Progressive contrastive learning for self-supervised text-independent speaker verification

self-supervised, text-independent, speaker, verification

Self-supervised speaker representation learning has drawn attention extensively in recent years. Most of the work is based on the iterative clustering-classification learning framework, and the performance is sensitive to the pre-defined number of clusters. However, the cluster number is hard to estimate when dealing with large-scale unlabeled data. In this paper, we propose a progressive contrastive learning (PCL) algorithm to dynamically estimate the cluster number at each step based on the statistical characteristics of the data itself, and the estimated number will progressively approach the ground-truth speaker number with the increasing of step. Specifically, we first update the data queue by current augmented samples. Then, eigendecomposition is introduced to estimate the number of speakers in the updated data queue. Finally, we assign the queued data into the estimated cluster centroid and construct a contrastive loss, which encourages the speaker representation to be closer to its cluster centroid and away from others. Experimental results on VoxCeleb1 demonstrate the effectiveness of our proposed PCL compared with existing self-supervised approaches.