| 摘要 |
<p>PROBLEM TO BE SOLVED: To permit high performance talker recognition according to the circumstances by using an angle formed by two feature vectors expressing a speech as a criterion of talker similarity at a time of recognizing a talker. SOLUTION: At a time of recognizing a talker, a switch 18 selects templates of all talkers registered in a speech feature information storage part 16 successively as subjects of comparison, and sends them to a degree of similarity calculation part 17. The degree of similarity calculation part 17 sequentially calculates crossed axes angles between parameter values of the templates 16A and acoustic vector rows of recognition data 12A. And, the calculated crossed axes angles are sent to a talker recognition part 19. In this case, the talker recognition part 19 judges a registered talker as an actual talker, who has a larger value than a corresponding threshold and also corresponds to the maximum value of each accumulated degree of similarity to the templates of all the registered talkers, and judges the talker as unregistered if the maximum does not exceed the threshold value, and outputs this as a talker recognition result 19A.</p> |
