YODAS2: YouTube-Oriented Dataset for Audio and Speech (long-form)
- Provider
- Xinjian Li, Shinnosuke Takamichi, Takaaki Saeki et al. (CMU WAVLab, ESPnet)
- License
- CC-BY-3.0 (source)
- Size
- 100,000 files
- Format
- JSON (in sampled files)
- Stored at
- Hugging Face (https://huggingface.co/datasets/espnet/yodas2)
- Access
- Free and public. No account with the data provider is needed.
- Category
- Speech & Audio
- Source
- https://arxiv.org/abs/2406.00899
About YODAS2: YouTube-Oriented Dataset for Audio and Speech (long-form)
YODAS (YouTube-Oriented Dataset for Audio and Speech) is a large-scale multilingual speech dataset of over 500,000 hours in more than 100 languages, collected only from YouTube videos carrying a Creative Commons licence. It has labelled subsets with manual or automatic subtitles for supervised ASR training and unlabelled subsets for self-supervised learning. YODAS2 provides the same data as espnet/yodas in long-form, video-level format: audio is not segmented and is encoded at 24 kHz instead of 16 kHz, and each record carries the whole video's audio, its duration and a list of utterances with text and start and end timestamps. It is one of the largest open multilingual ASR corpora. This repository holds a data folder with one subfolder per language shard (230 of them, such as en000, each with audio, duration and text subfolders) and a SQLite metadata file (measured 2026-11-14; the Hub reports about 60 TB). Load a subset with datasets.load_dataset('espnet/yodas2', 'en000', streaming=True). The YODAS paper lists all subsets as CC BY 3.0, the licence of the source videos; attribute the dataset and, where needed, the original channels.
How to load YODAS2: YouTube-Oriented Dataset for Audio and Speech (long-form) in Python
from huggingface_hub import hf_hub_download
path = hf_hub_download(repo_id="espnet/yodas2", filename="data/aa000/audio/00000000.tar.gz", repo_type="dataset")
print(path)Data structure
File types in sampled files: JSON
Files
Example file: data/aa000/audio/00000000.tar.gz
Analyze this dataset without downloading it
On Scigantic, YODAS2: YouTube-Oriented Dataset for Audio and Speech (long-form) opens in a hosted JupyterLab notebook with the data already mounted, so you can read files where they are stored instead of copying them first. Each notebook includes an AI assistant that can see the mounted files and write and run analysis code. Open the dataset page to launch a notebook, or browse the full catalog.
This page is also available as Markdown and as JSON.
Related Speech & Audio datasets
- LibriTTS-R: restored LibriTTS speech corpus
- LibriTTS multi-speaker English text-to-speech corpus
- LibriSpeech-Long: Long-Form Speech Benchmark of Contiguous Audiobook Segments up to 4 Minutes (Google DeepMind)
- CML-TTS multilingual speech synthesis dataset
- Omnilingual Asr Corpus
- VoiceBank+DEMAND noisy speech database (16 kHz)
- Bach Doodle Dataset: 21.6 Million User Melodies with Coconet Four-Voice Bach-Style Harmonizations (Google Magenta)
- CVSS: Multilingual Speech-to-Speech Translation Corpus, 21 Languages into English (CVSS-C and CVSS-T, Google Research)