Datasets:

zyk21
/

VoxDub

license: cc-by-nc-4.0
task_categories:
  - audio-to-audio
  - automatic-speech-recognition
  - video-classification
language:
  - zh
  - en
  - multilingual
tags:
  - audiovisual
  - speech
  - lip-sync
  - youtube
  - annotations
pretty_name: VoxDub
size_categories:
  - 100K<n<1M

VoxDub

VoxDub provides segment-level audiovisual annotations derived from public YouTube videos. This repository hosts the annotation archive (av_segments_v1.tar.zst). Raw media is not redistributed; reconstruct clips from YouTube using the companion pipeline.

Dataset: zyk21/VoxDub
Pipeline: see the av_pipeline tooling shipped with this release (download → cut → standardize → optional vocal separation)

Dataset summary

Item	Value
Videos (YouTube IDs)	~17,433
Segments	~766,708
Package	`av_segments_v1.tar.zst`
Layout	`datas/{video_id}/{seg_id}.json`
License	CC BY-NC 4.0

Files

av_segments_v1.tar.zst
└── datas/
    └── {youtube_video_id}/
        ├── S00001.json
        ├── S00002.json
        └── ...

Unpack:

# Python
import zstandard as zstd, tarfile
dctx = zstd.ZstdDecompressor()
with open("av_segments_v1.tar.zst", "rb") as f, dctx.stream_reader(f) as r:
    with tarfile.open(fileobj=r, mode="r|") as tar:
        tar.extractall("av_segments_v1")

Or with the pipeline:

python pipeline.py extract

Annotation schema

Each JSON file describes one temporal segment of a YouTube video.

Field	Type	Description
`id`	string	Segment id (e.g. `S00023`)
`start` / `end`	float	Time range in seconds
`text_whisper`	string	Whisper transcript
`text_paraformer`	string	Paraformer transcript
`language`	string	Detected language
`language_whisper_prob`	float	Language confidence
`wer`	float	Word error rate (ASR comparison)
`dnsmos`	float	DNSMOS speech quality score
`gender`	int	Speaker gender label
`multi_speaker`	float	Multi-speaker score
`av_offset`	int	Audio–visual offset (frames)
`sync_conf`	float	AV sync confidence
`origin_width` / `origin_height`	int	Source resolution
`scene_num`	int	Scene index
`faces`	object	Per-frame face tracks (`n_frames`, `score`, `bbox`, `landmarks`)

faces.n_frames is typically ≈ (end - start) * 25.

Example

{
  "id": "S00023",
  "start": 311.432,
  "end": 335.618,
  "language": "chinese",
  "origin_width": 1920,
  "origin_height": 1080,
  "faces": { "n_frames": 606, "score": [], "bbox": [], "landmarks": [] }
}

Reconstructing media

Annotations alone are not playable media. To obtain aligned clips:

Download the YouTube video whose id equals the directory name.
Cut [start, end).
Standardize to 25 fps, annotation resolution, 24 kHz mono audio.
(Optional) Run vocal separation for cleaner speech tracks.

Recommended tooling is provided in av_pipeline (pipeline.py). You will need yt-dlp, ffmpeg, and Deno (YouTube JS runtime) for reliable downloads.

python pipeline.py run --list-file ids.txt --limit 10

Intended uses

Audiovisual speech / lip-sync research
ASR and speech quality benchmarking on in-the-wild video
Training or evaluating dubbing / talking-head models (non-commercial under CC BY-NC)

Out-of-scope / limitations

Does not include video or audio binaries
YouTube videos may be deleted, geo-blocked, or privatized over time
Transcripts and scores are automatic estimates and may contain errors
Face landmarks are provided as metadata; respect privacy and platform policies

Ethical considerations

Use only for research and non-commercial purposes consistent with CC BY-NC 4.0 and YouTube Terms of Service. Do not attempt to re-identify private individuals beyond what is already public on YouTube.

Citation

If you use this dataset, please cite the Hub repository:

@misc{voxdub2026,
  title        = {VoxDub: Audiovisual Speech Segment Annotations},
  author       = {zyk21},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/zyk21/VoxDub}},
}

License

Annotations are released under CC BY-NC 4.0. Source media remains owned by the original uploaders and subject to YouTube ToS.