W-GVKT : within-global-view knowledge transfer for speaker verification

Abstract

Contrastive self-supervised learning has played an important role in speaker verification (SV). However, such approaches suffer from false-negative issues. To address this problem, we enhance the non-contrastive DINO framework by enabling knowledge transfer from the teacher network to the student network through diversified versions of global views and call the method Within-Global-View Knowledge Transfer (W-GVKT) DINO. We discovered that given the global view of the entire utterance, creating discrepancies in the student’s output through applying spectral augmentation and feature diversification to the global view can facilitate the transfer of knowledge from the teacher to the student. With negligible computational resource increases, W-GVKT achieves an impressive EER of 4.11% without utilizing speaker labels on Voxceleb1. When combined with the RDNIO framework, W-GVKT achieved an EER of 2.89%.