The DKU-MSXF Speaker Verification System for the VoxCeleb Speaker Recognition Challenge 2023

• DOI: 10.48550/arxiv.2308.08766
• Publication Date: 2023-01-01
• Authors: Ze Li; Yuke Lin; Xiaoyi Qin; Ning Jiang; Guoqing Zhao; Ming Li
• Keywords: Training set; Domain Adaptation; Performance Improvement; Speaker diarisation

This paper is the system description of DKU-MSXF System for track1, track2 and track3 VoxCeleb Speaker Recognition Challenge 2023 (VoxSRC-23). For Track 1, we utilize a network structure based on ResNet training. By constructing cross-age QMF training set, achieve substantial improvement in performance. 2, inherite pre-trained model from 1 conducte mixed by incorporating VoxBlink-clean dataset. In comparison to models data exhibit performance more than 10% relatively. Track3, semi-supervised domain adaptation task, novel pseudo-labeling method triple thresholds sub-center purification adopted make adaptation. The final submission achieves mDCF 0.1243 task1, 0.1165 2 EER 4.952% 3.