Trial-dependent miscalibration can severely affect the performance of speaker verification systems. Global calibration methods address the problem by incorporating side-information into the calibration model. Alternatively, score normalization approaches exploit statistics computed from scores of impostor cohorts. While effective in some scenarios, the latter approaches suffer from poor global calibration, and in some cases may even increase trial-dependent miscalibration with respect to unnormalized scores. While the former issue can be addressed through global calibration, the latter problem can result in degraded performance. In this work, we provide a theoretical framework for incorporating impostor score statistics as side information in discriminative calibration models. Our approach allows us to improve both global and trial-dependent calibration, without incurring in some of the issues of score normalization. Results on SRE 2019 and SITW datasets show that our approach achieves similar or better (up to 15% relative) results compared to state-of-the-art score normalization techniques. The model can also be trivially extended to incorporate additional side-information.
Cite as: Cumani, S., Sarni, S. (2022) Impostor Score Statistics as Quality Measures for the Calibration of Speaker Verification Systems. Proc. The Speaker and Language Recognition Workshop (Odyssey 2022), 25-32, doi: 10.21437/Odyssey.2022-4
@inproceedings{cumani22_odyssey,
author={Sandro Cumani and Salvatore Sarni},
title={{Impostor Score Statistics as Quality Measures for the Calibration of Speaker Verification Systems}},
year=2022,
booktitle={Proc. The Speaker and Language Recognition Workshop (Odyssey 2022)},
pages={25--32},
doi={10.21437/Odyssey.2022-4}
}