Rehearsal with Auxiliary-Informed Sampling (RAIS)

This repository contains the implementation of RAIS.

Installation

Python==3.12.3

pip install torch==2.4.1 torchvision==0.19.1 torchaudio==2.4.1 --index-url https://download.pytorch.org/whl/cu121
pip3 install -r requirements.txt

Dataset

Metadata for train, validation, and evaluation splits can be found in dataset/.

• ASVspoof 2019 LA ($\mathcal{E}_0$): Download LA.zip.
• VCC 2020 ($\mathcal{E}_1$): Combined datasets from
  • Source 1
  • Source 2
• InTheWild ($\mathcal{E}_2$): Download release_in_the_wild.zip.
• CFAD ($\mathcal{E}_3$): Download all CFAD zip files from Zenodo.
• OpenAI-LJSpeech ($\mathcal{E}_4$):
  • Fake audios generated using the OpenAI Text-to-Speech API.
  • Bonafide samples from LJSpeech.

Usage

• Run command:

  python main.py --method rais

Cite this work

@inproceedings{febrinanto25_interspeech,
  title     = {{Rehearsal with Auxiliary-Informed Sampling for Audio Deepfake Detection}},
  author    = {{Falih Gozi Febrinanto and Kristen Moore and Chandra Thapa and Jiangang Ma and Vidya Saikrishna and Feng Xia}},
  year      = {{2025}},
  booktitle = {{Interspeech 2025}},
  pages     = {{5358--5362}},
  doi       = {{10.21437/Interspeech.2025-2298}},
  issn      = {{2958-1796}},
}

Acknowledgments

This project includes code adapted from Avalanche.