PDF を翻訳

このツールでできること

This extracts a document's text in your browser, sends it to the AI provider you choose with your own key, and typesets the translation as a new PDF. What comes back is a clean document of translated text, not your original with the words swapped: German runs about a third longer than English and Indonesian shorter, so substituting in place would overflow some lines and leave holes in others. Paragraphs of prose survive that; images, tables, columns, headers, footers and the original pagination do not.

Use it when you need to read a document rather than reproduce it: supplier terms that arrived in a language you do not work in, a research paper you want to skim before paying to have it translated properly, an Indonesian regulation you need the substance of, or a manual whose instructions you have to follow today.

仕組み

1. Drop the PDF onto this page. Its text is extracted and shown to you, with an estimate of how much will be sent.
2. Choose your provider and paste your API key. It stays in this browser and goes only to the provider it belongs to.
3. Pick the target language. Thirteen are offered, and five of them come back as a text file rather than a PDF, for the reason below.
4. Set the register under formality if it matters, and the paper size for the typeset result, under advanced options.
5. Press Translate, then download the PDF, the plain text, or both.

The reason this produces a new document rather than an edited one is worth stating plainly, because plenty of tools imply otherwise. Text in a PDF is drawn at fixed coordinates; it does not sit in boxes that can grow. Replace an English sentence with its German translation and it is roughly a third too long for the space it has; replace it with Indonesian and there is a gap where the rest of the line used to be. Tools that do it anyway produce overlapping lines and clipped words. This one typesets the translation fresh, on the paper size you pick and with page numbers.

Five languages are delivered as text rather than as a PDF, and that is enforced rather than warned about. The font embedded in generated PDFs is Noto Sans, which covers Latin, Greek, Cyrillic and Vietnamese. Arabic, Chinese, Japanese, Hindi and Thai fall outside it, and a PDF built from glyphs a font does not have is a page of empty rectangles. So for those five you get a text file with a line at the top explaining why.

The text is cut into chunks of about 24,000 characters - smaller than the summariser uses, because a translation is roughly as long as its source and each request has to leave room for its own answer. Every chunk carries the same instruction: preserve meaning, structure and paragraph breaks, keep numbers, names and figures exactly, do not summarise and do not comment. Chunk boundaries fall between paragraphs, but a term rendered one way in one chunk can still come out differently in another.

Machine translation from a large model is good enough to work from and not good enough to sign. It handles ordinary prose well, technical vocabulary less reliably, and legal or contractual phrasing least reliably of all - which is to say it is weakest exactly where a translation matters most. For anything binding, this is a way to find out what a document says before engaging a translator, not a way to avoid engaging one.