Translator

About the Translator

Machine translation crossed a threshold a few years ago. It is now reliably good enough to understand a document written in a language you do not read, and to be understood by someone who does not read yours. It is not good enough to be trusted with a contract, a dosage, or a safety instruction, and no amount of improvement in the last few years has changed that, because the failures are rare and unannounced rather than frequent and obvious.

Within those limits this page covers twenty-eight languages, including eleven written in India. It works out the source language itself, so there is nothing to set on that side.

Two settings matter more than they look. The register control decides between formal and informal address, and in a great many languages that is grammar rather than style — Hindi has three levels of you, French and German two, and Japanese encodes it across the whole verb system. Getting it wrong does not produce a slightly off translation; it produces one that is rude or oddly intimate. The default matches whatever the original implied, which is right when you are translating something written by someone else and want their pitch preserved.

The romanisation switch writes the translation a second time in the Latin alphabet. It exists for a specific and common situation: you need to say something aloud in a language whose script you cannot read. It is skipped automatically when the target already uses Latin letters.

Idioms are translated as equivalents where one exists and as plain speech where none does, rather than word by word into nonsense. Proper nouns, brands, code, URLs and numbers are left alone. Where a term genuinely has no counterpart, the original word is kept with a short gloss in brackets, once. The one thing it will not do is improve what you gave it — a factual mistake in the source is translated faithfully rather than silently corrected.

How to use it

1. 1Paste the text. You do not need to say what language it is in.
2. 2Choose the language to translate into.
3. 3Set the register — formal or informal — if the situation calls for a specific one.
4. 4Turn on the Latin-alphabet version when you need to read the result aloud.
5. 5For anything consequential, have a fluent speaker check it before you rely on it.

Questions

How good is it, honestly?

Very good for prose of the kind people write to each other: emails, articles, instructions, messages. Weaker on poetry, wordplay, legal drafting and highly regional idiom. The characteristic failure is not gibberish, which you would notice, but a sentence that reads perfectly and means something slightly different — which is why anything that matters wants a human check.

Which Indian languages are covered?

Hindi, Bengali, Assamese, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi and Urdu. Quality is best on the ones with the most written material online, so Hindi and Bengali are stronger than Assamese, which is a fact about training data rather than about the languages.

Will it keep my formatting?

Paragraph breaks and list structure are preserved. Detailed markup is not guaranteed — if you are translating something with a lot of formatting, translate the text and reapply the formatting afterwards.

Can I translate a document rather than pasted text?

Not on this page, which takes text only. Extract the text with one of the document tools here, which run on your own machine, then paste it in. Nothing but the text you choose to paste ever leaves the device.