Investigation Of The Quality Of Pseudo-Labels For The Self-Supervised Speaker Verification Task

Optimizing a speaker embedding network in a discriminative fashion using clustering algorithm-driven pseudo-labels is one of the most widely used self-supervised speaker verification system training schemes. Although this kind self-supervised supervised training scheme showed impressive performance, recent studies have shown that label noise can significantly impact the performance. In this contribution, we explore various clustering algorithms to generate speaker pseudo-labels and conduct a fine-grained analysis on the relationship between the quality of the pseudo-labels and the speaker verification performance. Through experimental results, we also shed light on several previously unexplored and overlooked aspects of the pseudo-labels that can have an impact on the speaker verification performance. Furthermore, we observe that the performance of a self-supervised speaker verification system relies heavily on multiple qualitative aspects of the clustering algorithm used to generate the pseudo-labels. Additionally, we show that severe reduction in speaker verification performance can occur from overfitting to the noisy pseudo-labels and that the mixup data augmentation strategy can mitigate the memorization effects of label noise.