Speaker dependent model training set is in accordance with the relatively plurality of training speakers, one model per speaker, and model parameters are extracted, as set in the following order to form a super set of vectors, one per speaker. Then unique sound space eigenvectors super set of principal component analysis on the vector to produce a set of forming a is executed. If necessary, the number of vectors may be reduced to achieve data compression. Then, the new speaker provides application data to be super By vector on the basis of the maximum likelihood evaluation suppress the super-specific vector in speech space is configured. Final coefficient in the eigenspace of this new speaker is used to construct a new set of model parameters that are adapted model for that speaker is configured. Adaptation is executed By including environmental variability in the training data.
展开▼