metadata
license: apache-2.0
VidAudio-Bench: Benchmarking V2A and VT2A Generation across Five Audio Categories
1. Overview
VidAudio-Bench is a comprehensive multi-task benchmark designed for evaluating both Video-to-Audio (V2A) and Video-Text-to-Audio (VT2A) generation systems.
The benchmark comprises 1,634 video-text pairs spanning five audio categories:
- SFX: 400 samples
- Music: Instrument: 191 samples ; BGM: 231 samples
- Speech: 412 samples
- Singing: 400 samples
VidAudio-Bench aims to facilitate the evaluation of:
- Audio-video synchronization
- Semantic consistency between video and generated audio
- Text-guided controllable audio generation
- Fine-grained audio category understanding
2. Dataset Structure
The dataset is organized as follows:
VidAudio-Bench/
├── original_videos/ # original video files (.mp4)
├── silent_videos/ # muted video files (.mp4)
├── annotations/ # metadata files (.jsonl)
3. Data Format
Each sample in the annotation file follows the format below:
{
"video_name": "6TiZ1coebWw_76s",
"category": "SFX",
"V2A_positive_prompt": "Realistic foley sound synchronized with the video.",
"negative_prompt": "music, background music, speech, singing",
"VT2A_positive_prompt": "Realistic foley sound of a fan rotating."
}
4. Citation
If you use this dataset, please cite:
@article{zhang2026vidaudio,
title={VidAudio-Bench: Benchmarking V2A and VT2A Generation across Five Audio Categories},
author={Zhang, Qian and Cao, Yuqin and Gao, Yixuan and Min, Xiongkuo},
journal={arXiv preprint arXiv:2604.10542},
year={2026}
}