What is the FLEURS benchmark? It is a public test set of read speech in 102 languages, used to compare speech recognition systems on the same material. Each system's output is scored against reference text, which gives one comparable error rate per language.
Transcribe a recording
Send your voice message, audio or video to @potok_now_bot on Telegram and the text comes back in the chat.
Open @potok_now_bot in Telegram
Files up to 2 GB · export to TXT, MD, DOCX, PDF, SRT and VTT · speaker labels and timestamps · about 30 languages
Key facts
- The speakers read sentences aloud, so the data is cleaner than real meetings, calls or street recordings.
- Scores are usually reported as word error rate, or character error rate for languages written without spaces.
- A language with many test hours behind it can still score poorly if the model had little training data for it.
- Published scores come from the model makers or third parties, not from the product that quotes them.
- Because every system reads the same sentences, a score difference between two systems on one language is easier to interpret.
What it looks like
from datasets import load_dataset
// one language of the test split, English (US)
fleurs = load_dataset("google/fleurs", "en_us", split="test")
print(fleurs[0]["transcription"]) # reference text for scoring
print(fleurs[0]["audio"]["sampling_rate"])
More
See the full list of languages, the tools and the glossary.
Frequently asked questions
Who publishes the FLEURS data?
The dataset is released by Google and distributed through Hugging Face. Anyone can load it, so researchers can reproduce a published score.
Does a good FLEURS score predict my result?
Only roughly. Read speech in a quiet setting is easier than a noisy interview, so your own recording will usually score worse than the benchmark.
Why do benchmarks matter at all?
They let you compare systems on the same test material and the same scoring method, which is more informative than any marketing claim about quality.
Are FLEURS numbers our own measurements?
When Potok pages quote them they are published results from others, not our own tests. Use them as a guide and try a short file of your own.
What does the benchmark not cover?
Noisy rooms, overlapping voices, spontaneous speech and special vocabulary are thin in read-speech data, so check those cases on your own audio.
Sources
Each source was opened by hand on the date shown. Pages change, so open the source if a detail matters to you.
- FLEURS speech dataset (Google, Hugging Face), Hugging Face · checked
- Word error rate (Wikipedia), Wikipedia · checked
Try it on your own recording
Send your voice message, audio or video to @potok_now_bot on Telegram and the text comes back in the chat.
Open @potok_now_bot in Telegram