A Research On Wavelet-based Speech Signal Denoising Algorithms And Their Applications

Speech signal denoising is a fundamental problem which shall be resolved when speech recognition system being deployed in practical application. The research on the algorithms of speech signal denoising is certainly beneficial to this problem. In the wavelet-based speech signal denoising algorithm, speech signal shall firstly be transformed to the domain of wavelet, then by the mechanism of the scale difference of wavelet coefficients for pure speech signal and noise signal, the wavelet coefficients of speech signal with noises are processed. The essence of wavelet-based denoising algorithm is to perfectly complete the discrimination of source speech signal and noise signals by searching the optimal approximation of source signal according to provided measurement criteria for the function space formed by scaling and shifting of mother wavelet function.The correctness level of speech processing is heavily depended on the accuracy and quality of speech signal features which being extracted, and de-noising processing is very beneficial for improving the robustness of speech recognition system. Traditional de-noising methods have the drawback of trade-off between keeping local characteristics of speech signals and reducing noise. With more twenty years after its advent, the more and more mature wavelet techniques have been shown obvious advantages when being compared with traditional de-noising methods. It is may be due to wavelet is a good model for finer tempo-spectral characteristics, like a "mathematical microscope", meaningful signal can be extracted effectively from the mixed signal.This paper mainly studies the theory and methods of wavelet de-noising, analyzes the characteristics of wavelet transform, proposes a new threshold de-noising method and also the novel de-noising procedure for white noise and shot noise. This paper focuses on the analysis of the application in signal de-noising, compares several de-noising methods proposed in this paper with traditional methods, studies the threshold de-noising method, and put forward a new algorithm in allusion to two difficult points encountered in value de-noising method:the selection of wavelet basis and decomposition level as well as the reasonable selection of threshold. At the end, this paper uses the example of white noise and wavelet analysis toolbox of Matlab to propose an improved wavelet threshold de-noising and new wavelet de-noising method depending on the noise mode of the signal. The experimental results show that this method is applicable to additive white noise signal, simple and feasible, achieves the expected effect of noise reduction, and signal to noise ratio has been improved, the RSME has got a good decline. Consistent theoretical analysis and experimental results show that this method has a certain value for practical application.The first chapter introduces speech recognition, wavelet transform and signal de-noising.The second chapter introduces the different acoustic models for speech noise, characteristics of the noise, evaluation criteria of the noise and the Fourier transform.The third chapter introduces the basic principles, methods and properties of wavelet transform, commonly used wavelet functions and the concept of multi-resolution analysis.The fourth chapter introduces a variety of methods, principles and thoughts of wavelet threshold de-noisingThe last chapter, the fifth chapter, upon recording and establishing a certain number of speech signals and noise signals (white noise, shot noise), describes the simulation experiments in Matlab software environment for noising speech signals. Finally, analyzes and discusses the performance of two proposed algorithms for de-noising basing upon these simulation experiments.