摘要 |
<p>PROBLEM TO BE SOLVED: To improve verification performance by distinguishing a phoneme apt to express individuality from the phoneme difficult to express it. SOLUTION: This speaker verification device comprises a 1st feature extraction part 102 for extracting feature data of utterance of a standard speaker, a 2nd feature extraction part 105 for extracting the feature data of the utterances of many speakers, a matching part 107 for taking correspondences in the feature data between the standard speaker and many speakers, a feature data statistics analysis part 110 for calculating a covariance matrix from the plural feature data of many speakers corresponding to the feature data of each analysis frame of the standard speaker, and determine the covariance matrix between the speakers, and a mean/dispersion associating part 115 for registering a men value of the standard speaker feature data nearest to each time series data of the registered speaker feature data from a 3rd feature extraction part 113, and for registering the covariance matrix of the many speaker feature data considering dispersion among the feature data of all speakers of analysis frame numbers corresponding to the mean value of the standard speaker feature data.</p> |