Research On Mandarin Digit Speech Recognition Technology And Implement Approach

Mandarin digital speech recognition technology and implement approach is studied in this thesis.On the basis of the digital model of speech generation, the speech signal is analyzed in the time domain, frequency domain and the cepstrum domain. Then the fundamental theory and the main technology of speech recognition is expounded in the thesis. A computationally efficient algorithm on suppressin of acoustic noise in speech is adopted, which is RASTA-MFCCs. After comparing the theory of pattern recognition as DTW,HMM and ANN, Hidden Markov Model (HMM) is chosed in the thesis. We study HMM model applied in digital speech recognition and speech digital signal processing about speech recognition, and discuss its training and recognizing algorithm. As to the limitation of the classical HMM, then we discuss the Duration Distribution Based Hidden Markov Model (DDBHMM),which has taken duration distribution into account.The modified Viterbi algorithm is discussed, which is fit DDBHMM. And we also discussed the training and recognizing algorithm under the DDBHMM. Then, we simulat training and recognizing algorithm based on matlab6.5, and get the simulation main results. We also discuss programme design of digital speech recognition system, which is based on C++. We draw the system structure charm., design the class library programming and realize the interface between people and computer is discussed We disscuss how to build a digital speech database for training and testing digit model with the special recording tool made by our research group. We present the outcome of the experiment and analyze it, which proved that recognition effect of the DDBHMM is better than that of the classical HMM. That is to say, taking state duration distribution can improve the rate of correctness of recognition. But the actual time of the DDBHMM is worse than that of the classical HMM. At last, we point out the direction of the research improvement...