Document translation: PDF, Word, Excel and slides

Yes. The file is not rebuilt: DOCX, PPTX, XLSX and EPUB have only their text nodes replaced inside the original archive, and in a PDF the translation is drawn into the same page over the removed text. Tables, columns, styles, fonts, images and page breaks are never read and never recreated, so they come out exactly as they went in.

Slide text is translated, including tables and words inside shapes, along with speaker notes, chart titles and labels, and the text of SmartArt blocks. Slide masters and layouts are not: their text is prompt wording such as "Click to edit title", which an audience never sees, so translating it would spend your daily allowance on invisible strings and fill the review table with rows that match nothing on screen.

Formulas, cell references, numbers and sheet names are not touched at all: a translated SUM would stop adding up, and a renamed sheet would break every reference to it. What is translated is cell text, chart labels and comments. There is one honest side effect: a formula that MATCHES text — a VLOOKUP on a label, a SUMIF on a word, a data-validation list — will not find the translated label afterwards. The tool says so on the result screen.

A page with no text layer is recognized by a neural OCR engine in your browser, and the recognized text then travels the same path as ordinary text. The first such page takes longer while the recognition model downloads; after that it comes from cache. Those pages are flagged separately — recognition errs more often than reading a ready text layer, and they are worth re-reading.

Yes, and that is the point. When the translation is done you get a source-and-translation table and can edit any paragraph in it. Your version is what goes into the file. The table is searchable in both languages at once, so a term inside a four-hundred-paragraph contract is found by the word you remember — whichever language you happen to remember it in.

List those words under “Words that must not be translated”, one per line; use “=” to force your own wording. Note the cost: a protected word splits its sentence into parts that are then translated separately, so keep the list short and reserve it for names and codes rather than ordinary words.

Yes. In a PDF the translation is drawn as real text, not as a picture, so it can be searched, copied and read aloud by a screen reader. That is precisely why pages are not rasterised: a translation you cannot select is a photograph of a translation.

Identical paragraphs are grouped and translated once, and the result is placed everywhere it belongs. Contracts and books repeat themselves a great deal, so this saves both time and your daily allowance. The number of repeats found is shown in the summary above the table.

The standard mode translates paragraph by paragraph, and the translator never sees the neighboring paragraphs: the same term at the start and at the end of a contract can come out two different ways. Precise mode hands the document to a language model in consecutive batches — it sees the context, keeps one terminology from start to finish, follows the word list in the extra settings and applies your choice of register and subject area. A long document goes over in several parts, and the page says how many before it starts. For this the paragraph text is sent to our server and to the model (the recipients are listed on the AI providers page); the layout is then applied in the browser, from the original file.