FEATURE QUANTITY CORRECTION PARAMETER ESTIMATION DEVICE, VOICE RECOGNITION SYSTEM, FEATURE QUANTITY CORRECTION PARAMETER ESTIMATION METHOD, VOICE RECOGNITION METHOD AND PROGRAM

摘要

PROBLEM TO BE SOLVED: To provide a feature quantity correction parameter estimation device and method that stably estimate a feature quantity correction parameter through discriminative learning, in consideration of a degree of inconsistency between a feature of voice data for learning and that of voice data for recognition.SOLUTION: A feature quantity correction parameter estimation device includes: a correction parameter differential value calculation unit that obtains a differential value when using a feature quantity correction parameter to differentiate an objective function of dMMI discriminative learning criterion, on the basis of a language probability of an opposing candidate symbol series to be obtained by a language model, an acoustic score to be obtained by an acoustic model on the basis of the corrected feature quantity and the opposing candidate symbol series, a first margin parameter to be adjusted according to a degree of inconsistency between a feature of voice data for learning and that of voice data for recognition, a second margin parameter and a degree of difference; and a correction parameter update unit that updates the feature quantity correction parameter by changing the feature quantity correction parameter according to the differential value.