Multi-Dataset Co-Training with Sharpness-Aware Optimization for Audio Anti-spoofing

Official Repository for the paper "Multi-Dataset Co-Training with Sharpness-Aware Optimization for Audio Anti-spoofing", published in Interspeech 2023 (Paper)

Dependencies

pip install -r requirements.txt

Training

1. Download ASVspoof2019 dataset
2. Change the "database_path" in AASIST-L_SAM.conf file (or Move data to data folder)

python main.py --config ./config/AASIST-L_SAM.conf

Evaluation using pre-trained model

python main.py --eval --config ./config/AASIST-L_SAM.conf

• Performance
  • ASVspoof2019 evaluation(SAM w/o additional training datasets): EER 1.06%

Citation

@article{shim2023multidataset,
      title={Multi-Dataset Co-Training with Sharpness-Aware Optimization for Audio Anti-spoofing}, 
      author={Hye-jin Shim and Jee-weon Jung and Tomi Kinnunen},
      journal={Proc. Interspeech},
      year={2023}
}

Reference

This code refers to the the following repositories:

1. https://github.com/clovaai/aasist
2. https://github.com/davda54/sam

License

This code is licensed under MIT license