摘要 |
<p>PROBLEM TO BE SOLVED: To provide a speaker collating device for selecting a registered speaker in collation without damaging the real time property. SOLUTION: This speaker collating device comprises a sound characteristic quantity memory 106 for storing a set of short-time sound characteristics for indefinite speaker voice recognition which is formed of voices of a plurality of speakers and does not depend on speaker and vocabulary, and a means 108 for calculating the sound similarity of the characteristic quantity of each time of a voice input pattern to the characteristic quantity stored in the sound characteristic quantity memory part 106. The similarity of the input pattern to the registered pattern of a registered speaker is calculated by use of the sound similarity and the time series of the index of each preliminarily registered speaker, and a speaker used for normalization is selected to calculate the likelihood used for registering. After registering, the likelihood is determined, and the likelihood is compared with a preset threshold, whereby the speaker is received as the registered person or disposed.</p> |