Cluster-Guided Unsupervised Domain Adaptation for Deep Speaker Embedding

Recent studies have shown that pseudo labels can contribute to unsupervised domain adaptation (UDA) for speaker verification. Inspired by the self-training strategies use an existing classifier label unlabeled data retraining, we propose a cluster-guided UDA framework target clustering and combines labeled source pseudo-labeled train embedding network. To improve cluster quality, network dedicated minimizing contrastive center loss. The goal is reduce distance between its assigned while enlarging other centers. Using VoxCeleb2 as CN-Celeb1 domain, demonstrate proposed method achieve equal error rate (EER) of 8.10% on evaluation set without using any from domain. This result outperforms supervised baseline 39.6% state-of-the-art performance this corpus.