The Research Of Speech Enhancement Based On Wavelet Transform

Speech signal is inevitably interfered by noise from the surrounding environment and transmission medium. Speech quality degenerates as a result of noise, and the performance of systems such as speech coding, speech synthesis and speech suffers from it at some extent. Then speech enhancement becomes the core technology in the field of speech signal processing.In this paper, speech enhancement based on wavelet transform is systematically researched. Wavelet transform based speech enhancement is carried out in the following steps: first, the noisy speech signal is transformed into wavelet domain with wavelet coefficients as output. Then a series of nonlinear mathematical treatment operate on the wavelet coefficients. Finally, the noise is removed by inverse wavelet transform, and speech enhancement is successfully realized.Wavelet algorithms on speech enhancement lie in several vital problems -- choice of wavelet basis, wavelet decomposition levels and noise variance estimation. First of all, characteristics of several traditional wavelet functions are studied and the most feasible function is selected according to wavelet choosing principal and simulation results. Optimal level of wavelet decomposition can be given by analysis on the impact of wavelet decomposition levels. Improved method of noise variance estimation is proposed based on analysis of traditional noise variance estimation, the improved method remove the impact of speech wavelet coefficients on noise variance estimation by using the relevance of speech wavelet coefficients between scales.Improved methods on set of threshold and selection of threshold function- two core issues in wavelet threshold denoising algorithm are also proposed in this paper. Benjamini.Y's distribution-free false discovery rate (FDR) hypothesis test algorithm is used to wavelet coefficients selection, which outperforms other methods with the advantage: set of threshold is flexible, independent on signal length and and robust to noise variance. A perfect threshold function is also proposed to overcome the defects of traditional Soft-threshold function and Hard-threshold function.Finally, theoretical analysis and simulation with MATLAB are both taken into operation. Experimental results show that Sym10, Sym8, Coif4 and Db5 can give better performance than other wavelet basis; Sym10 performs the best in speech signal processing. Wavelet decomposition with 5 layers enjoys better signal-to-noise ratio. Improved noise variance estimation is more precise than the traditional method, especially when signal-to-noise ratio is low; the proposed FDR-threshold give better signal-to-noise ratio performance than the traditional method in speech enhancement, the proposed threshold function is not only feasible but also can give better signal-to-noise ratio performance, moreover, the similarity between enhanced speech signal and the original speech signal is good.