Back to Gap Map Model components / Language-specific evaluation sets

FLEURS

Google
open / Overall score: 4.0(strong)

FLEURS is an n-way parallel speech benchmark in 102 languages, made by recording people reading the FLORES sentences aloud. Each language has roughly ten hours of training speech plus separate development and test speakers, and it supports speech recognition, spoken language identification, speech translation and retrieval. Google built it as the speech counterpart of FLORES.

Openness

5 high confidence
5.0
license
cc-by-4.0
access
public
dataset_card
present

The recordings carry an attribution-only license and download from Hugging Face without a gate.

Adoption

4 high confidence
4.0

Hugging Face downloads over the trailing month for the FLEURS repository. Copies pulled through evaluation toolkits that mirror it are not counted.

Capability

not assessed

FLEURS is a speech test, and the map has no agreed way yet to turn hours of recorded speech into a measure of test depth, so it shows no capability score for now. Its overall score reflects adoption alone.

  • https://arxiv.org/abs/2205.12446 recorded 2026-11-08

    "an n-way parallel speech dataset in 102 languages ... with approximately 12 hours of speech supervision per language".

  • https://arxiv.org/abs/2305.13516 recorded 2026-11-08

    "more than halves the word error rate of Whisper on 54 languages of the FLEURS benchmark".

  • https://arxiv.org/pdf/2205.12446 recorded 2026-11-08

    "FLoRes-101 contains 3001 sentences extracted from English Wikipedia and these sentences have been translated in 101 languages by human translators"; "we collected three recordings by three different native speakers"; "each recording is evaluated by additional workers to assesses whether the recording corresponds to the input sentence"; "we focus on speech recognition, speech language identification and speech-text retrieval"; retrieval keys are "collected from the FLEURS test set"; Table 9 gives each language's ISO 639-3 code, listing "Filipino (Tagalog)" as tgl

  • https://huggingface.co/datasets/google/fleurs recorded 2026-11-08

    "We use 2009 n-way parallel sentences from the FLoRes dev and devtest publicly available sets, in 102 languages. Training sets have around 10 hours of supervision."

  • https://huggingface.co/datasets/google/fleurs/raw/main/README.md recorded 2026-11-08

    the card's dataset_info gives every language config's split sizes in num_examples (af_za: train 1032, validation 198, test 264), the same figures the Hub's size endpoint reports for all 102 language configs; "Speakers of the train sets are different than speakers from the dev/test sets"; "FLEURS-LangID is splitted in train/valid/test for each language"; "Results have to be reported on the test sets of Retrieval"

Verified 2026-09-24