sphereface2_speaker_verification

official implementation of sphereface2 for speaker verification in Exploring Binary Classification Loss for Speaker Verification, this code is based on wespeaker

advantages:

1. better performance, especially on hard trials
2. robust to noisy labels
3. natural parallelization for the classifier layer

Running

SphereFace2 has been supported in Wespeaker toolkit, see #173. Welcome to use wespeaker for developing and research.

Results

ResNet34-TSTP-emb256

Model	Margin	Params	LM	AS-Norm	vox1-O-clean	vox1-E-clean	vox1-H-clean
AAM	0.2	6.63M	×	×	0.867	1.049	1.959
	0.2		×	√	0.787	0.964	1.726
	0.5		√	×	0.797	0.937	1.695
	0.5		√	√	0.723	0.867	1.532
C-Sphereface2	0.2	6.63M	×	×	0.904	0.973	1.737
	0.2		×	√	0.835	0.931	1.652
	0.3		√	×	0.830	0.862	1.510
	0.3		√	√	0.755	0.833	1.449
A-Sphereface2	0.15	6.63M	×	×	0.835	0.975	1.742
	0.15		×	√	0.761	0.938	1.630
	0.25		√	×	0.766	0.899	1.590
	0.25		√	√	0.686	0.852	1.480

Citations

If you find its useful, please cite it as

@inproceedings{han2023exploring,
  title={Exploring Binary Classification Loss for Speaker Verification},
  author={Han, Bing and Chen, Zhengyang and Qian, Yanmin},
  booktitle={ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages={1--5},
  year={2023},
  organization={IEEE}
}

@InProceedings{wen2021sphereface2,
  title = {SphereFace2: Binary Classification is All You Need for Deep Face Recognition},
  author = {Wen, Yandong and Liu, Weiyang and Weller, Adrian and Raj, Bhiksha and Singh, Rita},
  booktitle = {ICLR},
  year = {2022}
}