摘要 |
PROBLEM TO BE SOLVED: To create a model considering peculiarity of each learning speaker in speech recognition. SOLUTION: A reducde dimensionality eigenvoice analysis method is used for structuring a context-dependent acoustic model concerning allophones. This eigenvoice method is also used during execution of analysis about speech of a new speaker. By this method, characteristics peculiar to the speaker are excluded, and more general and firmer allophonic model is created. In one embodiment, the eigenvoice method is utilized for specifying a centroid of each speaker, and the centroid is subtracted from recognition error. In another embodiment, a maximum likelihood method is used, and a decision tree structure commonly usable to all speakers is composed in the case of composing the eigenvoice expression of speaker space. |