A Novel Self-supervised Representation Learning Model for an Open-Set Speaker Recognition

Speaker recognition is an important problem in behavioral biometric domain. Supervised speaker recognition systems have rapidly evolved since the development of deep learning (DL) architectures. Despite advancements in supervised speaker recognition, an open-set speaker clustering remains a challenging problem. This paper proposes a novel self-supervised representation learning architecture that laverages bi-modal architecture based on CNN and MLP sub-networks. A novel combination of angular prototypical loss and cosine similarity loss ensure an excellent clustering parity, while data augmentation results in a better generalization of the model. The experimental results convincingly demonstrate that the proposed archietcture outperforms state-of-the-art speaker verification methods on VoxCeleb1 and LibriSpeech datasets.