Contrastive Information Maximization Clustering for Self-Supervised Speaker Recognition

Pseudo-labels (PLs) generated through clustering are extensively employed to optimize speaker embedding (SE) networks and to train self-supervised speaker verification (SV) systems. However, the effectiveness of PL-based self-supervised training is contingent on the quality of the PLs, and achieving high clustering performance often requires time-consuming and resource-intensive data augmentation regularization. In this paper, we introduce an efficient, general-purpose multi-objective clustering algorithm that outperforms all other baseline methods for clustering SEs. Our approach, named Contrastive Information Maximization Clustering (CIMC), circumvents the need for explicit data augmentation, enabling rapid training with minimal memory and computational resource usage. CIMC is founded on three key principles: (1) Self-Augmented Training, which ensures representation invariance and maximizes the information-theoretic dependency between samples and their predicted PLs (2) Virtual Mixup Training, which enforces local-Lipschitzness and upholds the cluster assumption (3) Supervised contrastive learning, which fosters the learning of more discriminative features and enhances robustness to natural corruptions by bringing together samples of the same class while separating those of different clusters. We present a comprehensive comparative analysis of our clustering method against baselines using various clustering metrics, conduct an ablation study to assess the contribution of each component, and demonstrate that our multi-objective approach provides beneficial complementary information. Furthermore, utilizing the generated PLs to train our SE system enables us to achieve high SV performance.