11hu83/audioset20k

AudioSet-20K: raw video in WebDataset shards

The balanced-train and eval splits of AudioSet ("AudioSet-20K"), as the clips downloaded from YouTube: untouched mp4 files at their original resolution and codecs, with no re-encoding, resizing, frame extraction or audio resampling.

What that means in practice (measured over all 33,949 files):

• Video codec: h264 31,056 · av1 1,662 · vp9 1,231. Resolution varies per clip (from 120 px up to 3840 px high).
• Audio: AAC, Opus or E-AC-3; 48, 44.1 or 32 kHz; mono, stereo or 5.1. Three clips have no audio track (NlBQ4kLpdzk_0260000, d8WgfWSf1VM_0280000 in train; chc33hy6iIU_0510000 in eval).
• Duration: median 10.03 s, but not always 10 s: 2,463 files are shorter than 9.9 s (down to 1.0 s) and 3,122 are longer than 10.1 s (up to 63.7 s). The AudioSet label refers to the segment [start_s, end_s] of the YouTube video, given per clip in clips.csv and the clip's .json; duration_s is the length of the stored file.

split	listed upstream	present here	missing (gone from YouTube)
train (balanced_train_segments.csv)	22,160	17,726	4,434
eval (eval_segments.csv)	20,371	16,223	4,148

62 shards, ~62 GB, 527 classes (AudioSet ontology mids).

Layout

audioset20k-train-000000.tar ... audioset20k-train-0000NN.tar
audioset20k-eval-000000.tar  ... audioset20k-eval-0000NN.tar
    <key>.mp4     the original clip bytes
    <key>.json    {"key", "split", "ytid", "start_s", "end_s", "labels", "duration_s", "width", "height", "vcodec", "bytes"}
index.json        {"version": 1, "shards": [...], "members": {"<split>/<key>.mp4": [shard_index, byte_offset, size]}}
clips.csv         one row per clip: key, split, ytid, start_s, end_s, labels, bytes, duration_s, width, height, vcodec, shard
MISSING.csv       segments listed upstream but not available
SHA256SUMS        sha256 of every file above

<key> is <ytid>_<start in ms, 7 digits>, e.g. --PJHxphWEs_0030000; it contains no dot, so stock WebDataset readers group .mp4 and .json correctly. labels is the comma-separated list of AudioSet mids from the official segment CSVs. index.json gives byte offsets, so a single clip can be read with one seek without unpacking a shard.

Use

Download and verify:

hf download 11hu83/audioset20k --repo-type dataset --local-dir audioset20k-wds-v1
cd audioset20k-wds-v1 && sha256sum -c SHA256SUMS

With yavih this is the audioset20k benchmark; its setup.md points here.

Stream with WebDataset:

import webdataset as wds
ds = wds.WebDataset("audioset20k-train-{000000..000031}.tar").to_tuple("mp4", "json")

Provenance

Clips were downloaded from YouTube by the dataset maintainers in September–October 2026 from the official AudioSet segment lists and packed deterministically (sorted by key, tar mtimes 0); re-packing the same download yields byte-identical shards.

Citation

@inproceedings{gemmeke2017audioset,
  title     = {Audio Set: An ontology and human-labeled dataset for audio events},
  author    = {Gemmeke, Jort F. and Ellis, Daniel P. W. and Freedman, Dylan and Jansen, Aren and
               Lawrence, Wade and Moore, R. Channing and Plakal, Manoj and Ritter, Marvin},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  year      = {2017}
}

Licence and takedown

The AudioSet annotations are CC BY 4.0. The media are third-party YouTube content, redistributed here only for non-commercial research reproducibility; rights remain with the uploaders. If you are a rights holder and want a clip removed, open a discussion on this repository with the clip key.