| 摘要 |
<p>A speaker recognition system tests data during enrolment by preparing a template for a word from repeated utterances, comparing further repeated utterances with the template and updating the template with accepted utterances, and finally accepting or rejecting the updated template on the basis of how many of the further utterances are accepted and how many rejected. This protects the quality of the template formed from the enrolment utterances. In addition to storing final versions of the templates for each of a plurality of words, data is stored relating to the co-variance of utterances of different words. During analysis of the spoken inputs, the pitch of voiced sections is determined and this is used both to enable determination of the start and end points of the spoken utterance and also to determine the timing of analysis frame periods. During speaker recognition, a plurality of test utterances are obtained from the unknown speaker, and these are analysed separately and the analysis results combined using the co-variance data to make a recognition decision. The system is particularly useful for speaker verification to verify a previous claimed identity. The system has particular application to the recognition of individuals over the telephone.</p> |
