The audio configs carry each clip's own licence and credit line; clips of the noncommercial config may be used for non-commercial purposes only. The links config holds links, timestamps and machine-made annotations, no audio: the linked media belong to their owners and remain under their own terms, and you are responsible for complying with them when you fetch any media.

LAION-Audio-XL

LAION-Audio-XL is the corpus of 366,007,518 audio clips of at most 30 s with captions that the LAION CLAP2 models were trained on. This repository holds 4,364,123 of them as audio, in three configs by licence class, and the other 361,620,512 as links, with the same captions and analysis fields; 462 source datasets in all.

repository	content	clips	access
laion/laion-audio-xl	audio (configs open, sharealike, noncommercial) and links (config links)	365,984,635	automatic approval
laion/laion-audio-xl-embeddings	CLAP2 audio and caption embeddings (models L, M, S, XS, XXS)	364,985,867	automatic approval

Paper and models

LAION-Audio-XL is released with the paper LAION-Audio-XL: A Web-Scale Corpus and Scaling Studies for Contrastive Language-Audio Pretraining (arXiv link follows shortly). On it we trained the LAION CLAP2 models, five sizes from XXS to L, each on 1B samples (about 2.8 passes over the corpus): laion/clap2-xxs-audioxl-s1B, laion/clap2-xs-audioxl-s1B, laion/clap2-small-audioxl-s1B, laion/clap2-medium-audioxl-s1B, laion/clap2-large-audioxl-s1B. They load with OpenCLIP (naflexclap_<size>, pretrained tag laion_audio_xl_s1b); their audio and caption embeddings of every clip are in laion/laion-audio-xl-embeddings.

Loading

from datasets import load_dataset, Audio
ds = load_dataset("laion/laion-audio-xl", "open", streaming=True)    # or sharealike / noncommercial / links

config	content	clips	hours
`open`	audio whose licences allow commercial use: CC0, public domain, CC BY, permissive licences (MIT, Apache-2.0, BSD, ODC-By) and CC BY-ND (verbatim excerpts)	2,101,816	3,343
`sharealike`	audio under share-alike licences (CC BY-SA, GPL, ODbL); commercial use allowed, adaptations under the same terms	990,829	1,770
`noncommercial`	audio of academic datasets whose licences allow non-commercial use only (CC BY-NC and similar)	1,271,478	3,952
`links`	where to fetch the audio of every other clip (our YouTube and web crawls, YouTube-based datasets, datasets that forbid re-hosting, closed-model audio), with the same captions and analysis fields; no audio	361,620,512	1,380,150

Each config has one split. In the audio configs, column audio holds the clip as FLAC (32 kHz, mono, 16 bit), typed as a datasets Audio feature; .cast_column("audio", Audio(decode=False)) gives the FLAC bytes. A field that does not apply to a clip is null. audio_id is the clip key everywhere. The rows of one media item in links are in the same file, sorted by start.

Columns

`open`, `sharealike`, `noncommercial`

Identifiers

column	type	meaning
`audio`	audio (FLAC bytes)	The clip as FLAC: 32 kHz, mono, 16 bit, at most 120 s. `datasets` decodes it on access; cast the column to `Audio(decode=False)` for the raw FLAC bytes.
`audio_id`	integer	Our stable numeric clip id, unique across all configs.
`source_dataset`	text	The source dataset the clip comes from, for example `CSS10` or `nsynth`, or one of our own crawls: `YouTube (bulk)`, `Common Crawl (bulk)`, `BVD (bulk)`. The credit for each source is a row in `sources.csv`.

Captions and text

column	type	meaning
`text`	text	The main description of the sound. Either a caption written by a captioning model (then `captioning_model` is set) or the source dataset's own caption or label sentence.
`long_caption`	text	A detailed caption of several sentences, written by an audio-language model. Only on clips we captioned ourselves.
`short_caption`	text	A one-sentence caption of the same clip.
`alt_text`	text	A second text taken from the source's metadata or built from a template, for example a class label, `Audio recording`, or for speech `Speech in Bengali: <transcript>`.
`captioning_model`	text	Exact Hugging Face id of the model that wrote the captions, for example `Qwen/Qwen3-Omni-30B-A3B-Instruct` or `OpenMOSS-Team/MOSS-Audio-8B-Thinking`. `null` when the text did not come from a captioning model.
`text_type`	text	Which caption `text` holds: `long` or `short`.

