Skip to content

Dataset · espnet

YODAS v3 100+ language speech corpus

espnet/yodas3

Over 1.1 million hours of web audio in 100+ languages for ASR, TTS, and audio representation learning.

What's inside

YODAS v3 is a large web audio corpus collected from videos, containing over 1.1 million hours of recordings. The audio was originally published under the CC-BY-3.0 license and covers over 100 languages. Videos for v3 were collected separately from v1 and v2 to avoid overlaps. The data is suitable for ASR, TTS, and audio representation learning.

What you can do with it

  • Train ASR models for many languages
  • Train TTS and audio representations
  • Validate multilingual speech pipelines
Tasks
Audio → audio, Automatic speech recognition, Text → speech, Translation
Size
1M–10M rows
Downloads this month
253.4K
Likes
263

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("espnet/yodas3")