AudioSet

Google's AudioSet with the audio: 1,780,876 of its 2,084,320 labelled 10 s YouTube segments as 48 kHz FLAC, each with a record of where its audio came from, measured quality, duplicate and eval-overlap flags, and a reason for every segment that could not be found.

clips with audio	hours	classes	format	size
1,780,876 of 2,084,320 (85.4%)	4,904	527	48 kHz, 24-bit FLAC	2.25 TiB

Two versions of every clip: Opus and AAC

YouTube stores the audio of each video as two separate streams, and this repository ships both:

YouTube stream	format	typical quality	where it is here	clips
Opus (itag 251)	48 kHz, variable bitrate (median about 120 kbps)	full band, up to 20 kHz	audio48k/ (shard, member), native_48k = true	1,564,753
AAC (itag 140)	44.1 kHz, about 130 kbps constant	usually low-passed near 16 kHz	audio_v1/ (aac_shard, aac_member), as downloaded	1,347,236

• audio48k/ is the release to train on. Each clip is the Opus version wherever one could be fetched and verified. The other 216,123 clips, mostly older uploads that have no Opus stream at all, are the AAC audio resampled to 48 kHz and flagged native_48k = false.
• audio_v1/ keeps the AAC originals at 44.1 kHz, untouched. 1,150,863 clips have both an Opus and an AAC version, so the two codecs can be compared clip by clip; for the 196,373 AAC-only clips it holds the original before resampling.

# clips = metadata/clips.parquet, loaded as in the quick start below
row = clips[clips.available & clips.aac_shard.notna()].iloc[0]
opus = (row.shard, row.member)          # 48 kHz, from the Opus stream
aac = (row.aac_shard, row.aac_member)   # 44.1 kHz, the AAC stream of the same clip

Highlights

• 48 kHz throughout. 87.9% of the clips are decoded from Opus streams (full band to 20 kHz). The rest exist only as AAC on YouTube; they are resampled to 48 kHz and flagged, never passed off as native.
• Provenance on every clip. Where the audio came from, which codec it was decoded from, and whether it was resampled.
• Measured. Bandwidth, energy above 16 kHz, peak and RMS level, clipping and silence for every clip.
• Deduplicated. Audio fingerprinting finds re-uploads and flags train clips whose audio also appears in the eval set.
• Accounted for. Each of the 303,444 missing segments carries one specific reason (removed, private, blocked, ...).
• Both YouTube versions. Opus in audio48k/, the AAC originals in audio_v1/, linked per clip.

Contents

• Two versions of every clip: Opus and AAC
• Quick start
• How it compares
• Dataset structure
• How the audio was built
• Quality measurements
• Duplicates and eval overlap
• Missing segments
• Limitations
• Licence and terms
• Citation

Quick start

Access is gated: accept the terms on this page, then log in with huggingface-cli login.

Browse and filter the metadata

import pandas as pd
from huggingface_hub import hf_hub_download

clips = pd.read_parquet(hf_hub_download("Muno459/AudioSet", "metadata/clips.parquet", repo_type="dataset"))
audio = clips[clips.available]

# a clean training pool: no overlap with eval, native 48 kHz only, one clip per group of identical audio
train = audio[(audio.split != "eval") & ~audio.eval_overlap & audio.native_48k]
train = train[train.dup_group.isna() | ~train.duplicated("dup_group")]

Read a clip

import io, tarfile, soundfile as sf

row = train.iloc[0]
with tarfile.open(hf_hub_download("Muno459/AudioSet", row.shard, repo_type="dataset")) as tar:
    wav, sr = sf.read(io.BytesIO(tar.extractfile(row.member).read()))   # sr == 48000

Stream the shards with WebDataset

import webdataset as wds
from huggingface_hub import get_token, hf_hub_url

auth = f"-H 'Authorization:Bearer {get_token()}'"
urls = [f"pipe:curl -s -L {auth} {hf_hub_url('Muno459/AudioSet', shard, repo_type='dataset')}" for shard in audio.shard.unique()]
dataset = wds.WebDataset(urls, shardshuffle=True).decode(wds.torch_audio)   # key "<ytid>_<start_ms>", audio under "flac"

