Datasets:

FoleyBench
/

foleybench

license: cc-by-nc-sa-4.0
task_categories:
  - any-to-any
  - text-to-audio
language:
  - en
tags:
  - video-to-audio
  - foley
size_categories:
  - 1K<n<10K
extra_gated_prompt: >
  The use of the FoleyBench dataset and the original videos is governed by the
  Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC
  BY-NC-SA 4.0) license. This dataset is intended for academic research purposes
  only. Commercial use in any form is strictly prohibited.


  Before using the dataset, you must read and agree to the following conditions:


  Compliance with Original Licenses - FoleyBench is a collection of videos from
  various sources, each with its own Creative Commons license. Any use of the
  videos, in whole or in part, must adhere to the terms of their original
  licenses, including any attribution requirements. We provide provenance
  information for each data point to facilitate this.


  Data Removal and Updates - The dataset will be updated periodically to remove
  data upon request. By accessing this repository, you agree to -


  (1) Update your version of the dataset to the most recent version specified by
  the maintainers.


  (2) Immediately delete any specific videos from your records upon notification
  from the dataset maintainers.


  No Warranty and Content Disclaimer - The videos in this dataset are provided
  "as is." We do not own the copyright to any of the raw video files. Given the
  nature of this dataset, it is impossible for us to review all the video files.
  We do not vouch for or warrant the accuracy, completeness, or usefulness of
  any file and are not responsible for the contents of any files. Some files may
  contain objectionable material.


  Sharing and Distribution - If you host, share, or otherwise provide access to
  this dataset, you must include these Terms and Conditions and require all
  users to agree to them.


  Contact Information - By clicking "Access repository," you agree that your
  contact information (email address and username) may be shared with the
  dataset maintainers. We may notify users via email when the dataset is
  updated.


  Infringement - If you believe that any content in this dataset infringes on
  your rights, please contact us at satvikdixit7@gmail.com to request its
  removal. We will address any such requests immediately.


  By accessing the repository, you acknowledge that you have read, understood,
  and agree to be bound by these terms and conditions.
dataset_info:
  features:
    - name: key
      dtype: int64
    - name: duration
      dtype: float64
    - name: dataset
      dtype: string
    - name: width
      dtype: float64
    - name: height
      dtype: float64
    - name: caption
      dtype: string
    - name: discrete_vs_rest
      dtype: string
    - name: source_type
      dtype: string
    - name: sound_type
      dtype: string
    - name: ucs_category
      dtype: string
    - name: audioset_category
      dtype: string
    - name: metadata
      dtype: string
    - name: video_path
      dtype: string
    - name: video_data
      dtype: string
  splits:
    - name: train
      num_bytes: 13189754295
      num_examples: 5000
  download_size: 13000000000
  dataset_size: 13189754295
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
viewer: false

FoleyBench: A Benchmark For Video-to-Audio Models

Generating realistic sound for video (V2A) is a major challenge, especially for Foley—the art of creating sound effects synchronized with on-screen actions. Current benchmarks for this task are flawed. They are often contaminated with speech, music, and off-screen sounds, making it impossible to truly evaluate a model's ability to generate accurate, synchronized Foley. This misalignment leads to misleading conclusions about model performance.

We introduce FoleyBench, the first large-scale benchmark designed for evaluating Foley-style V2A generation.

High-Quality, Foley-Focused: Contains 5,000 video-audio-text triplets meticulously curated for Foley. All content is non-speech/non-music, with strong causal links between visible actions and their sounds.
Diverse Category Coverage: Ensures a broad and balanced distribution across the Universal Category System (UCS), addressing a key limitation of previous datasets.
Rich Metadata for Deep Analysis: Each clip is annotated with source complexity (single vs. multi-source) and sound type (discrete vs. continuous), enabling fine-grained analysis of model strengths and weaknesses.

Dataset Description

Total Videos: 5,000
Video Duration: ~10 seconds each
Total Size: ~13 GB
Format: MP4 with embedded audio
Resolution: Various (mostly 640x360)

Dataset Structure

foleybench.csv: Metadata file with video information including captions, duration, and source metadata
data/train-0-of-9.parquet ... data/train-9-of-9.parquet: Ten shards with embedded base64 video content (500 videos per shard)

Features

Each example in the dataset contains:

key: Unique identifier for the video
duration: Video duration in seconds
dataset: Source dataset name
width: Video width in pixels
height: Video height in pixels
caption: Text description of the video content
discrete_vs_rest: Discrete/rest label
source_type: Source type label
sound_type: Sound type label
ucs_category: UCS category label
audioset_category: AudioSet category label
metadata: JSON string with source information (YouTube ID, timestamps, etc.)
video_path: Relative path of the source video
video_data: Base64-encoded video content

Usage

from datasets import load_dataset
import base64

# Load the dataset
dataset = load_dataset("FoleyBench/foleybench")

# Access video data
first_example = dataset["train"][0]
print(f"Caption: {first_example['caption']}")
print(f"Duration: {first_example['duration']} seconds")

# Decode video
video_b64 = first_example["video_data"]
video_bytes = base64.b64decode(video_b64)

# Save video to file
with open("video.mp4", "wb") as f:
    f.write(video_bytes)

Data Access

This dataset is gated. Visit the repository page and click "Access repository" to agree to the terms and gain access.

License

This dataset is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0) license.

Important: This dataset is intended for academic research purposes only. Commercial use is strictly prohibited.

Contact

For questions, concerns, or data removal requests, please contact: satvikdixit7@gmail.com

Citation

If you find FoleyBench useful in your research, please consider citing our paper:

@misc{dixit2025foleybenchbenchmarkvideotoaudiomodels,
      title={FoleyBench: A Benchmark For Video-to-Audio Models}, 
      author={Satvik Dixit and Koichi Saito and Zhi Zhong and Yuki Mitsufuji and Chris Donahue},
      year={2025},
      eprint={2511.13219},
      archivePrefix={arXiv},
      primaryClass={cs.SD},
      url={[https://arxiv.org/abs/2511.13219](https://arxiv.org/abs/2511.13219)}, 
}