More data does not help much if its captions and reuse rights are opaque. This corpus tackles both issues at scale: it combines audio-text examples from 462 source datasets, adds generated captions where needed, and makes each clip’s licence and attribution part of the dataset record.
What Sets It Apart
- The corpus contains about 366 million clips, but only around 4.36 million are hosted as audio; the rest are represented by links and metadata. This distinction matters when estimating what can be downloaded and used directly.
- Audio is split into open, share-alike, and non-commercial configurations. Each clip carries its source licence and attribution, so access to a file should not be mistaken for permission to use it without conditions.
- The data supports the LAION CLAP2 scaling study and its family of contrastive audio-text models. Short and long captions, source text, and model-generated descriptions provide different levels of supervision for retrieval and classification research.
Who It’s For
A good fit for researchers studying audio-text retrieval, sound classification, or training-data scaling who can filter examples by licence and source. Look elsewhere if you need a uniformly licensed, fully downloadable corpus, or if your task depends on long recordings: clips are capped at 30 seconds, and linked media remain subject to their owners’ terms.
