A Web-Scale Corpus and Scaling Studies for Contrastive Language‑Audio Pretraining

LAION-Audio-XL is a corpus of 365M audio-text pairs (1.39M hours), selected from 1.48B audio samples of academic datasets, internet datasets, web collections, Common Crawl and YouTube. CLAP2 are contrastive audio-text models in five sizes (27M to 340M parameters) trained on it. This repository holds the data pipeline, the training and evaluation recipes, and the experiments.

Quickstart

Models

# pip install "open_clip_torch>=4.2.1" soundfile
import torch, open_clip
from open_clip.audio.io import load_audio

hub = "hf-hub:laion/clap2-medium-audioxl-s1B"  # sizes: xxs, xs, small, medium, large
model, preprocess = open_clip.create_model_from_pretrained(hub)
tokenizer = open_clip.get_tokenizer(hub)

audio = preprocess()(load_audio("dog.wav"))  # any sample rate and channel count
texts = ["A dog barks.", "A door slams shut.", "People are applauding."]
with torch.no_grad():
    logits, _ = model.get_logits({k: v[None] for k, v in audio.items()}, tokenizer(texts))
print({t: round(p, 3) for t, p in zip(texts, logits.softmax(-1)[0].tolist())})

Dataset

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

Clips that cannot be redistributed are listed as links in the links config of the same repository; 0_download/web_youtube.py and 0_download/web_links.py stream its rows and download and cut the clips into WebDataset tars, one folder per source dataset, ready as input IN of the pipeline.

Pipeline

pip install -e .
bash 1_preprocessing/run.sh IN WORK OUT   # IN: folder of WebDataset tar folders, OUT: shards of clips up to 30 s

run.sh reads OLAF (upstream Olaf build), EVAL_DL (benchmark download folder), QWEN_ENV and MOSS_ENV (captioning server environments). The selection target and budget (5_selection/config.yaml) and --holdout (7_keep_30s/export.py) are set for the paper's selection and must be set for your pool.

Folder	Step
0_download/raw/	rebuild the collection: sources.csv (425 sources), download handlers, converter, text rules; Common Crawl WARC extraction; YouTube
1_preprocessing/0_convert/	32 kHz 16-bit mono, long recordings cut at silences
1_preprocessing/1_filter_silent/	per-clip signal statistics, silence and duration gate
1_preprocessing/2_deduplication/	0_hash_matching (PCM MD5), 1_fingerprint (Olaf), 2_semdedup (Dasheng, SemDeDup)
1_preprocessing/3_remove_eval_data/	screen against the evaluation data of 46 datasets: identifiers, landmark and sub-band fingerprints, neural fingerprint, exhaustive search, waveform verification
1_preprocessing/4_annotation/	0_taxonomy (4,416-node taxonomy, classifier), 1_language, 2_speaker, 3_emotion, 4_music, 5_species
1_preprocessing/5_selection/	coverage-aware selection with exact-size trim
1_preprocessing/6_captioning/	short and long captions with Qwen3-Omni and MOSS-Audio
1_preprocessing/7_keep_30s/	export of the clips up to 30 s with annotations and captions
2_training/	training with OpenCLIP, warmup-stable-decay schedule
3_evaluation/	MAEB, MSEB and HEAR with CLIP_benchmark
4_experiments/	appendix studies: segmentation, taxonomy_classifier, captioner_comparison, metadata_prompt, audio_text_ratio, scaling_law, design_studies

External tools: Olaf (built separately, AGPL-3.0), 7z, a C++17 compiler. Captioning servers, training and evaluation use their own environments (see 1_preprocessing/6_captioning/serve.sh, 2_training/, 3_evaluation/).

Licence

Code: MIT. Models: CC BY-NC 4.0. Audio and captions: per source, see the dataset card.

Citation

@misc{wijngaard2026laionaudioxl,
  title     = {{LAION-Audio-XL}: A Web-Scale Corpus and Scaling Studies for Contrastive Language-Audio Pretraining},
  author    = {Wijngaard, Gijs and F{\"o}rster, Felix and Wightman, Ross and Nezhurina, Marianna
               and Schuhmann, Christoph and Jitsev, Jenia and Koepke, A. Sophia},
  year      = {2026},
  note      = {Preprint},
}