Research Of Speech Recognition Based On CDHMM/SOFM Neural Network

Speech recognition, as a comprehensive subject, is involved in many fields. Its applications refer to industry, military, medicine, civil, especially computer, information processing, communication and electric power system, auto-control, etc. These years, speech recognition technology has already been a hot topic in research field. In order to improve the discrimination and robustness of speech recognition system, this thesis does some deeper research.Firstly, the characteristics of speech signal in time and frequency domain are analyzed; pre-weighting, en-framing, window adding in preprocessing stage and endpoints detecting are researched; extraction methods of LPCC and MFCC, which reflect spectrum characteristics of speech signal, are systematically analyzed and researched.And then, two methods of speech recognition are researched: Hidden Markov Mode1 (HMM) and Artificial Neutral Net (ANN). HMM has a strong ability of sequential model building, and it is a kind of dynamic information processing method based on sequential accumulative probability. But it needs prior knowledge of speech signal and its classification and decision making capability is weak. ANN is widely used in speech recognition field due to its self adapting, parallelism, non-linearity, robustness and learning ability.At last, HMM and ANN is combined in this thesis, making full use of the strong abilities of sequential model building of HMM and classification and decision making of ANN. Bring foreword a new model of speech recognition—CDHMM/SOFMNN, combined by Continuous Density HMM (CDHMM) and Self-organizing Feature Mapping Neural Network (SOFMNN). Speech recognition and simulations based on this new model are given. It is provedthat: compared with conventional CDHMM, this new model can increase discrimination and robustness apparently.