| 摘要 |
<p>PURPOSE:To improve the collating rate by deciding a registered speaker when the value of the output unit of a collation neural network is larger than a certain value. CONSTITUTION:The system is constituted by coupling a voice input part 10, a preprocessing part 20, a collation neural network 30, and the discrimination neural network 40. Then the discrimination neural network 40 is equipped with the output unit which outputs a discrimination output indicating which of a registered speaker group and an unregistered speaker group a speaker is in and further has threshold values thetax for registered speaker discrimination making by registered speakers, and those threshold values thetax are learnt. Consequently, the decision neural network is structured while differences in output patterns among speakers are reflected, so the collating rate is improved.</p> |
