PRIOR DISTRIBUTION CALCULATION APPARATUS, VOICE RECOGNITION APPARATUS, PRIOR DISTRIBUTION CALCULATION METHOD, VOICE RECOGNITION METHOD, AND PROGRAM

摘要

PROBLEM TO BE SOLVED: To provide a prior distribution calculation apparatus capable of generating prior distribution allowed to be used for both of application of model/feature spaces and capable of reducing a calculation volume.SOLUTION: The prior distribution calculation apparatus includes: a feature vector extraction part for extracting a feature vector for each speaker from input voices of a plurality of speakers; a first transformation matrix estimation part for estimating a first transformation matrix for each speaker by fMLLR by using the feature vector and an initial acoustic model previously learned from all speaker data; a feature vector transformation part for transforming a feature vector of a corresponding speaker by using the first transformation matrix of each speaker; an MLE acoustic model learning part for learning an acoustic model by a most likelihood method by using the feature vector transformed by the feature vector transformation part; a second transformation matrix estimation part for estimating a second transformation matrix for each speaker by the fMLLR by using the feature vector transformed by the feature vector extraction part and the acoustic model learned by the MLE acoustic model learning part; and a prior distribution calculation part for calculating multivariate normal distribution of a matrix by using the second transformation matrix.