Dataset Card for VCapAV

VCapAV is a large-scale audio-visual deepfake detection dataset focused on non-speech environmental sounds. It introduces new multimodal deepfake scenarios using both Text-to-Audio (TTA) and Video-to-Audio (V2A) pipelines, together with Text-to-Video (TTV) synthesis.
The dataset contains 90,990 clips, totaling 252.75 hours, and supports audio-only, visual-only, and audio-visual detection tasks.

Dataset Description

VCapAV addresses the lack of multimodal deepfake data involving environmental sounds. Unlike existing datasets focused on speech or face-centric manipulations, VCapAV introduces a comprehensive set of environmental audio generation methods and high-fidelity video forgeries.

• Curated by: Duke Kunshan University, University of Yamanashi, Wuhan University
• Funded by: DKU Foundation Project “Emerging AI Technologies for Natural Language Processing”
• Shared by: Authors of the VCapAV paper
• Language(s): English (captions)
• License: MIT License

Dataset Sources

• Repository: https://github.com/wailywang/VCapAV/
• Paper: VCapAV: A Video-Caption Based Audio-Visual Deepfake Detection Dataset
• Demo: https://vcapav.github.io/

Dataset Uses

• Audio anti-spoofing research
• Audio-visual deepfake detection
• Evaluation of general-purpose audio generation methods
• Studying modality consistency between vision and sound
• Research on multimodal synchronization, scene-aware generation, and cross-modal alignment

Dataset Creation

Most deepfake datasets focus on speech or human faces. VCapAV fills this gap by focusing on general environmental audio and video–audio consistency, enabling research on non-speech deepfake detection.

The dataset is constructed from a subset of VGGSound (15,446 videos).

Citation

@inproceedings{wang2025vcapav,
  title={VCapAV: A Video-Caption Based Audio-Visual Deepfake Detection Dataset},
  author={Wang, Yuxi and Wang, Yikang and Zhang, Qishan and Nishizaki, Hiromitsu and Li, Ming},
  booktitle={Interspeech},
  year={2025}
}

---