Adaptive Margin Circle Loss for Speaker Verification

TL;DR

This paper introduces an adaptive margin circle loss function for speaker verification, improving angular discrimination and achieving state-of-the-art results on Voxceleb and SITW datasets.

Contribution

It proposes a novel adaptive margin circle loss with stage-based and chunk-based margins, enhancing training flexibility and convergence in speaker verification systems.

Findings

Achieves 1.31% EER on Voxceleb1

Achieves 2.13% EER on SITW

Demonstrates improved angular discrimination

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.