Licence and attribution

column	type	meaning
`license`	text	SPDX id of the licence (a short standard code such as `CC-BY-4.0`, `CC0-1.0`, `CC-BY-SA-3.0`, `CC-BY-NC-SA-3.0`); `LicenseRef-…` for licences with no SPDX id. Audio configs: the licence of this clip. Links config: the licence of the linked media as known for the source; `LicenseRef-see-source` means the media owner's own terms apply, `mixed` means the items of that source carry different licences.
`license_url`	text	Link to the legal text of that licence.
`attribution`	text	A ready-to-print credit line: title, authors, year, link and licence. For sources that publish a licence per recording (Freesound, xeno-canto, Jamendo, FMA, …) it names the recording and its author; otherwise it credits the source dataset. Licences that demand a copyright notice carry it here.

Signal statistics

column	type	meaning
`duration`	number	Length of the clip in seconds (at most 120 in the audio configs).
`rms_db`	number	Average loudness of the whole clip in decibels relative to digital full scale (dBFS), so always negative; closer to 0 is louder.
`mean_rms_db`	number	Loudness averaged over short 25 ms frames. Much lower than `rms_db` when the clip has long quiet stretches.
`peak_db`	number	Level of the loudest single sample in dBFS. 0 means the clip touches full scale.
`dynamic_range_db`	number	`peak_db` minus `rms_db`: how far the peaks stand out above the average level.
`dc_offset`	number	Absolute mean sample value (0 to 1). 0 means the waveform is centred; larger means a constant bias.
`silence_pct`	number	Share of 25 ms frames quieter than −60 dBFS. Despite the name it is a fraction from 0 to 1, not a percentage.
`clip_frac`	number	Share of samples at or above 0.99 of full scale (clipping), a fraction from 0 to 1.
`bandwidth_hz`	number	Frequency below which 95 % of the clip's spectral energy lies. Low values mean dull or band-limited audio (telephone, low-bitrate MP3).
`spectral_centroid_hz`	number	The centre of mass of the spectrum in Hz; higher sounds brighter.

Model predictions

column	type	meaning
`tax_l1`	text	Top-level class of our sound taxonomy, for example `Speech`, `Music`, `Animal - bioacoustics`. Predicted by our classifier, not a human label. Class names contain a hyphen padded with two spaces on each side; match them exactly.
`tax_l1_prob`	number	The classifier's confidence in `tax_l1` (0 to 1).
`tax_l1_top3`	list of {label, prob} objects	The (up to) three most likely top-level classes with their probabilities, most likely first.
`tax_l3`	text	Most specific class, as a path `level 1/level 2/level 3`, for example `Music/Electronic music/Atmospheric electronic - dream pop`.
`tax_l3_prob`	number	The classifier's confidence in `tax_l3` (0 to 1).
`pred_age`	number	Estimated speaker age in years. Speech clips only. Model: audeering wav2vec2 age/gender model.
`pred_gender`	text	Estimated speaker class: `female`, `male` or `child`. Speech clips only.
`lang_id`	text	Spoken language as a three-letter ISO 639-3 code (`spa`, `fra`, `ben`, `cmn`, …), predicted by Meta's MMS-LID model. Speech clips only.
`lang_prob`	number	Confidence in `lang_id` (0 to 1).
`pitch_hz`	number	Median pitch of the voice in Hz from a pitch tracker; 0 means no voiced pitch was found. Measured, not a class prediction, so it has no `pred_` prefix. Speech clips only.
`pred_emotion`	text	Emotion label from emotion2vec+ large: `angry`, `disgusted`, `fearful`, `happy`, `neutral`, `other`, `sad`, `surprised` or `unknown`. Speech clips only.
`maest_tags`	list of text	The five most likely music styles from the MAEST model, as Discogs `Genre---Style` labels such as `Electronic---Ambient`. Music clips only.
`maest_probs`	list of numbers	Probabilities of those five tags, in the same order.
`pred_tempo_bpm`	number	Estimated tempo in beats per minute (librosa). Music clips only.
`pred_music_key`	text	Estimated musical key, for example `D minor` (librosa). Music clips only.
`perch_top_species_id`	list of text	The five species that Google's Perch bioacoustics model rates most likely, as Perch class labels. Animal-sound clips only.

