Singing Voice Deepfake Detection

SingGraph and our follow-up work on singing voice deepfake (SingFake) detection

Papers · Datasets · Citation · Contact

Research line

• SingGraph is the first method proposed specifically for singing voice deepfake detection.
• Instrumental Music Analysis studies what detectors learn from accompaniment.
• Sing-HiResNet extends detection to high-resolution audio.

Papers

Singing Voice Graph Modeling for SingFake Detection

Xuanjun Chen, Haibin Wu, Jyh-Shing Roger Jang, Hung-yi Lee · Lightning Talk at SLT 2024

• SingGraph: The first method proposed specifically for SingFake detection, combining MERT for pitch and rhythm with wav2vec 2.0 for lyrics through graph modeling.
• Music-informed augmentation: Uses RawBoost and beat matching for singing voice augmentation.
• Result: Improves EER on the SingFake dataset by 13.2% for seen singers, 24.3% for unseen singers, and 37.1% for unseen singers with different codecs, relative to the previous state of the art.

Code: This repository contains the official SingGraph code. Set the dataset and model checkpoint paths in utils/SingGraph.conf, then run:

pip install -r requirements.txt
bash run.sh

---

How Does Instrumental Music Help SingFake Detection?

Xuanjun Chen, Chia-Yu Hu, I-Ming Lin, Yi-Cheng Lin, I-Hsiang Chiu, You Zhang, Sung-Feng Huang, Yi-Hsuan Yang, Haibin Wu, Hung-yi Lee, Jyh-Shing Roger Jang

• Behavioral analysis: Tests different backbones, unpaired instrumental tracks, and frequency subbands.
• Key finding: Instrumental accompaniment acts mainly as data augmentation rather than providing intrinsic cues such as rhythm or harmony.
• Representational analysis: Fine-tuning increases reliance on shallow speaker features while reducing sensitivity to content, paralinguistic, and semantic information.

---

Joint Fullband-Subband Modeling for High-Resolution SingFake Detection

Xuanjun Chen*, Chia-Yu Hu*, Sung-Feng Huang, Haibin Wu, Hung-yi Lee, Jyh-Shing Roger Jang

• High-resolution analysis: Presents the first systematic analysis of 44.1 kHz audio for SingFake detection.
• Sing-HiResNet: Combines a fullband model for global context with subband experts for fine-grained synthesis artifacts.
• Result: Shows that high-frequency subbands provide essential complementary cues, significantly outperforming 16 kHz models on WildSVDD.

Datasets

Dataset	Used by	Access
SingFake	SingGraph, Instrumental Music Analysis	Not open source because of copyright. Please follow the instructions in the SingFake paper to download it.
WildSVDD	Sing-HiResNet	Because of copyright, the data are re-collected following the official protocols.

Citation

If you find our work useful, please consider citing:

@inproceedings{chen24o_interspeech,
  title     = {Singing Voice Graph Modeling for SingFake Detection},
  author    = {Chen, Xuanjun and Wu, Haibin and Jang, Jyh-Shing Roger and Lee, Hung-yi},
  booktitle = {{INTERSPEECH 2024}},
  pages     = {4843--4847},
  year      = {2024},
  doi       = {10.21437/Interspeech.2024-1185}
}

@inproceedings{chen2026how,
  title     = {How Does Instrumental Music Help {SingFake} Detection?},
  author    = {Chen, Xuanjun and Hu, Chia-Yu and Lin, I-Ming and Lin, Yi-Cheng and Chiu, I-Hsiang and Zhang, You and Huang, Sung-Feng and Yang, Yi-Hsuan and Wu, Haibin and Lee, Hung-yi and Jang, Jyh-Shing Roger},
  booktitle = {ICASSP 2026 - IEEE International Conference on Acoustics, Speech and Signal Processing},
  pages     = {19072--19076},
  year      = {2026},
  doi       = {10.1109/ICASSP55912.2026.11460829}
}

@inproceedings{chen2026joint,
  title     = {Joint Fullband-Subband Modeling for High-Resolution {SingFake} Detection},
  author    = {Chen, Xuanjun and Hu, Chia-Yu and Huang, Sung-Feng and Wu, Haibin and Lee, Hung-yi and Jang, Jyh-Shing Roger},
  booktitle = {{INTERSPEECH 2026}},
  year      = {2026},
  doi       = {10.21437/Interspeech.2026-1614}
}

Contact

If you have any questions, please contact d12942018@ntu.edu.tw.