| The speech recognition is a science that investigates how to make the computer understand the human language as a most convenient, natural and quick exchanges method, and it is also the most important step in the man-machine dialogue. As an interdisciplinary field, it is also theoretically very valued. On the basis of combining theory with practice, the speech recognition based on Hidden Markov Model (HMM) is researched systematically in this paper. The details are studied as follows:First, according the basic structure of the speech recognition systems based on HMM, the principles of the main module, the speech signal collecting, preprocessing, the Mel-Frequency Cepstrum Coefficient (MFCC) feature extraction, HMM-based training and recognition, are presented in detail. Some of the problems existing in HMM model, such as Model initialization and data overflow, are analyzed, and the corresponding measures, which can improve system performance, are introduced.Second, a non-specific and isolate speech recognition system, on the basis of the theory method of the speech recognition, is designed and developed within the Visual C + + 7.0 platform. Auxiliary tools, such as development and audio processing tools, is introduced. The basic functions of identification system are showed according to the visualized interface. And the key modules of C++ algorithm are presented.Third, in the platform of the speech recognition system, the experiments of training and identification based on HMM are carried out on the four words, such as one, two, three, and four and so on. In the training stage, the average training time of each model with 16 samples is about 6.206 second, meeting the real-time requirements. In the identification stage, during the total of 32 times, four times are identified wrongly. With the recognition accuracy of 84.38% and the identification time of each word of 0.313 second, the presented method in this paper is tested the effectiveness which has good application value. |
PDF Full Text Request