Source links

column	type	meaning
`start_time`	number	Seconds in the source recording where the clip starts, for sources that cut their clips from longer recordings (meetings, parliament sessions, talks, videos).
`end_time`	number	Seconds in the source recording where the clip ends (same sources as `start_time`).
`original_id`	text	The source dataset's own id of the recording, where it publishes one.
`original_duration`	number	Length in seconds of the source file or video the clip was cut from.

`links`

Identifiers

column	type	meaning
`audio_id`	integer	Our stable numeric clip id, unique across all configs.
`source_dataset`	text	The source dataset the clip comes from, for example `CSS10` or `nsynth`, or one of our own crawls: `YouTube (bulk)`, `Common Crawl (bulk)`, `BVD (bulk)`. The credit for each source is a row in `sources.csv`.

Captions and text

column	type	meaning
`text`	text	The main description of the sound. Either a caption written by a captioning model (then `captioning_model` is set) or the source dataset's own caption or label sentence.
`long_caption`	text	A detailed caption of several sentences, written by an audio-language model. Only on clips we captioned ourselves.
`short_caption`	text	A one-sentence caption of the same clip.
`alt_text`	text	A second text taken from the source's metadata or built from a template, for example a class label, `Audio recording`, or for speech `Speech in Bengali: <transcript>`.
`captioning_model`	text	Exact Hugging Face id of the model that wrote the captions, for example `Qwen/Qwen3-Omni-30B-A3B-Instruct` or `OpenMOSS-Team/MOSS-Audio-8B-Thinking`. `null` when the text did not come from a captioning model.
`text_type`	text	Which caption `text` holds: `long` or `short`.

Licence and attribution

column	type	meaning
`license`	text	SPDX id of the licence (a short standard code such as `CC-BY-4.0`, `CC0-1.0`, `CC-BY-SA-3.0`, `CC-BY-NC-SA-3.0`); `LicenseRef-…` for licences with no SPDX id. Audio configs: the licence of this clip. Links config: the licence of the linked media as known for the source; `LicenseRef-see-source` means the media owner's own terms apply, `mixed` means the items of that source carry different licences.
`license_url`	text	Link to the legal text of that licence.
`attribution`	text	A ready-to-print credit line: title, authors, year, link and licence. For sources that publish a licence per recording (Freesound, xeno-canto, Jamendo, FMA, …) it names the recording and its author; otherwise it credits the source dataset. Licences that demand a copyright notice carry it here.

Signal statistics

column	type	meaning
`duration`	number	Length of the clip in seconds (at most 120 in the audio configs).
`rms_db`	number	Average loudness of the whole clip in decibels relative to digital full scale (dBFS), so always negative; closer to 0 is louder.
`mean_rms_db`	number	Loudness averaged over short 25 ms frames. Much lower than `rms_db` when the clip has long quiet stretches.
`peak_db`	number	Level of the loudest single sample in dBFS. 0 means the clip touches full scale.
`dynamic_range_db`	number	`peak_db` minus `rms_db`: how far the peaks stand out above the average level.
`dc_offset`	number	Absolute mean sample value (0 to 1). 0 means the waveform is centred; larger means a constant bias.
`silence_pct`	number	Share of 25 ms frames quieter than −60 dBFS. Despite the name it is a fraction from 0 to 1, not a percentage.
`clip_frac`	number	Share of samples at or above 0.99 of full scale (clipping), a fraction from 0 to 1.
`bandwidth_hz`	number	Frequency below which 95 % of the clip's spectral energy lies. Low values mean dull or band-limited audio (telephone, low-bitrate MP3).
`spectral_centroid_hz`	number	The centre of mass of the spectrum in Hz; higher sounds brighter.

