| 摘要 |
PROBLEM TO BE SOLVED: To generate an acoustic model which raises the rate of voice recognition using a generated acoustic model in comparison with a conventional example. SOLUTION: A prescribed learning algorithm is used to generate an HMM of a single Gaussian distribution on the basis of the speaking voice data of a plurality of specific speakers, and one state is divided in a context direction, a time direction, or a pass direction to retrieve a state which has the maximum increase of a likelihood expected value. After the state which has the maximum increase of the likelihood expected value is divided in the context direction, the time direction, or the pass direction corresponding to the maximum increase of the likelihood expected value, the prescribed learning algorithm is used to generate an HMM of a single Gaussian distribution, and this processing is repeated until the state in the HMM of the single Gaussian distribution cannot be divided any more ore the division number reaches a preliminarily determined value, thereby generating an HMM. The generated HMM is used to perform voice recognition.
|
