摘要 |
PROBLEM TO BE SOLVED: To prepare a normalized likelihood ratio model by voice unit, and to calculate the reliability of normalized voice recognition using the likelihood ratio model.SOLUTION: This device for preparing a likelihood ratio model calculates acoustic feature quantity of voice data for each frame using an acoustic model 280, development data 200 including voice data and a correct answer label, and a GMM 210. The device calculates correct answer likelihood of the voice unit included in the correct answer label using the correct answer label and the acoustic model 280 for the acoustic feature quantity. The device calculates GMM likelihood by the GMM 210 for the acoustic feature quantity, and calculates the likelihood ratio by voice unit between the correct answer likelihood and GMM likelihood for each frame. The device prepares a normalized probability distribution function (likelihood ratio model by voice unit) obtained by normalizing the probability distribution function using, as the stochastic variable, likelihood ratio by voice unit corresponding to the voice unit for each voice unit included in the development data 200. The reliability of a voice recognition result is calculated using the likelihood ratio model by voice unit. |