Model predictions

column	type	meaning
`tax_l1`	text	Top-level class of our sound taxonomy, for example `Speech`, `Music`, `Animal - bioacoustics`. Predicted by our classifier, not a human label. Class names contain a hyphen padded with two spaces on each side; match them exactly.
`tax_l1_prob`	number	The classifier's confidence in `tax_l1` (0 to 1).
`tax_l1_top3`	list of {label, prob} objects	The (up to) three most likely top-level classes with their probabilities, most likely first.
`tax_l3`	text	Most specific class, as a path `level 1/level 2/level 3`, for example `Music/Electronic music/Atmospheric electronic - dream pop`.
`tax_l3_prob`	number	The classifier's confidence in `tax_l3` (0 to 1).
`pred_age`	number	Estimated speaker age in years. Speech clips only. Model: audeering wav2vec2 age/gender model.
`pred_gender`	text	Estimated speaker class: `female`, `male` or `child`. Speech clips only.
`lang_id`	text	Spoken language as a three-letter ISO 639-3 code (`spa`, `fra`, `ben`, `cmn`, …), predicted by Meta's MMS-LID model. Speech clips only.
`lang_prob`	number	Confidence in `lang_id` (0 to 1).
`pitch_hz`	number	Median pitch of the voice in Hz from a pitch tracker; 0 means no voiced pitch was found. Measured, not a class prediction, so it has no `pred_` prefix. Speech clips only.
`pred_emotion`	text	Emotion label from emotion2vec+ large: `angry`, `disgusted`, `fearful`, `happy`, `neutral`, `other`, `sad`, `surprised` or `unknown`. Speech clips only.
`maest_tags`	list of text	The five most likely music styles from the MAEST model, as Discogs `Genre---Style` labels such as `Electronic---Ambient`. Music clips only.
`maest_probs`	list of numbers	Probabilities of those five tags, in the same order.
`pred_tempo_bpm`	number	Estimated tempo in beats per minute (librosa). Music clips only.
`pred_music_key`	text	Estimated musical key, for example `D minor` (librosa). Music clips only.
`perch_top_species_id`	list of text	The five species that Google's Perch bioacoustics model rates most likely, as Perch class labels. Animal-sound clips only.

Source links

column	type	meaning
`youtube_id`	text	YouTube video id of the media, set for every row whose media is on YouTube, otherwise null.
`page_url`	text	The web page on which the audio was found (Common Crawl).
`warc_file`	text	The Common Crawl archive file (WARC) that holds the crawled page.
`source_type`	text	How the audio was found on the page: `audio_tag` (an HTML audio element) or `a_link` (a plain link).
`scene_index`	integer	Number of the scene or segment inside the source video (BVD).
`original_duration`	number	Length in seconds of the source file or video the clip was cut from.
`media_id`	text	Identifier of the media item the clip is cut from: YouTube video id, web media URL, or file name inside a release (see the table below). Clips of one item share it.
`url`	text	Where to fetch that item: YouTube watch URL, the media URL itself, or the landing page of a release that is only distributed as a whole.
`start`	number	Where the clip starts inside the media item, in seconds.
`end`	number	Where the clip ends inside the media item, in seconds.

Licences and attribution

Each audio clip's licence is its source dataset's licence, or the clip's own licence where the source publishes one per recording (Freesound, xeno-canto, FMA, ...).
attribution is a ready-to-print credit line (title, author, source, licence). For per-recording sources it names the recording and its author; otherwise it credits the source dataset. Licences that require a copyright notice (e.g. the Tampere University licence) carry it in this line.
We cut clips from their source recordings and re-encode them; no other changes to the audio.
In links, license and attribution describe the linked media as far as we know them; the media stay under their owners' terms.
sources.csv: one row per source dataset with its credit, licence and number of clips here.
licences.csv: number of clips per licence in each audio config.

Downloads last month: 138

Total file size:

940 GB

Collection including laion/laion-audio-xl

LAION-AUDIO-XL & CLAP2

Collection

Release of the LAION-Audio-XL dataset and the CLAP2 model achitectures • 7 items • Updated about 22 hours ago • 2

You need to agree to share your Hugging Face information to access this dataset

LAION-Audio-XL

Paper and models

Loading

Columns

open, sharealike, noncommercial

Identifiers

Captions and text

Licence and attribution

Signal statistics

Model predictions

Source links

links

Identifiers

Captions and text

Licence and attribution

Signal statistics

Model predictions

Source links

Licences and attribution

Collection including laion/laion-audio-xl

`open`, `sharealike`, `noncommercial`

`links`