Self-supervised learning with multi-head multi-mode knowledge distillation for speaker verification

Abstract

Training speaker verification (SV) systems without labeled data is challenging. To tackle the challenge, we propose Multi-Head, Multi-Mode (MeMo) self-supervised learning based on knowledge distillation. Unlike DINO, the teacher in MeMo uses two distinct architectures to learn collaboratively, and so does the student. MeMo employs two distillation modes: self- and cross-distillations, with the teacher and student having the same and different architectures, respectively. To reduce the output discrepancy caused by different architectures, we divide the projection head into self- and cross-heads so that each head is responsible for distillation in its respective mode. We also discover that contrastive learning at the embedding level is supportive only in early training stages. To address this issue, we propose dynamically stopping the contrastive learning while continuing knowledge distillation. MeMo achieves an impressive EER of 3.10% on Voxceleb1 using a small ECAPA-TDNN backbone.