STATISTICAL ACOUSTIC MODEL ADAPTATION METHOD, ACOUSTIC MODEL LEARNING METHOD SUITED FOR STATISTICAL ACOUSTIC MODEL ADAPTATION, STORAGE MEDIUM STORING PARAMETERS FOR CONSTRUCTING DEEP NEURAL NETWORK, AND COMPUTER PROGRAM FOR STATISTICAL ACOUSTIC MODEL ADAPTATION

摘要

PROBLEM TO BE SOLVED: To provide a statistical acoustic model capable of achieving efficient DNN-based acoustic model adaptation using learning data in specific conditions with high accuracy.SOLUTION: An acoustic model speaker adaptation method based on DNN includes: a step of separately storing speech data 90 to 98 on different speakers in a first storage device; a step of preparing hidden layer modules 112 to 120 for each speaker; a step of performing preparatory learning of all layers 42, 44, 110, 48, 50, 52, and 54 of a DNN 80 while switchably selecting the speech data 90 to 98 and dynamically replacing the specific layer 110 by the hidden layer modules 112 to 120 corresponding to the selected speech data; a step of replacing the specific layer 110 of the DNN completed with the preparatory learning by an initial hidden layer; and a step of fixing parameters of the layers other than the initial hidden layer, and performing DNN learning with voice data on a specific speaker.