Study On Speech Enhancement Algorithm Based On Wavelet Transform

In recent years, with the rapid development of computer and digital signal processingtechnology, voice processing technology has matured and has been widely used in real lifethrough the mobile phone and network video, video conferencing and other forms. However,various interference noise will affect the effect of the use of voice communications equipment,even not use normally, hinder the normal work and life of people. Therefore, it is of practicalsignificance solving the noise interference effectively, and enhancing the quality of the voicesignal. For decades, more and more domestic and foreign experts and scholars come into thisstudy them, and there have been a lot of effective voice enhancement algorithms. On the basisof a study of previous work, this paper does deep research and discussion on speechenhancement algorithm in wavelet.Firstly, the paper discusses the background and significance of domestic and foreignresearch status on speech enhancement, history of the development of the wavelet analysistheory, and three categories of wavelet transform in speech signal processing, includingde-noising algorithm based on modulus maxima, de-noising algorithm based on the spatialcorrelation and de-noising based on wavelet thresholding algorithm.Secondly, voice model and noise classification is briefly described some classic speechenhancement algorithm is studied. Specifically, traditional spectral subtraction and itsimproved algorithm, Wiener filtering, Kalman filter method and the minimum mean squareerror estimation algorithm are discussed. Several algorithms are implemented and theiradvantages, disadvantages and the scope are summarized, and then several speechenhancement algorithm evaluation criteria are introduced, including subjective evaluation andobjective evaluation.Then, based on mastering the theory of wavelet de-noising, four key factors of waveletthreshold de-noising algorithm are focused on, and a new threshold function is proposedcombined with dead negative region theory, simulation comparison is made. Experimentalresults show that, the threshold function improved in this paper has the advantage of maintainthe edge characteristics and continuity of smoothness and the de-noising effect is obvious.Finally, the advantages and disadvantages of the minimum statistical noise spectralestimation algorithm in noise estimation algorithm is studied, a new method which canadaptively track short-time power spectrum of the noisy speech minimum for the lack of fixedparameter smoothing algorithm which is easy to produce a larger deviation is proposed with spectral smoothing and speech presence detection, and then it is applied to the waveletthreshold noise reduction algorithm, comparative experiments are did between it andtraditional noise reduction algorithm. The results show that the algorithm can effectivelyremove the white noise of the voice signal, signal-to-noise ratio in the Segment-SNR andLog-Spectral-Distortion was significantly better than the traditional speech enhancementalgorithm. Combined with improved threshold function proposed in the previous chapter, ithas better noise reduction effect.