SPECIFYING METHOD FOR CHARACTERISTIC SPACE REPRESENTING MULTIPLE LEARNING SPEAKERS

摘要

PROBLEM TO BE SOLVED: To provide a method for guaranteeing that all parameters of supervectors of individual speakers have similar arrangements in voice recognition. SOLUTION: This is a method for representing multiple learning speakers; and unspecified speaker models SI are generated as to the respective learning speakers by using learning voice data of the individual learning speakers and specific speaker models SD of one group of models are described with multiple model parameters. Then the model parameters of the models of the individual learning speakers are linked with related supervectors respectively to display combined models by the speakers in a high-dimensional space. Then the dimensions of the model space is decreased and conversion is performed so as to restore a bas vector Ee in the characteristic space.