METHOD, DEVICE, AND PROGRAM FOR GENERATING ACOUSTIC MODEL AND VOICE RECOGNITION DEVICE

摘要

PROBLEM TO BE SOLVED: To improve recognition performance by optimizing distributed numbers in each state constituting an HMM (for example, a syllable HMM). SOLUTION: An acoustic model generation device has; a distributed number setting means 5 which successively increases a distributed number in each state of the HMM; an alignment data generation part 7 which makes respective states set to a certain distributed number correspond to voice data for learning to generate alignment data; a description length calculation part 8 which uses a description length minimum reference to obtain a description length with respect to each state of the HMM having alignment data and a present distributed number and obtains a description length with respect to each state of the HMM having the just preceding distributed number: and an optimum distributed number determination means 9 which sets an optimum distributed number for each state on the basis of description lengths obtained with respect to respective states of the HMM having the present distributed number and sizes of description lengths obtained with respect to respective states of the HMM having the just preceding distributed number. COPYRIGHT: (C)2005,JPO&NCIPI