PROBLEM TO BE SOLVED: To provide an acoustic-analysis-frame-reliability calculation device for calculating a reliability measure for each acoustic analysis frame that is a minimum unit for handling acoustic data, and to provide an acoustic-model adaptation device and a voice recognition device each including an acoustic-analysis-frame-reliability calculation device.;SOLUTION: A voice recognition unit extracts an acoustic feature amount from each acoustic analysis frame, calculates acoustic likelihood using the acoustic feature amount and an initial acoustic model, assigns a maximum-likelihood phoneme ID to the acoustic analysis frame, and outputs an acoustic feature amount sequence with phoneme IDs and acoustic likelihood in which the acoustic feature amount is made to correspond to the acoustic likelihood and the phoneme ID. A frame-reliability-measure calculation unit classifies the acoustic feature amounts for each phoneme ID to create an acoustic feature amount set, determines an outlier score by performing an outlier detection on each acoustic feature amount set, and outputs an acoustic feature amount sequence with frame reliability in which a value of a function using the outlier score is assigned to the acoustic feature amount for each acoustic analysis frame as the reliability of the acoustic feature amount.;COPYRIGHT: (C)2015,JPO&INPIT
展开▼
