Preprint
FLEURS: Few-shot Learning Evaluation of Universal Representations of Speech
Alexis Conneau et al., arXiv (Computation and Language), 2022
Read the originalRead on October 17, 2026
- What it is
- The paper that introduced FLEURS, a speech benchmark built from the FLoRes-101 machine translation sentences and read aloud in 102 languages. It defines the benchmark and gives baselines for speech recognition and other tasks.
- How much weight it carries
- Peer-reviewed-style research from Google, published on arXiv and used widely as a multilingual test set. The sentences are read speech from Wikipedia-derived text, so it does not test conversation or noise. Baselines are from 2022 models.
Facts taken from this source
- The FLEURS benchmark is a parallel speech dataset in 102 languages, with about 12 hours of speech supervision per language.
“FLEURS is an n-way parallel speech dataset in 102 languages built on top of the machine translation FLoRes-101 benchmark, with approximately 12 hours of speech supervision per language.” (Abstract)
Pages that rest on this source
- Dictation Languages: What Each App and Device SupportsGuide
- Hindi Voice Typing and Speech to TextGuide
- Speech Recognition Benchmarks and LeaderboardsGuide
- Whisper Supported Languages and Accuracy by LanguageGuide
- Speech to Text in French, German, Russian and MoreGuide
- What are the best datasets for speech recognition?Question