AudioSet-Strong-10

A ten-class, strong-label, 10-second-clip subset of AudioSet-Strong, with clip-relative event onsets and offsets. Built for streaming / bounded-latency sound event detection, where you need per-event times rather than clip tags.

Classes

Class	AudioSet MID
Alarm	/m/07pp_mv
Bark	/m/05tny_
Cough	/m/01b_21
Explosion	/m/014zdl
Gunshot	/m/032s66
Laughter	/m/01j3sz
Screaming	/m/03qc9zr
Vehicle horn	/m/0912c9
Whispering	/m/02rtxlg
Whistling	/m/01w250

The official strong annotations contain 15,052 train and 2,217 eval unique 10-second clips for these ten classes.

Coverage in this release

Split	Official clips	In this release	Coverage
train	15,052	14,521	96.5%
eval	2,217	2,161	97.5%
total	17,269	16,682	96.6%

37,832 clip-relative event rows in total. Every clip has at least one event. missing.tsv lists the 404 segment ids that could not be retrieved from YouTube or from the mirror (the underlying videos are deleted).

Event rows per class (train / eval):

Class	train	eval
Alarm	1,509	351
Bark	6,014	541
Cough	2,032	368
Explosion	1,095	186
Gunshot	3,396	353
Laughter	11,250	1,918
Screaming	1,326	270
Vehicle horn	1,832	297
Whispering	1,581	308
Whistling	2,804	401

Contents

This repository releases annotations and identifiers, not audio:

train_events.csv                    path,onset,offset,label
eval_events.csv
missing.tsv                         official clips that were unavailable
short_windows.tsv                   clips whose source audio ends early
oversize_windows.tsv                clips kept at full length

Event rows use paths of the form audio/<split>/<bucket>/<ytid>_<start_ms>.wav, where <bucket> is the first two hex digits of a CRC32 of the file stem. That is the layout the audio takes in the companion repository, so the tables apply directly once the audio is mounted there.

Audio

The audio lives in a companion repository, LakoreAI/AudioSet-Strong-10-audio (16,682 clips, same layout), released for research use subject to the rights of the original uploaders. The annotation tables here apply to it unchanged.

<ytid>_<start_ms> is the official AudioSet segment_id: the YouTube id and the start of the 10-second window in milliseconds. onset/offset are seconds inside the clip (0-10), so a row is directly usable as a strong label.

One clip can have several rows (several labelled events). Every clip has at least one event from the ten classes above; other AudioSet classes that occur in the same window are not annotated here.

Provenance and completeness

1. Segment ids, labels and event times come from the official audioset_train_strong.tsv / audioset_eval_strong.tsv (storage.googleapis.com/us_audioset), filtered to the ten MIDs.
2. Audio is the 10-second window containing the labelled event, downloaded from YouTube with yt-dlp and resampled to 16 kHz mono.
3. Clips whose YouTube video is deleted are recovered from the public CLAPv2/audioset_strong mirror (same official windows). Annotation times still come from the official TSVs; the mirror only supplies audio bytes. Harvested clips were cross-checked against the mirror's own event list (zero label mismatches).

YouTube attrition is real: a share of the official videos is permanently unavailable. missing.tsv lists the segment ids that could not be obtained from either source at build time. Coverage is reported in the dataset card table below and should be re-checked before any comparison with published numbers on the complete strong set.

Licence and credit

• Annotations: CC BY 4.0, Google AudioSet (Gemmeke et al., 2017).
• Audio: downloaded from YouTube; rights remain with the original uploaders. This repository grants no rights over the audio and is released for research use, mirroring how AudioSet itself distributes the corpus. It may be removed on request.
• The ten-class selection follows the AudioSet-Strong subset used by the RED+EPN line of work (Schmid et al.).

Citation

@misc{audioset_strong_10,
  title  = {AudioSet-Strong-10: ten-class strong labels with 10-second audio},
  author = {LakoreAI},
  note   = {Subset of AudioSet-Strong (Gemmeke et al., 2017)},
  year   = {2026},
  url    = {https://huggingface.co/datasets/LakoreAI/AudioSet-Strong-10}
}

Clips whose source audio is short or over-long

183 clips in audio_oversize/ are kept exactly as downloaded: the source audio runs past the 10 s window, so they cannot enter the labelled corpus, but they are not discarded. Clips whose source audio ends slightly before the window are padded with silence to 10.00 s; their true length is listed in short_windows.tsv and the events are clipped to it.

Why clips are missing

Each missing clip was checked with a metadata request. Of the 404:

• 218 deleted or unavailable at the source
• 125 private
• 59 removed (copyright)
• 2 other errors

None is recoverable: the videos no longer exist in any accessible form, so no credential or client can fetch them. missing_reasons.tsv lists the per-clip cause.