| With the development of information technology, speech recognition technology has been used more and more widely in all fields of our life. Because of the difference between training environment and identify environment, and the complexity of real environment, all the existing voice recognition systems will have a much worse performance when the speech data has some noise. It is also one of the greatest difficulty in the large-scale promotion of speech recognition. Therefore, the research of de-noising technology in speech recognition has important theoretical and practical significance.Most of the traditional speech de-noising methods are based on the Fourier transform, for the time-frequency unitary of Fourier transform, these methods can not get a well effect especially when used for the non-stationary and complex signals de-noising. Wavelet transform is a new discipline developed progressively from the late '80s of 20th century, it has a good time-frequency localization property which can give detailed analysis and processing on the time-frequency space, so the wavelet de-noising in non-stationary signal processing can obtain a better result. This paper firstly introduces the basic theory of wavelet analysis, gives some research on the basic principles and common methods of wavelet de-noising, especially on several key steps like the choice of base wavelet and threshold function. Then considering the special requirements of speech recognition de-noising system and the traditional hard and soft threshold functions and some other modified threshold de-noising methods, an improved threshold de-noising function is proposed in this paper. The following experiments show that the improved function has a better performance in speech de-noising and leaves more original speech information, we also give an application example of this improved de-noising algorithm. At last, Summarizes the algorithm is summarized and the future research plan is given out. |
PDF Full Text Request