Listen in the viewer, or load the preview with datasets

from datasets import load_dataset
preview = load_dataset("Muno459/AudioSet", "preview", split="balanced_train")   # 5,000 clips with embedded audio and label names

How it compares

agkphysics/AudioSet is the most complete earlier release, and this one builds on it (it is the source of 1,284,135 of the clips here).

	this dataset	agkphysics/AudioSet
segments with audio	1,780,876	1,774,481
sample rate	48 kHz for every clip	as downloaded; 44.1 kHz AAC for at least 214,890 clips
source and codec per clip	recorded (source, source_codec, native_48k)	not recorded
AAC-only clips	resampled and flagged (216,123)	mixed in, not distinguished
quality measurements	bandwidth, level, clipping, silence	none
digitally silent clips	removed	1,185 (replaced here from YouTube where possible)
missing segments	one specific reason each	not listed
duplicates and eval overlap	2,609 train clips overlap eval, flagged	not flagged

Dataset structure

Files

path	contents
audio48k/<split>/<split>-NNNN.tar	this release: one FLAC per clip, members <ytid>_<start_ms>.flac, about 4 GB per shard
audio_v1/<split>/<split>-NNNN.tar	the first 2026 download as fetched, mostly 44.1 kHz AAC (see aac_shard, aac_member)
metadata/clips.parquet	one row per AudioSet segment, all 2,084,320
metadata/duplicates.parquet	every verified pair of clips with the same or partly the same audio
metadata/*_segments.csv, class_labels_indices.csv, ontology.json, qa_true_counts.csv	Google's release files, unchanged
subsets/noise_bank.parquet	clips labelled with neither speech nor music, with a coarse category; a clip can carry several
preview/*.parquet	5,000 balanced-split clips with embedded audio, label names and provenance, for the viewer

Splits

split	segments	with audio	share	shards
balanced_train	22,160	18,765	84.7%	7
unbalanced_train	2,041,789	1,744,882	85.5%	566
eval	20,371	17,229	84.6%	6

metadata/clips.parquet

column	description
ytid, split, start_s, end_s	the AudioSet segment
labels	ontology ids (/m/...), names in class_labels_indices.csv
available	the clip has audio in this release
shard, member, bytes	tar shard, member name and FLAC size
aac_shard, aac_member	the same clip as first downloaded (44.1 kHz AAC), where that differs from this release
source	youtube_2026 (fetched from YouTube, August to September 2026) or agkphysics_2023 (from agkphysics/AudioSet)
source_codec	opus or aac: the stream the audio was decoded from
native_48k	false if resampled from 44.1 kHz AAC (no content above the AAC encoder's cutoff, typically about 16 kHz)
sample_rate, channels, duration_s	as stored; sample rate is always 48000
peak_dbfs, rms_dbfs	sample peak and RMS level over all channels
clipped_fraction	share of samples at or above 0.999 of full scale
silent_fraction	share of 50 ms frames of the mono mix below -60 dBFS
bandwidth_hz	highest frequency whose smoothed level is within 50 dB of the 90th percentile level between 200 Hz and 4 kHz
hf16_energy_db	energy above 16 kHz relative to the total, in dB
dup_group, dup_group_size	near-identical clips (the same audio at the same position, within 1 s) share a group id
same_audio_in, partial_audio_in	splits of the clips this clip matches over the whole overlap, or over at least 3 s of it
eval_overlap	a train clip that matches an eval clip, or an eval clip that matches a train clip
reason, reason_group, reason_checked	for missing segments: why, the group it counts under, and the date it was last checked

