Noise is the speech recognition technology to the biggest obstacle to widespread practical use .In the actual process of speech recognition applications, access and transmission of the original speech inevitably be transferred from the surrounding environment and the media noise. The jamming will decline in voice quality, resulting in the performance of speech recognition system significantly. Therefore, the speech signal is noise de-noised thereby reducing the noise of the noisy speech signal, enhanced voice clarity, a wide range of speech recognition technology into practical use the key.After years of continuous research, people have a variety of speech de-noising algorithms, such as spectral subtraction and Wiener filtering method. Wavelet transform is the rapid development of the late '80s a new type of mathematical analysis tool, it has the characteristics of multi-resolution analysis in time domain, frequency domain also has good localization properties. By stretching, translation and other computing functions for multi-scale refinement of signal analysis, wavelet analysis can efficiently extract from the signal useful information. Because the wavelet transform in the analysis of the advantages of non-stationary signals, in recent years it has been widely used in the field of speech de-noising. Wavelet de-noising method can be divided into three main categories: modulus maxima de-noising method, the correlation method and threshold de-noising method. Which based on wavelet threshold de-noising method, the calculation is simple and clear and de-noising have been widely used.In this paper, a speech enhancement technology into the discussion, describes and compares various existing speech enhancement methods, the main analysis of the wavelet transform is applied to the relevant speech de-noising theory, wavelet threshold based speech enhancement method. This article focuses on methods of wavelet threshold on wavelet bases, wavelet decomposition levels, threshold and threshold function, reasonable selection of four key parameters of the problem, conducted a systematic study. Threshold function reflects the different treatment strategies wavelet coefficients, the effect of wavelet de-noising has a great influence. At present, widely used in practice is the hard and soft threshold function threshold function. Both methods have inherent shortcomings, the hard threshold function is discontinuous, which results in the de-noised signal pseudo-Gibbs phenomenon occurs, the signal still contains significant noise; using soft threshold method, although the continuity is good, but the estimated wavelet coefficients and wavelet coefficients of noisy signals between the constant bias. In this paper, wavelet threshold de-noising method of soft and hard threshold function, this paper proposes a new threshold function, not only to overcome the hard threshold function is not continuous disadvantage, and solve the soft threshold function exists a constant bias problems.The simulation shows that using this improved speech de-noising threshold function, can effectively remove the speech signal with white noise, the SNR is better than traditional indicators threshold function method can obtain better de-noising effect, thus proving that the proposed threshold function of the superiority and effectiveness.Finally, based on wavelet threshold method to the bath a large number of research, design and implementation of a better noise immunity, the voice command control system. From the speech signal pre-processing, Wavelet speech enhancement, endpoint detection, speech feature extraction, HMM training and recognition, and so a detailed analysis of several aspects of the speech recognition system design process. |