A Generative Model for Duration-Dependent Score Calibration

In this work we introduce a generative score calibration model for speaker verification systems able to explicitly account for utterance-dependent miscalibration sources, with a focus on segment duration. The model is theoretically motivated by an analysis of the effects of distribution mismatch on the scores produced by Probabilistic Linear Discriminant Analysis (PLDA), and extends our previous investigation on the distribution of well-calibrated PLDA log-likelihood ratios. We characterize target and non-target scores by means of Variance-Gamma densities, whose parameters represent effective between and within-class variabilities. Experimental results on SRE 2019 show that the proposed method improves both calibration and verification accuracy with respect to duration-agnostic models and to duration-aware discriminative methods.