Collation device, collation method, and collation program

摘要

The accuracy of speaker verification is improved for short time utterances.The collation device 10 comprisesAudio signalA first NN 142 which converts the characteristic of each frame constituting the voice signal andA second nn143 for recognizing the speaker of the frame based on the feature of the transformed frameA speaker recognition model comprising a third nn144 that identifies the phonemes of the frame based on the feature of the transformed frameThe output of the second nn143 approaches correct answer dataThe output of the third NNN 144 is learned so as not to approach the correct answer data.The collation section 146 performs speaker collation using the feature quantity output from the first NN 142 or the feature output from the intermediate layer of the second nn143 in the speaker recognition model after learning.Diagram