Selection and analysis of HMM´s state-number in speech recognition

It is known that the process of pronunciation is a stochastic process. From the viewpoint of information theory, it is an information source, which generates stochastic vectors of speech. While in recognition, the hidden Markov model (HMM) is another generator of the stochastic sequence. But for different states of HMM, the distortion of above two information sources, pronunciation itself and HMM, is different. This paper establishes a simplified model to study the principle of selection of the number of the state in HMM. Finally, three conclusions on HMM information entropy are drawn. It is found that when the states of HMM amount to 6, the information entropy of HMM is very close to that of pronunciation itself. Therefore, the result is obtained that the number of the state in HMM of about 6 is the best selection