Research Of The Speech Recognition Algorithm Based On HMM

With the progress of modern computer technology, we use computer more and more. Because communicating through speech with computer is the most comfort way, the speech technology became a mark of the development of science. Speech recognition and synthesis become one of the important research field.The technology of the speech recognition contains of more different field. The development of these field has made contribution for the development of speech recognition. Nowadays, most speech recognition system are still in their infancy and have problems if migrated from lab, and are much far from practicality. The ultimate reasons for restricting practicality can be classified to two kinds, precision for recognition and complexity of the system. This paper is just researching for the theory and technology problems which practicality of Chinese speech recognition is faced with, and also validates some part of them with experiments.This paper consists of 4 parts. The first one clarifies the significance of speech recognition research, and then introduces the history and actuality of speech recognition. Finally it introduces primary contents of this paper and the results of them. The second part is a system with speech recognition. It introduces mostly contents of speech recognition system for research and both the keystone and difficulty of speech signal processing. In Chapter 3, I introduce a speech recognition system basing on HMM, and I work over mostly how to mend traditional algorithms when modeling with HMM in practice. The last part is arithmetic for speech recognition. The work is to research methods for matching speech signals after found models.The main contents for research are as follows:1. Research for the construction of speech recognition system and the primary technology.2. Analyze the technology keystone and difficulty when HMM is applied to speech recognition system.3. Put forward the method for endpoint detection with FER4. Improve on optimized method of parameter B in the process of HMM training.5. Analyze the shortage of the training data and influence from speaker to the model in practice, and then Put forward a method for it.6. Improve on optimized recognition algorithms and cut off low-belief embranchments as to get higher recognition probability and shorter matching time.