On the impact of several regularization techniques on label noise robustness of self-supervised speaker verification systems

Clustering-based Pseudo-Labels (PLs) are widely used to optimize Speaker Embedding (SE) networks and train Self-Supervised (SS) Speaker Verification (SV) systems. However, this SS training scheme relies on highly accurate PLs. In this paper, we perform a large investigative study of the effect of several regularization techniques (mixup, label smoothing, employing sub-centers) on the label noise robustness of SSSV systems. We study these techniques and apply them on various recent metric learning loss functions for better generalization of SSSV systems. In particular, we investigate the effect of these losses and regularizations on the robustness of the self-supervised SV task against label noise using the CAMSAT clustering model to generate PLs. We provide a thorough comparative analysis of the performance of these techniques using different numbers of clusters and show that some of them are effective against label noise and lead to considerable improvements in SV performance.