviation of the estimated value wavelet coefficient from the noised signal wavelet coefficient. It will be too smooth when the noise is very irregular. A new thresholding function is presented based on the shortage of soft and hard thresholding function. It not only overcomes the discontinuity of hard thresholding function but also solves the constant deviation of soft thresholding function.At last, by experiment simulation, it is demonstrated that the improved thresholding function can remove the white noise in which the speech signal contain so effectively . And its signal-to-noise ratio is superior to the traditional thresholding function. It can obtain better de-noising effect. So the thresholding function in this thesis has a lot of superiority and efficiency.