Language coverage

Honest multilingual quality grid.

Most dictation apps claim 100+ languages without telling you how good the transcription actually is. We test on the FLEURS benchmark and publish the truth.

7excellent25good~40functionalrestlimited

Tier A — Excellent

<5% WER

Production-ready. Use it for transcripts, drafts, and dictation in any setting.

English
Spanish
Mandarin
French
German
Portuguese
Italian

Tier B — Good

5–12% WER

Reliable for everyday dictation. Expect occasional cleanup on proper nouns and technical terms.

Hindi
Arabic
Russian
Japanese
Korean
Polish
Dutch
Turkish
Vietnamese
Indonesian
Bengali
Tamil
Ukrainian
Romanian
Greek
Czech
Hungarian
Swedish
Norwegian
Danish
Finnish
Catalan
Hebrew
Thai
Malay

Tier C — Functional

12–25% WER

Usable but expect to edit. We list 15 representative languages — about 40 land in this tier.

Swahili
Urdu
Telugu
Marathi
Punjabi
Gujarati
Filipino
Khmer
Lao
Burmese
Mongolian
Pashto
Sinhala
Welsh
Icelandic

Tier D — Limited

>25% WER

Most low-resource languages land here. Works but rough — your mileage may vary.

Most low-resource languages. We don't list them individually — quality varies too much across speakers and dialects to give an honest tier per language.

WER = Word Error Rate, measured against the FLEURS benchmark. Lower is better. Updated April 2026.

Methodology

Numbers, not adjectives.

Tested on the FLEURS evaluation set — 102 languages, clean read speech.

Scored with jiwer against normalized transcripts.

Missing a language? Tell us.

We benchmark Whisper large-v3-turbo on Google's FLEURS evaluation set: 102 languages, clean read speech, reference transcripts hand-verified by native speakers.

WER is computed using standard jiwer tooling against normalized transcripts (lowercased, punctuation stripped, Unicode-normalized). Lower-resource languages and accented English typically score worse — we report that instead of hiding it.

If your dialect or domain matters, tell us — we'll add it to the next benchmark run.