摘要 |
PROBLEM TO BE SOLVED: To quickly achieve the practical level without requiring a large quantity of training data on a new speaker by generating adaptation model by use of voice input from a new speaker to train an adaptation model while using a set of main component vectors. SOLUTION: A training speaker 20 of a T person gives document for training data so that a native space is constructed, and the training data is used for training a speaker dependence(SD) model 24. After a super vector is constructed for each training speaker, main component analysis is executed (32). The main component analysis on T-number of super vectors gives T-number of characteristic vectos (34). N-number of vectors are extracted arbitrarily from the T-number of characteristic vectors (36) to construct a native space for parameters decreased in number of dimensions (38). When the native space is once constructed, speaker adaptation or environment adaptation can be easily achieved. In this case, adaptation is comparatively simple computing operation, and it can be executed even during the use of the system by a new speaker. |