GPT translator for text, documents and images

2. Everything you can open: 8 formats, 50 MB, 15 MB

The GPT translator takes eight kinds of file. Most tools tell you what they support; fewer tell you where they stop. Here is both.

Text — paste it into the box at the top of this page. No account, no character cap.

PDF — the layout gets rebuilt rather than flattened into a wall of running text. Columns stay columns.

Word — headings, numbered lists, footnotes and the numbering that depends on them.

Excel — cell contents translate. Formulas are left alone. Column and sheet structure holds.

PowerPoint — slide text and speaker notes, in place.

JSON — values translate, keys do not. Your app still boots.

Images — screenshots, photographs of printed pages, scanned forms.

Every format at once — mixed batches in a single pass.

Documents cap at 50 MB. Images cap at 15 MB. If your file is bigger than that, splitting it is the honest answer, and we would rather tell you here than fail after you have waited through an upload.

4. See the difference: six worked examples

These are real translation problems, each shown three ways. The literal column is what phrase-matching produces. The GLOTX column is what happens when a GPT translator reads the sentence in context first.

Idiom — English to German

What changed. The literal German says something close to "we should let sleeping dogs lie down on the pricing question." That is not an idiom in German. It is an image of dogs being placed on a topic, and a reader will stop and reread it. German has no proverb that carries "leave a settled matter alone," so there is no equivalent phrase to swap in. The substitution approach has nowhere to go. The GLOTX version drops the picture and translates the intention: better not to raise the subject again. The folksiness is lost, and that is the right trade. An idiom that lands as nonsense costs you more than plain phrasing ever will.

Formal register — English to Japanese

What changed. Politeness in Japanese lives in grammar, not vocabulary. The literal version uses 教えて, the plain imperative you would use with a friend, and あなた, a second-person pronoun that reads as blunt or oddly distant when aimed at a client. Both are defensible translations of the English words. Both damage the relationship. The GLOTX version uses the honorific お知らせいただけますでしょうか and drops the pronoun entirely, because Japanese business writing addresses its reader by implication rather than by name. English signals register through word choice. Japanese signals it through verb morphology, and anything that only swaps words will miss it every single time.

Ambiguous pronoun — English to French

What changed. English lets "it" stay vague. French does not. Both conseil and comité are masculine singular, so il can point at either, and a French reader will default to the more prominent noun, which here produces the reading that the board dissolved itself. The GLOTX version uses ce dernier, "the latter," pinning the reference to the committee. Notice what happened: the source sentence never resolved the ambiguity, and the translation had to, because the target language forces a choice the source language allowed you to duck. That is a judgement, not a substitution. A system matching phrases has no mechanism for making it.

Medical terminology — Spanish to English

What changed. Three false friends in a single sentence. Cuadro does mean picture, but in a clinical note it is the presentation, and English discharge summaries say "presented with." De esfuerzo is conventionally exertional, not "effort." And control in Spanish clinical writing means follow-up. The literal version tells an English-speaking clinician that something was being controlled, when what the record actually says is that the patient should come back in a month. Word by word, every literal choice is defensible. Read together, they change what the next clinician does.

Legal terminology — English to Spanish

What changed. Inquilino and propietario are the words you use talking about a flat. Arrendatario and arrendador are the words a Spanish lease uses about its own parties, and an agreement that calls them anything else reads as a translation rather than as an instrument. Asignar is assignment in the arithmetic sense; transferring a lease is ceder. Retener means holding something physically, which misses the clause entirely. This is about refusing consent, so denegarse. The literal version is comprehensible Spanish. It is not enforceable-sounding Spanish, and opposing counsel will notice before your client does.

