Bootstrap Equilibrium and Probabilistic Speaker Representation Learning for Self-Supervised Speaker Verification

In this paper, we propose self-supervised speaker representation learning strategies, which comprise of a bootstrap equilibrium in the front-end and an uncertainty-aware probabilistic embedding training back-end. stage, learn representations via scheme with uniformity regularization term. back-end embeddings are estimated by maximizing mutual likelihood score between speech samples belonging to same speaker, provide not only but also data uncertainty. Experimental results show that proposed strategy can effectively help outperforms conventional methods based on contrastive learning. Also, demonstrate integrated two-stage framework further improves verification performance VoxCeleb1 test set terms EER MinDCF.