Chinese whispered speech recognition is widely applied on national security, judicial departments, financial sector, medicine system and our daily communication. Now the research is still in its early stage both at home and abroad. Because of the specificity in acoustic characteristics of the whispered speech, many issues in its recognition need to be solved.This paper uses digital speech as research object. The acoustic characters of Chinese whispered speech are analyzed deeply. A baseline system of speech recognition using Hidden Markov Model is built. In this system, the feature distance and the recognition performance are both compared between normal and whispered speech. According to the results of comparison, the feature parameters based on MFCC are improved and it increases the recognition rate by 6.3 percent. The conclusion is researched by experimental comparison that the most suitable number of HMM states and mixtures are 4 and 3. Finally, a classified and multistage system is built for Chinese whispered speech recognition to improve the identification of several numbers. Through the research and improvements above, the recognition rates of Chinese digital whispered speech 0-9 could reach 97.9%(in the training set)and 86.4%(out of the training set). |