ACOUSTIC MODEL PERFORMANCE EVALUATION DEVICE, METHOD AND PROGRAM

摘要

PROBLEM TO BE SOLVED: To provide an acoustic model performance evaluation device in which it can be evaluated how stable an acoustic model is as to voice data for which an entire recognition target is not transcribed.SOLUTION: A learning set division part divides a learning set which is voice data with subscription to be used for acoustic model learning into M (M≥2) learning subsets; and an acoustic model successive learning part outputs m(m=an integer of 1 to M) successive learning acoustic model groups each of which is an acoustic model obtained by way of learning in which a base acoustic model and learning subsets are used while gradually increasing the number of learning subsets. Then, a stability evaluation set division part divides voice data different from the learning set into N (N≥2) stability evaluation subsets. The acoustic model performance evaluation part obtains, by using the successive learning acoustic model group and the stability evaluation set, a reliability of the successive learning acoustic model group for each stability evaluation set so as to calculate a stability indicative of variation of the reliability.