| 摘要 |
PROBLEM TO BE SOLVED: To provide a voice recognition system capable of highly precisely estimating a variance of an input signal so as to improve an effect of MVN. ! SOLUTION: A mean/variance calculation unit 8 estimates a mean/variance of a feature vector on the basis of a previous feature vector 5, which is accumulated in a feature vector accumulation unit 4, and a prior distribution parameter 7 by utilizing MAP(Maximum a posteriori) estimation. A feature quantity normalization unit 12 normalizes a feature vector 3 using an estimated mean/variance vector 9 and a mean/variance vector 11 of a feature vector of an acoustic model. A voice recognition unit 18 performs voice recognition using a normalized feature vector 13, an acoustic model 15, and a language model 17. ! COPYRIGHT: (C)2015,JPO&INPIT |