How the audio was built

Google publishes AudioSet as YouTube ids with time stamps; the audio has to be fetched per segment. It was fetched in August and September 2026 and then rebuilt clip by clip, preferring Opus (48 kHz, full band to 20 kHz) over AAC (44.1 kHz, typically low-passed near 16 kHz), since a paired comparison on the same segments found the Opus stream never measurably worse:

1. Opus copy from agkphysics/AudioSet, accepted only if it is the same recording as the 2026 download (normalised cross-correlation of at least 0.5 after alignment), not silent, not shorter, and not narrower below 20 kHz.
2. Opus stream from YouTube, fetched again under the same checks when agkphysics had no usable copy. Noisy recordings lose waveform shape through two different codecs, so a correlation between 0.3 and 0.5 is also accepted when the audio fingerprints match (bit error rate at most 0.20 at the aligned offset).
3. Resampled AAC, only when no Opus version exists (older uploads often have none) or the Opus copy failed the checks: polyphase resampling from 44.1 to 48 kHz (Kaiser window, beta 10), flagged native_48k = false.

Segments YouTube no longer serves were filled from agkphysics/AudioSet under the same checks. Every file was decoded back, its sample rate and duration verified, and rejected if its peak was at or below -70 dBFS. Nothing is normalised, and the channel layout is as served.

audio	source	source_codec	native_48k	clips	share
Opus copy from agkphysics/AudioSet (2023)	agkphysics_2023	opus	true	1,268,580	71.2%
Opus stream from YouTube (2026)	youtube_2026	opus	true	296,173	16.6%
AAC stream from YouTube (2026), resampled	youtube_2026	aac	false	200,568	11.3%
AAC copy from agkphysics/AudioSet (2023), resampled	agkphysics_2023	aac	false	15,555	0.9%

Quality measurements

Every clip was measured from the final file (definitions in the column table above).

	value
clips with content up to at least 19 kHz	26.0%
median bandwidth_hz, native 48 kHz clips	15,574 Hz
median bandwidth_hz, resampled AAC clips	14,484 Hz
clips with more than 0.1% clipped samples	116,646

Many recordings are narrower than their codec allows (phone and camera microphones, old uploads); bandwidth_hz shows this directly, so band-limited audio can be filtered out or used on purpose.

Duplicates and eval overlap

Every clip was fingerprinted (32-bit sub-fingerprints from band-energy differences of the 8 kHz mono mix, 32 ms hop, after Haitsma and Kalker) and looked up against every other clip at any time offset. A pair is the same audio when the bit error rate over the whole overlap (at least 3 s of active audio) is at most 0.25, and partly shared when some 3 s stretch matches at 0.12 or less; unrelated audio sits near 0.5. In an audit of 100 random accepted pairs across the full dataset, 75 were confirmed as the same recording by local waveform correlation (19 of 20 below a bit error rate of 0.05); most of the rest are the same music with tempo, mix or phase differences that waveform correlation cannot follow. Treat matches near the 0.25 limit as likely rather than certain: metadata/duplicates.parquet carries the error rate.

	clips
clips in a group of near-identical clips	53,772 in 19,425 groups
extra copies (dropped by keeping one per group)	34,347
train clips with the same audio as an eval clip	2,425
train clips sharing at least 3 s with an eval clip (eval_overlap)	2,609
eval clips sharing at least 3 s with a train clip (eval_overlap)	922

The overlaps come from YouTube itself: one video uploaded under several ids, compilations that reuse clips, popular music reused across many videos, and channel intros, each labelled by AudioSet as an independent segment.

Groups (dup_group) hold near-identical clips: every member matches the group's representative clip directly (bit error rate at most 0.20, offset within 1 s, at least 8 s of overlap), so songs reused at different offsets across many videos cannot chain unrelated clips into one group. same_audio_in, partial_audio_in and eval_overlap use every verified pair. metadata/duplicates.parquet lists every pair with its bit error rate, offset and overlap.

