Dataset · google
FLEURS multilingual ASR benchmark
google/fleurs
Speech data in 102 languages for training and evaluating ASR on parallel sentences.
What's inside
FLEURS is the speech version of the FLoRes benchmark: 2009 parallel sentences across 102 languages. Train has approximately 10 hours of supervision per language, and train speakers differ from dev/test. Format is audio and text transcriptions; data is available in many languages.
What you can do with it
- Training multilingual ASR models
- Evaluating ASR across 102 languages
- Cross-lingual experiments
- Tasks
- Automatic speech recognition
- Languages
- Afrikaans, Amharic, Arabic, Assamese, Asturian, Azerbaijani, Belarusian, Bangla and 4 more
- Size
- 100K–1M rows
- Downloads this month
- 220.2K
- Likes
- 541
Models trained on it
- facebook/mms-1b-allAutomatic speech recognition 232.5K
- facebook/mms-lid-1024Audio classification 205.8K
- facebook/mms-lid-4017Audio classification 60.2K
- neuphonic/neucodecAudio → audio 55K
- SWivid/Habibi-TTS 44.5K
- facebook/mms-lid-256Audio classification 27.7K
- facebook/mms-lid-2048Audio classification 21.8K
- facebook/mms-lid-126Audio classification 15.3K
How to load it
from datasets import load_dataset
ds = load_dataset("google/fleurs")