Determination of threshold for speaker verification using speaker adaptation gain in likelihood during training

This paper describes methods to determine thresholds for speaker verification. Setting an appropriate threshold a priori is difficult because likelihood verification covers a wide range and the appropriate threshold for each speaker is different. We propose new methods to determine the speaker verification threshold depending on the "adaptation degree" for each speaker. We use the gain in likelihood during the adaptive training process from speaker-independent models as the "adaptation degree" and determine the threshold by its linear function. We evaluate the proposed methods in text-prompted speaker verification experiments using connected digit speech to show that the estimated coefficients of the linear function are relatively constant regardless of the amount of training data and that thresholds set by our proposed methods are stable and reliable. Consequently, use of our new methods improves verification performance and reduces the error rate by 30 percent.