This report presents the submission of UNISOUND for both Track 1 and Track 2 of the VoxCeleb Speaker Recognition Challenge 2023 (VoxSRC 2023). Our system, which is trained solely on VoxCeleb2-dev, utilizes two architectures: Large-scale ResNet and RepVGG. Additionally, we introduce a novel consistency-aware score calibration method that capitalizes on the stability of audio voiceprints through a Consistency Measure Factor (CMF). The incorporation of CMF significantly enhances our system’s performance in this challenge. Our final submission combines six models and achieves first place in Track 1 and second place in Track 2 of VoxSRC 2023. Notably, our submission boasts a minDCF of 0.0855 and an EER of 1.5880%.
A Study on the UNISOUND System for VoxCeleb Speaker Recognition Challenge 2023 (arXiv:2308.12526v1 [eess.AS])
