YODAS2

YODAS2 is the long-form re-release of YODAS, the YouTube-Oriented Dataset for Audio and Speech assembled by the ESPnet team (Xinjian Li, Shinji Watanabe and colleagues). It gathers over 500,000 hours of speech across 149 languages from YouTube videos that carried Creative Commons licences at download time. Where the original YODAS shipped short clips, YODAS2 keeps each video as full-length audio with segment-level timestamps, which is what makes it suitable for long-form speech recognition, and it raises the sampling rate to 24 kHz.

Access is through Hugging Face. The corpus splits into labelled subsets, whose manual or automatic subtitles support supervised training, and unlabelled subsets aimed at self-supervised learning. You can stream the data or download it in bulk (the full set is tens of terabytes), loading the WAV audio and its utterance metadata with standard audio tooling.

For RAG and AI work this is a training and fine-tuning corpus rather than a text retrieval source. It suits multilingual speech recognition, speech foundation models, and self-supervised pretraining, and the long-form structure helps models learn context beyond a few seconds. You could also index the transcripts for spoken-language retrieval experiments.

Treat the licence with care. The dataset card is tagged CC BY 3.0 and attribution is required, but each video was published under its own Creative Commons terms, so some items carry share-alike obligations and the underlying licence varies item by item. The audio is YouTube-sourced, which brings takedown and terms-of-service risk, so verify individual items before any commercial use. It is broader than our Emilia entry, whose Emilia-YODAS portion draws only partly on this data; if you need speech generation specifically, Emilia is the more curated option, while YODAS2 is the wider, rawer multilingual pool.