Current browse context:
cs.SD
Change to browse by:
References & Citations
Computer Science > Sound
Title: Adaptive Margin Circle Loss for Speaker Verification
(Submitted on 15 Jun 2021)
Abstract: Deep-Neural-Network (DNN) based speaker verification sys-tems use the angular softmax loss with margin penalties toenhance the intra-class compactness of speaker embeddings,which achieved remarkable performance. In this paper, we pro-pose a novel angular loss function called adaptive margin cir-cle loss for speaker verification. The stage-based margin andchunk-based margin are applied to improve the angular discrim-ination of circle loss on the training set. The analysis on gradi-ents shows that, compared with the previous angular loss likeAdditive Margin Softmax(Am-Softmax), circle loss has flexi-ble optimization and definite convergence status. Experimentsare carried out on the Voxceleb and SITW. By applying adap-tive margin circle loss, our best system achieves 1.31%EER onVoxceleb1 and 2.13% on SITW core-core.
Link back to: arXiv, form interface, contact.