Missing segments

303,415 of the 303,444 missing segments were retried from YouTube between 2026-09-14 and 2026-09-15; each missing segment carries the single most specific reason YouTube gave.

reason group	segments	share of AudioSet
removed	206,398	9.9%
private or restricted	88,600	4.3%
blocked by region or claim	6,457	0.3%
stream refused	1,024	<0.1%
no usable audio	936	<0.1%
other	29	<0.1%

All reasons

group	reason	segments
removed	video unavailable, no reason given (removed)	136,969
removed	uploader account terminated	53,338
removed	removed for a Terms of Service violation	6,676
removed	removed after a copyright claim	4,997
removed	removed for a Community Guidelines violation	2,794
removed	removed under YouTube's policy on violent or graphic content	853
removed	removed under YouTube's policy on nudity or sexual content	274
removed	removed under YouTube's policy on harassment and bullying	163
removed	removed by the uploader	153
removed	removed under YouTube's policy on spam, deceptive practices, and scams	66
removed	removed after a privacy claim	65
removed	removed as a duplicate of another video	40
removed	removed after a trademark claim	10
private or restricted	private video	87,794
private or restricted	video not available, no reason given	561
private or restricted	age-restricted, sign-in required	134
private or restricted	sign-in required	70
private or restricted	channel members only	41
blocked by region or claim	blocked by a content claim	5,415
blocked by region or claim	rights holder blocked the fetch region	788
blocked by region or claim	not available in the fetch region	175
blocked by region or claim	uploader blocked the fetch region	79
stream refused	audio stream refused by the CDN (HTTP 403)	530
stream refused	stream served only up to its first MiB, segment lies past it	486
stream refused	temporarily unavailable when fetched	7
stream refused	audio fetch failed	1
no usable audio	audio track is digital silence	572
no usable audio	the labelled 10 s window is digital silence	240
no usable audio	segment starts after the end of the video	124
other	unplayable, no reason given	20
other	YouTube error, no reason given	9

Limitations

• Labels are Google's. They are clip-level and weakly verified; qa_true_counts.csv is Google's own estimate of label quality per class.
• Resampled clips are not full band. The 216,123 clips with native_48k = false are 48 kHz files carrying AAC audio, with nothing above the AAC cutoff. Filter on native_48k when that matters.
• The audio is today's YouTube audio, checked against our own 2026 download, not against what Google's raters heard in
  2017. Uploaders can re-edit videos; a small number of segments may no longer match their labels.
• Duplicate detection is a threshold. An audit confirmed 75 of 100 random matches as the same recording by waveform correlation, with the rest mostly music the check cannot follow; the pair table carries the bit error rate so stricter cut-offs are one filter away. Eval-overlap flags err on the side of excluding a clip.

Licence and terms

The labels and the ontology are © Google, released under CC BY 4.0. The audio is YouTube content owned by its uploaders and is not covered by that licence. It is shared for non-commercial research under gated access, and requesters agree not to redistribute it. Rights holders who want a clip removed can open a discussion on this repository with the YouTube id.

Citation

If you use this dataset, please cite AudioSet:

@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 = {2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages     = {776--780},
  year      = {2017},
  doi       = {10.1109/ICASSP.2017.7952261}
}

and this release:

@misc{audioset48k2026,
  title        = {AudioSet at 48 kHz: audio, provenance, quality measurements and duplicate flags for the AudioSet segments},
  author       = {Muno459},
  year         = {2026},
  howpublished = {Hugging Face},
  url          = {https://huggingface.co/datasets/Muno459/AudioSet}
}

The duplicate search follows:

@inproceedings{haitsma2002robust,
  title     = {A Highly Robust Audio Fingerprinting System},
  author    = {Haitsma, Jaap and Kalker, Ton},
  booktitle = {Proceedings of the 3rd International Conference on Music Information Retrieval (ISMIR)},
  year      = {2002}
}

Acknowledgements

AudioSet was created by the Sound Understanding group at Google Research (research.google.com/audioset). Clips with source = agkphysics_2023 come from agkphysics/AudioSet, without which many segments whose videos have since disappeared would be lost.