Label-Efficient Self-Supervised Speaker Verification With Information Maximization and Contrastive Learning

Abstract

Most state-of-the-art self-supervised speaker verification systems rely on a contrastive-based objective function to learn speaker representations from unlabeled speech data. We explore different ways to improve the performance of these methods by: (1) revisiting how positive and negative pairs are sampled through a "symmetric" formulation of the contrastive loss; (2) introducing margins similar to AM-Softmax and AAM-Softmax that have been widely adopted in the supervised setting. We demonstrate the effectiveness of the symmetric contrastive loss which provides more supervision for the self-supervised task. Moreover, we show that Additive Margin and Additive Angular Margin allow reducing the overall number of false negatives and false positives by improving speaker separability. Finally, by combining both techniques and training a larger model we achieve 7.50% EER and 0.5804 minDCF on the VoxCeleb1 test set, which outperforms other contrastive self supervised methods on speaker verification.

Trinet: Stabilizing Self-Supervised Learning From Complete or Slow Collapse

“…Among them, W-MSE [14], Barlow-Twinss [15], and VICReg [10] attempt to produce embedding variables that are decorrelated from each other, whereas CorInfoMax [13] does not constrain the variables to be uncorrelated but instead avoids covariance matrix degeneracy by using log-determinant as a regularizer loss function. However, recent investigations show that these regularization terms worked effectively only if given specific SSL structural settings [10] and strong data augmentation [16]. Note that all these regularization methods [10,13,14,15] adopt an SSL-no-SG structure, where "no-SG" means the branch networks are both learnable with no stop-gradient.…”