Confidence Estimator for Speaker Verification Using Dual DET Curves

摘要

In speaker verification, the result of a trial is traditionally summarised as an arbitrary score, where a higher score indicates stronger evidence in favour of the speaker hypothesis. However this is difficult to interpret. It is useful to convert this score into a 'confidence level', i.e. the posterior probability that the speaker hypothesis is correct, given the score. One of the simplest formulae to obtain a confidence level is using a logistic curve, but this requires the assumption that the true and impostor speaker scores are distributed according to a Normal distribution. In this report I propose a new formula, called the dual Detection Error Trade-Off (DET) curve, since it represents the same information as a DET curve. This formula avoids the assumption of normally distributed target and impostor scores. Experiments on the NIST 99 data prove the dual DET curve performs slightly better than the logistic curve.