Chinese Continuous Digital Speech Recognition In Noisy Environment

At present Speech Recognition System can achieve very high precision to pure speech. However, Environment noise would bring great effects to speech recognition system and lead to recognition rate decrease. So anti-noise problem is a key problem that whether speech recognition could achieve practical.First, traditional method of speech endpoint detection is improved. The paper combines short-term energy with zero-crossing rate and set limitation and judgement method according to experimental results. And the paper proposals new endpoint detection algorithm combining spectral subtraction method in strong noisy environment.Second, introduce common denoising methods: binomial weighted, Wiener filter, spectral subtraction method. To spectral subtraction method's three issues, the paper combines new endpoint detection algorithm with MFCC and proposals the second denoising algorithm. By comparison speech recognition rate of several methods, find that the new algorithm has an obvious effects.At last, based on the result of the above analysis, in the matlab7.0 experimental environment, compare recognition rates of the isolated number of four denoising methods. On this basis, building a speech recognition system of non-fixed-length Chinese digital string based on HTK. And then By using the four method to denoise and comparing recognition rate, prove that new denoising algorithm is valuable.