| 摘要 |
PURPOSE: A method for recognizing a voice using structure decision of an HM-net(hidden markov network) model by combination of state division and morpheme decision tree is provided to gradually improve a recognition rate according to a change in state numbers after executing state division. CONSTITUTION: A triphone context independent HMM(hidden markov model) is learned. Parameters of the context independent HMM coincided in center phoneme are set up as initial values of a context dependent HMM is learned. A large context dependent HMM is learned. A structure of an initial model of the HM-net is defined and state distribution parameters of the large context dependent HMM are obtained through a maximum likelihood estimation(MLE). The mixed distribution number of each state is increased up to a random number, and learning voice data is used to execute relearning. |
