SingFake: Singing Voice Deepfake Detection

|

Official Repository for paper "SingFake: Singing Voice Deepfake Detection", accepted by ICASSP 2024. [Project Webpage]

Updates

• Feb 2024: We have launched the inaugural Singing Voice Deepfake Detection (SVDD) 2024 Challenge! Check out the challenge website here! Ongoing!
• Nov 2023: We released our metadata annotation tool (annotation-tool/) designed as a Chrome Extension to speed up the annotation process.
• Sep 2023: We released our training and evaluation scripts, as well as trained model checkpoints for reproducibility due to copyright concerns, we do not release our trained model checkpoints.

Directory Structure

dataset/ contains scripts related to preparing the dataset. Assuming you have a directory filled with downloaded FLAC files, you could run them first through separate.py to generate separated vocal stems and generate VAD timecodes, then use split.py to generate separated audio clips for training. We also provide simulate_codec.py, which is being used for generating our T03 subset.

models/ contains script for training and evaluation of our four baseline models. feat_resnet contains implementation for Spectrogram+ResNet; lfcc_resnet contains implementation for LFCC+ResNet; AASIST and wav2vec2+AASIST contains their corresponding implementations.

Annotation Tool

The metadata annotation tool is designed as a Chrome Extension built on top of a Google Firestore backend. To use this, you can start a free-tier project under Google Firebase, enable Firestore, and fill in your credentials under annotation-tool/background.js's firebaseConfig variable.

Citation

@inproceedings{zang2024singfake,
  title={SingFake: Singing Voice Deepfake Detection},
  author={Zang, Yongyi and Zhang, You and Heydari, Mojtaba and Duan, Zhiyao},
  booktitle={Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages={1--5},
  year={2024},
  organization={IEEE}
}