On the influence of metric learning loss functions for robust self-supervised speaker verification to label noise

While clustering-driven Pseudo-Labels (PLs) are commonly employed to optimize Speaker Embedding (SE) networks and facilitate training of self-supervised Speaker Verification (SV) systems, the efficacy of PL-based self-supervised training hinges on the accuracy of these generated labels. In this paper, we perform a large-scale comparative study of a wide range of recent metric learning loss functions for better generalization of self-supervised SV systems. In particular, we investigate the effect of these losses on the robustness of the self-supervised SV task against label noise using various real-life clustering-based PLs. We present an extensive comparative evaluation of the performance of these loss functions using different numbers of clusters and show that our proposed selection of loss functions is effective against label noise and leads to considerable improvements in SV performance. Moreover, using our selected losses combined with the adopted CAMSAT clustering algorithm-based PLs to train our SE system allows us to achieve state-of-the-art self-supervised SV performance. Code of our experiments will be made publicly available.