| 摘要 |
PROBLEM TO BE SOLVED: To provide a sound model adaptation device which generates a sound model exhibiting a high adaptation effect in view of a phoneme error tendency of adaptive data, and a voice recognition device.SOLUTION: A phoneme error tendency vector generating part receives a plurality of voices as input, generates a phoneme error tendency vector for each of the voices, and outputs, as a set, the voice and the phoneme error tendency vector. A clustering part, while classifying the voices and the phoneme error tendency vectors into a predetermined number of classes of clusters in accordance with a degree of similarity between the phoneme error tendency vectors, obtains a centroid, which is an average vector of phoneme error tendency vectors in each cluster, and outputs the cluster and the centroid as a pair. A base acoustic model adaptation part causes a base acoustic model to be subjected to adaptation without a teacher for each cluster so as to generate an acoustic model after adaptation, and records it in an after-adaptation acoustic model recording part. Further, a voice recognition device performs a voice recognition process by using an application acoustic model which is selected in accordance with the degree of similarity between phoneme error tendency vectors. |
