Fast Adaptation of Pretrained Speaker Verification System for Source Speaker Tracking

Traditional speaker verification system aims at distinguish speaker identity in real world audio, and has achieved satisfying performance in many scenarios. However, it is also very vulnerable, and can be easily attacked by voice anonymization system. In this report, we describe how to fast adapt a pretrained speaker verification model to source speaker tracking task with pretrained feature and Lora [1] technique. It significantly reduce EER on voice anonymization system, as well as keep its performance in real world audio intact. Experiment on Attacker Challenge [2] shows that our system successfully reduce baseline EER by 32% in average, and achieve lowest EER in all voice anonymization system except T8-5.