Predicting Biometric Error Behaviour from Speaker Embeddings and a Fast Score Normalization Scheme

Abstract

For the task of speaker recognition from audio, it is known that speakers experience different levels of error rates. In this work, predicting the proneness to false alarm and false reject of a given speaker embedding is investigated. Although exact prediction of biometric error behaviour appears to be a difficult problem, it is seen that the tendency to false alarm and false reject errors can be predicted directly from embeddings by training a neural network in a supervised manner. This prediction might be useful for several applications such as normalization of verification scores, incorporating those characteristics in embedding training or using it as an adversarial objective. We have utilized this predicted behaviour for a fast score normalization method. Our approach is compared to the frequently employed biometric normalization method that is s-norm which is a cohort-based technique and accounts only for imposter calibration. The proposed normalization is not only faster than s-norm, but also it outperforms s-norm by 8% and 3% for male and female speakers, respectively.