CLAP2-L

CLAP2-L

A 340M-parameter audio-text model from LAION that places sounds and English text in one embedding space, for zero-shot audio classification, audio-text search and general audio embeddings.

Made by laion on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: CLAP2-L is the largest of five CLAP2 models from LAION, which learn to match audio clips with English text. It pairs an audio encoder with a text encoder, and both place their output in the same vector space. The files come in OpenCLIP and Transformers formats.

You can use it to classify sounds without training, to search audio with text, or to get embeddings for general audio work. The audio side accepts clips of any length with no cropping or padding to a fixed size.

Who it suits: Developers and researchers who work with audio, such as sound search, tagging and similarity, and who want an open model instead of a hosted service.

Industry | Who uses it | How they use it | Example
Media | Archive teams | Search sound libraries with text | Finding clips described as applause
Software | App developers | Tag or classify audio without labeled data | Labeling uploaded sound effects
Research | Audio researchers | Use embeddings as features | Clustering field recordings
Music and games | Sound designers | Retrieve sounds by description | Finding a door slam in a library

You can skip it if you need speech transcription or audio generation, since this model only matches sound and text.

Where to use it: It runs in Python through OpenCLIP, and the card says Transformers format is also available.

• Python with open_clip_torch and soundfile
• Windows, macOS and Linux
• Notebooks and back ends that handle audio files

Use case | Who | What they get back
Zero-shot classification | Developers | Probabilities for each text description you supply
Audio-text retrieval | Search teams | Clips ranked against a text query
Audio embeddings | Researchers | One vector per clip for downstream use

How to set it up: You need Python and the open_clip_torch package, plus soundfile to read audio. The first call downloads the model from the Hugging Face hub.

• Python with pip
• open_clip_torch 4.2.1 or newer
• soundfile for reading wav files
• About 1.4 GB of disk for the F32 weights

1. Install Python.
2. Install the packages with the command below.
3. Run the quick start script on a wav file of your own.

pip install "open_clip_torch>=4.2.1" soundfile

pip install -U huggingface_hub
hf download laion/clap2-large-audioxl-s1B

How to use it: Load the model and tokenizer from the hub, preprocess a clip, and compare it with several text prompts. The script prints a probability for each prompt.

import torch, open_clip
from open_clip.audio.io import load_audio

hub = "hf-hub:laion/clap2-large-audioxl-s1B"
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())})

To encode several clips at once, pad them to the longest clip. A mask keeps the padding from changing the embeddings.

from torch.nn.utils.rnn import pad_sequence
items = [preprocess()(load_audio(f)) for f in ["dog.wav", "rain.wav"]]
batch = {k: pad_sequence([x[k] for x in items], batch_first=True) for k in items[0]}
with torch.no_grad():
    audio_embeds = model.encode_audio(batch, normalize=True)

Tool | What it does
create_model_from_pretrained | Loads the model and its preprocessing from the hub
get_logits | Scores the audio against your text prompts
encode_audio | Returns an embedding for each clip
encode_text | Returns an embedding for each prompt

Example: You supply a recording of a dog and three prompts: a dog barks, a door slams shut, and people are applauding. The card shows that for a dog-barking clip, the first prompt gets nearly all the probability.

For newcomers and experts: If you are new, start with the quick start script and a short wav file. Write prompts as full sentences, such as a cello playing a slow melody, rather than bare labels.

If you know the field, the audio encoder is a NaFlex-style vision transformer fed with log-mel patches, producing 50 tokens for every second of sound. Input is averaged to mono and resampled to 32 kHz. Average several prompt embeddings per class for zero-shot work.

Load audio of any length
Resample to 32 kHz mono
Log-mel patches into the audio tower
Text goes through the text tower
Compare embeddings in a shared space

Good to know: The card lists five sizes trained for the same number of samples. The chart shows the MAEB-30 score of each, where higher is better.

Title: MAEB-30 score by model size, percent
L, this model: 53.6 percent
M: 52.3 percent
S: 51.8 percent
XS: 49.1 percent
XXS: 46.2 percent

• License: Apache 2.0
• Cost to run: none named
• Data it reaches: the audio files you load and the model files you download
• Not for: non-English text prompts, since the text side handles English

On Hugging Face, 13 November 2026

• Task: feature extraction
• Library: open_clip
• Parameters: 340M
• Weights: about 1.4 GB in F32; about 170 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 396
• Likes: 6
• License: apache-2.0
• Languages: en
• Quantized versions: 1
• Tags: open_clip, safetensors, clap2, audio, clap, contrastive, audio-text-retrieval, zero-shot-audio-classification, feature-extraction

Links

• Model: https://huggingface.co/laion/clap2-large-audioxl-s1B
• Files: https://huggingface.co/laion/clap2-large-audioxl-s1B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download laion/clap2-large-audioxl-s1B

Then follow the usage in its model card. Nothing in this kit runs by itself.