Number and measure words — English to Chinese (Simplified)

What changed. Chinese counts nouns with measure words matched to the noun's class, and the literal version omits both of them. 三签合同 and 一复印件 are not awkward Chinese. They are ungrammatical. The GLOTX version supplies 份 for documents and 张 for sheets, and adds 已 to 签署 to mark the contracts as already signed, which English carries invisibly inside the participle. None of this is stylistic preference. A phrase-matching system that has not seen this exact sentence before has no way to choose the right measure word, because the correct measure word depends on what the noun is rather than on what surrounded it last time.

6. Who uses the GPT translator

Legal teams with a hundred-page supply agreement that arrived in the wrong language on a Friday. What matters is not elegance. It is that the defined terms hold from the definitions clause through to the final schedule.

Clinics and hospitals receiving discharge notes thick with abbreviations and regional drug names. Reading control en un mes as "control" rather than "follow-up" is a scheduling error with a patient attached to it.

E-commerce teams pushing forty thousand listings into six markets. If the size attribute renders three different ways, faceted search breaks and nobody can filter by it.

Universities and schools distributing the same problem set in four languages, where question numbering and mathematical notation have to survive intact or the instructions stop matching the problems.

Support teams taking tickets in twenty languages, written fast and full of product shorthand. "It keeps dying on me" is a hardware fault, not a bereavement, and the routing depends on knowing that.

Media and subtitling teams needing a first pass in eight languages. Character names, running jokes and who-calls-whom-what have to stay consistent, so the human subtitler is correcting a draft instead of starting again.

8. What is a GPT translator, and how is it different from machine translation?

A GPT translator runs on a large language model — the same class of system behind a chat assistant, pointed at one job. The name comes from the transformer architecture underneath it, which processes a passage as a whole and weighs how each word relates to every other word in it.

The systems that came before worked differently, and the difference explains most of what you have experienced. Statistical machine translation matched your text against enormous archives of previously translated documents and picked the most probable rendering. Neural machine translation improved the fluency of the output substantially, but still worked in short segments, translating a sentence with only limited sight of what came before it.

That segment boundary is where the errors live. A pronoun whose antecedent sat two sentences back. A term defined on page one and used on page thirty. An idiom that is only an idiom because of the sentence in front of it. All six examples above are versions of the same problem.

A GPT translator does not solve translation. It moves the failure point. It will still get low-resource languages wrong, and it will do so fluently, which is worse than getting them obviously wrong. Section 12 is about where the line is.

10. What we can and cannot tell you about accuracy

We do not publish an accuracy percentage. Not because the number would be unflattering, but because we have not run an evaluation we would be willing to defend, and a figure without a published method behind it is decoration.

Here is what that means in practice. Every GPT translator in this category quotes a number. Almost none of them say which language pairs it covers, which document types, which metric, how many samples, or who scored it. A 99% that turns out to mean sentence-level BLEU on news text tells you nothing about a Spanish discharge summary or a German lease.

What we can tell you is specific and checkable. The six examples above are real outputs, shown next to what literal translation produces, in six languages and six domains. Run the same sentences yourself in the free text box. If the output does not match what is on this page, we would rather hear about it than have you take our word for it.

We are building a published evaluation with a documented method and a fixed document set. When it exists, the results will appear here, including the cases where we come off worse.

12. When you still need a human translator

Certified and sworn translation. A court, a registry office or an immigration authority will require a translator with recognised credentials and a signed attestation. That is a legal category rather than a quality bar. No GPT translator output satisfies it, and anyone telling you otherwise is selling you a document that will be rejected.

Marketing transcreation. A campaign has to land, not merely mean the same thing. That work usually means writing fresh copy against the original brief, in the target culture, by someone who lives in it. Translation is the wrong tool and a good translator will tell you so.

Poetry and literary prose. Sound, metre, and ambiguity that the author put there deliberately. Translating these means choosing what to lose, and defending the choice. A person makes that call.

Low-resource languages. Where there is little training text, quality falls off sharply — and it stays fluent while it falls, which is the dangerous part. Confidently wrong output is harder to catch than obviously wrong output. If you are working into a language with a small digital footprint, have a speaker read it.

14. Frequently asked questions