PROBLEM TO BE SOLVED: To realize adaptation between a speaker and environment without requiring many pieces of training data about a new speaker.;SOLUTION: The speaker is trained by one model per speaker, a model parameter is extracted in a prescribed order and one set of super vectors per speaker is constituted. Main component analysis is performed regarding the set of super vectors, a set of unique vectors which define a unique voice space is generated and the number of vectors can be reduced for data compression if it is desired. The new speaker provides adaptation data by which the super vectors are constituted by restricting the super vectors so as to exist in the unique voice space based on the maximum likelihood evaluation. A coefficient of a result in the unique space of the new speaker is used for constituting a new set of a model parameter by which the adaptation model is constituted for the speaker. Environmental adaptation is achieved by including environmental change in the training data.;COPYRIGHT: (C)2004,JPO
展开▼