Study Of Speech Enhancement Algorithms

Speech enhancement is one of the most important parts of the speech signal processing. It plays an important role in speech recognition and speech coding. The purpose of the speech enhancement is to take the pure originality speech from speech with noises and rise the SNR, Then improve speech quality. This thesis introduces the different characteristics between speech, hearing and noise, discuss various speech enhancement methods, and do some research on the traditional spectral subtraction method.In this paper, an adaptive filtering in speech enhancement and the traditional spectral subtraction method are discussed. According to the theory of wavelet transform and the idea of spectral subtraction, a method on wavelet domain is presented in this thesis.Firstly, this paper concentrates on the basic principle and methods of speech enhancement, and discusses the theories, conditions and problems about several methods of speech enhancement.Secondly, this paper introduces the adaptive filter in speech enhancement, and combine RLS algorithm with IIR filter, then the better method of single channel adaptive filter in speech enhancement is introduced. Through simulation this paper identifies the speech enhancement algorithm.Thirdly, the traditional spectral subtraction method is validated, found that common spectral subtraction method can filter a little noise, but has "music noise". On the other hand the enhanced method gets better SNR than traditional spectral subtraction method after it adjusts the parametersαandβ. The "music noise" is reduced.Fourthly, after decompounding the speech signal with noise through wavelet transform. The proportions are different between signal and noise on different measures. So through setting different threshold, noises are reduced. The speech after reducing noise is fluent. Before calculating threshold, the improved wavelet transform distinguishes surd and sonant depend on characteristics of speech signal. This method not only improved speech signal's SNR availably but also held surd's important information perfectly. The simulation results indicate the method is effective. The algorithm is simple and easy to realize.