| 摘要 |
The invention relates to a method and an arrangement for adapting models in speaker verification systems or similar systems using models based on data collected from a person during a certain time period. If a simple model is utilized a less reliable verification is obtained, but if, on the other hand, a more complex model is utilized, the problem is a long training time period. The present invention solves this problem by using a plurality of different models in the same speaker verification system. The verification is put into operation using the model requiring a smaller amount of speech data. During the use, more speech data is collected continuously. This material is used to further train either only the more complex model or both the simpler model already put into operation and the more complex model. At suitable intervals, a comparison is made of the performance capacities of the models. Once the more complex model yields a more reliable verification result it will take over the verification in the operating situation. The subsequent model unit may be put into operation either instantaneously or gradually, e.g. by using a weight function. <IMAGE> |
