| With the rapid development of artificial intelligence, the technology of speech recognition is becoming the focus topic of research institutions around the world. People commit to make a machine that can understand human speech instruction, and also hope that we can control the machine through voice. Research and development of speech recognition will greatly convenient people’s life in the near future.The processes of speech recognition include roughly: feature extraction, acoustic model training, linguistic model training and recognition algorithm. The whole process in this speech recognition is researched in this thesis, including feature extraction, model training and recognition algorithm. Based on the characteristics of Chinese phonetic pronunciation and computation problems such as comprehensive consideration, the Chinese voice initials and finals are selected as the basic acoustic unit. Besides this, semi-continuous HMM model is used as the acoustic model of initials and finals. Mel cepstrum coefficient is selected as the characteristic parameters in this thesis, and this paper deeply analyzes the MFCC parameters extraction process, and proposed an improved MFCC extraction algorithm that reduced calculation by nearly 50% than the traditional MFCC extraction algorithm, which greatly improved the efficiency of feature extraction.Sphinx system is a mature international system in speech recognition area. In this thesis, the acoustic model training tool sphinxtrain internal implementation and mainly relates to conduct the thorough research to the algorithm, and by adjusting the parameters to adapt the demand of Chinese speech recognition. In the end, based on CMU Sphinx this thesis builds and implements the mandarin continuous speech recognition system, and achieves pretty good effect. And the recognition rate of the numeric string comes to 98%, besides this, compare the recognition effect with different code book number, state number and feather extraction algorithm. |
PDF Full Text Request