OR-Gate

Official PyTorch implementation of the paper "Robust Training for Speaker Verification against Noisy Labels" in INTERSPEECH 2023.

Data preparation

./required_data.sh

Training

./run.sh

• main.py: Main program framework;
• parser.py: Parameters and experimental settings.
  • train_list_path: Training file path;
  • device: GPU device;
  • warm_up: The first stage trains the number of epochs (default: warm_up=5), when warm_up=max_epochs is baseline;
  • topk: The value of k in the top-k mechanism (default: topk=90).

Results

VoxCeleb 1 (EERs):

noisy rate	0%	5%	10%	20%	30%	50%
Baseline	4.36	5.30	6.24	7.99	9.78	14.39
OR-Gate	4.08	4.07	4.22	4.28	4.41	5.53

VoxCeleb 2 (EERs):

noisy rate	0%	5%	10%	20%	30%	50%
Baseline	1.69	1.72	1.90	2.21	2.88	4.32
OR-Gate	1.64	1.65	1.62	1.67	1.72	1.97

Acknowledge

This code is based on Lantian Li/Sunine.

Reference

@inproceedings{fang23_interspeech,
  author={Zhihua Fang and Liang He and Hanhan Ma and Xiaochen Guo and Lin Li},
  title={{Robust Training for Speaker Verification against Noisy Labels}},
  year=2023,
  booktitle={Proc. INTERSPEECH 2023},
  pages={3192--3196},
  doi={10.21437/Interspeech.2023-452}
}

Contact

• E-mail: [email protected]
• Wechat: PunkMale