Study Of Mandarin Digit Speech Recognition Algorithm Based On HMM Model

Speech recognition is a crossed subject covered a wide range of subjects. It closely related with phonetics, linguistics, mathematic statistics and neurophysiology and is one of the most quickly developed fields in information research. The task of mandarin digit recognition is to recognize ten numbers from zero to nine toward unspecified person mandarin digit recognition. With an eye to the main problems of mandarin digit recognition, this text studies the key technology of mandarin speech words recognition in order to increase the speech discrimination and convergence rate of the model of cognition.Firstly, this text analysis the development actuality of speech recognition technology at present. Based on this, the text introduces the basic theory of speech recognition including the mathematical model for speech signal producing and feature analysis to mandarin speech. Secondly, improved end-point detection method based on energy and short-time zero crossing rates to speech signal is presented and analyzed with experiment result. In addition, the selection of feature parameters has a great influence to all speech recognition system's real-time feature and robustness. The text elaborates feature parameters extraction method of Linear Predictive Cepstrum Coefficients (LPCC) and Mel-Frequency Cepstrum Coefficients (MFCC), and compares the two parameters associoated with experiments. Experiments show that feature parameters based on MFCC have better recognition rate than LPCC parameters.In mandarin word recognition, Dynamic Time Warping (DTW) theory and Hidden Markov Model (HMM) theory are the commonly used methods. Based on the traditional analysis to DTW, a high-efficient arithmetic of DTW which can reduce calculating burden and storage space is presented. Next the text analyzes three basic problems of HMM arithmetic and solves underflow in practical Viterbi and Baum-Welch arithmetic settled by logarithm and scaling. In the end, the text realizes the mandarin digit word recognition system by MATLAB programming based on HMM, then compares the recognition rate under two kinds of parameters and compares with experiment result of isolated word recognition based on DTW. The text separately indicates superiorities and deficiencies of them and presented the improvement orientation of the subject research in the future.