Translate.
African languages, in and out of English and each other — with the quality of every language published beside it, so you know what you are getting before you rely on it.
Nothing matches that.
Nothing matches that.
The translation appears here.
How is this translation?
Ratings decide which sentences get reviewed first.
Suggest a better translation
A spelling, a word choice, the whole sentence — whatever is wrong. A speaker reviews it before it goes anywhere.
Machine output, uncorrected. For anything that matters — a clinic leaflet, a legal notice — have a speaker check it.
Every language, and how well we do it.
chrF++ translating from English into each language, on a held-out test set. Higher is better — it counts character overlap, which is fairer to languages that build long words out of parts than a word-level score.
Showing of 61 languages
No matching language yet — tell us what to add next.
More languages are added as evaluation sets are built for them. A score is only published once there is a held-out set to measure it on — an unmeasured language is listed without a number rather than given a flattering guess.
Three steps, no account.
Nothing to install, nothing to sign. Paste, pick, translate.
01
Pick your pair
Set the language you are writing in and the one you want. Either side can be any language on the list — English is a convenience, not a requirement.
02
Paste up to 2,000 characters
A paragraph or two at a time. Longer passages translate better broken into pieces, because the model keeps context within a passage rather than across one.
03
Take the output
Copy it, swap the pair to read it back the other way, or edit and translate again. Nothing is stored — close the tab and it is gone.
Trained on text people checked.
Most models learn these languages from whatever could be scraped. This one is trained on material speakers corrected and other speakers signed off — which is slower, and is why the scores are published rather than described.
61
languages with a published score, each measured on a set built for it rather than translated from somebody else's
Corrected, then checked again
A speaker fixes what the model got wrong; a second speaker either clears the fix or sends it back. Nothing enters training on one person's judgement.
Held out, not held back
Every score comes from sentences the model never saw while training. The weak languages are on the list beside the strong ones.
Which is why the numbers move. A language improves when people work on it, and the score published here is the one from the last time it was measured — not a target.
Nobody should have to
change language to be understood.
A nurse explaining a dosage, a farmer reading a warning, a parent at a school gate — each one currently switches into somebody else's language, or goes without. That is a daily tax on hundreds of millions of people, and it is paid quietly, by the people who can least afford it.
Every African language
The list on this page is a starting point, not a limit. A language belongs here whether it has forty million speakers or four hundred thousand.
Then the global south
The same gap runs through South Asia, Southeast Asia and the Pacific. Nothing about the method is particular to one continent.
Owned by its speakers
The people who correct a language decide what ships in it. The data is not extracted from them; it is made by them.
Which is also why every number on this page is published rather than summarised. Nobody should have to take our word for the quality — including us.
The things people ask first.
Short answers, including to the awkward ones.
Yes, for translating on this page. No account, no card, no trial that expires. There is a per-minute limit so one person cannot occupy the service, and if you need volume there is an API key with a free tier behind it.
Any language on the list can be either side. Worth knowing, though: pairs that do not involve English are generally weaker, because there is far less data connecting two African languages directly than connecting each of them to English.
Partly to keep the service responsive for everyone, and partly because quality falls off on very long passages — the model holds context within a passage better than across one. A long document translates better broken into sections anyway.
It varies enormously by language, which is why every score is published above rather than summarised into one headline number. Some languages are strong. Some are weak enough that you should treat the output as a first draft. Look yours up before you rely on it.
A measure of how closely the model's output matches a human translation of the same sentence, counted in runs of characters rather than whole words. Characters matter here: many African languages build long words out of parts, and a word-level score would call a nearly-right word entirely wrong. Higher is better.
Because somebody is going to use this for something that matters, and they deserve to know which languages we have not got right yet. A page showing only the good numbers would be more flattering and less useful.
Very likely, but not overnight — a language needs a body of reviewed text before a model can learn it, and that comes from speakers doing the work. If you can bring a few people who write it well, write to us and we will help you start.
Not unless you choose to send feedback. A plain translation is processed and forgotten — close the tab and it is gone. If you rate an output or suggest a correction, that submission is stored along with the sentence it refers to, because a correction without its context is useless.
We would rather you did not. Nothing is stored, but the text does travel to a server to be translated, and this is a free public tool rather than a service with a confidentiality agreement behind it. For patient records, legal papers or anything similar, talk to us about a private deployment instead.
A speaker of that language reviews it. Corrections that hold up join the data the next model is trained on, which is how the scores above move. It is not a suggestion box — it is the same pipeline our annotation communities work in, with a shorter on-ramp.
Yes — there is an API with a free tier, and the datasets behind the models are released openly. Please do not scrape this page to do it; the endpoint is easier for you and kinder to everyone else using the service.
Something we have not answered? Ask us directly.
Need this inside your own product?
The same model is available behind an API key — one endpoint to translate a string, another to send one prompt into several languages at once.
Part of the EveryLanguageMatters platform — building foundational multilingual AI infrastructure for every community, in every region and language.