# YODAS v3: 1.1-million-hour multilingual web speech corpus

## Facts

- Provider: ESPnet / CMU WAVLab
- License: CC-BY-3.0 (https://huggingface.co/datasets/espnet/yodas3)
- Size: 13,373 files
- Format: Parquet (in sampled files)
- Stored at: Hugging Face, https://huggingface.co/datasets/espnet/yodas3
- Access: public, readable without an account
- Category: Speech & Audio
- Source: https://www.isca-archive.org/interspeech_2026/chen26d_interspeech.html
- Page: https://scigantic.com/archive/8f123c47-5cb4-502f-9a88-545700d04797
- Keywords: Speech & Audio, speech recognition, text-to-speech, Parquet, Hugging Face

## About

YODAS v3 is a large web-crawled speech dataset of over 1.1 million hours of audio in over 100 languages, all originally released under a CC-BY-3.0 licence, published by the ESPnet team as the third version of YODAS (YouTube-Oriented Dataset for Audio and Speech). The videos were crawled as a distinct set from YODAS v1 and v2, so there is no overlap with those releases. For each audio file it provides segment-level transcripts where available, English translations where available, language IDs, audio length and effective bandwidth. It is intended for automatic speech recognition, speech translation, text-to-speech and audio representation learning at foundation-model scale. This repository holds a data folder with one subfolder per language code (104 of them, for example data/en with audio and metadata subfolders) (measured 2026-11-14; the Hub reports about 56 TB). Stream it with the datasets library or read the files directly from the mounted repo; see the card for per-language hours. Described in an Interspeech 2026 paper. The publisher states the audio was released under CC BY 3.0; attribute the dataset and the original creators.

## Load the data

```python
from datasets import load_dataset
ds = load_dataset("espnet/yodas3")
print(ds)
```

## Data structure

File types in sampled files: Parquet

Example file: `data/ab/metadata/0000.parquet`

## Open it

- In a hosted notebook with the data mounted: https://scigantic.com/archive/8f123c47-5cb4-502f-9a88-545700d04797
- Full record as JSON (schema card included): https://api.scigantic.com/api/archives/8f123c47-5cb4-502f-9a88-545700d04797
- Over MCP: `get_schema_card` with id `8f123c47-5cb4-502f-9a88-545700d04797` at https://api.scigantic.com/mcp

## Related topics

- [Parquet datasets](https://scigantic.com/topics/parquet)
- [Speech & Audio datasets](https://scigantic.com/topics/speech-and-audio)
- [Speech recognition datasets](https://scigantic.com/topics/speech-recognition)
- [Text-to-speech datasets](https://scigantic.com/topics/text-to-speech)
