Skip to content
TranslationAtlas
Tutti gli articoli

PDF and Scanned Document Translation

Pubblicato il 29 luglio 2026 · 6 min di lettura

A PDF isn't a text format, it's a presentation format, and that distinction causes more quoting confusion than almost anything else in this business. Some PDFs are exported directly from Word and contain clean, selectable text. Others are a photograph of a paper document, wrapped in a PDF container, with no text in them at all — just pixels. Those two files look identical in a preview pane and cost completely different amounts to translate.

What OCR actually does, and where it breaks

Optical character recognition converts an image of text into machine-readable text a translator can work with directly instead of retyping by hand. It works well on clean, high-resolution scans of typed documents: a well-lit office scanner, standard fonts, good contrast between text and background. It breaks down fast on anything else — a phone photo taken at an angle, a document scanned at low DPI, a stamp or watermark overlapping the text it's meant to preserve, carbon-copy forms with faded ink, or handwriting, which OCR still handles poorly regardless of how good the underlying model is. Non-Latin scripts add another failure point: OCR trained mainly on Latin-alphabet documents misreads Arabic, Chinese, or Cyrillic characters at a noticeably higher rate, especially with unusual fonts or a low-quality source.

When OCR fails silently — producing garbled but plausible-looking text rather than an obvious error — that's worse than it failing loudly, because it can slip past a translator working from the extracted text rather than the original image. This is why a competent translation workflow always keeps the original scan available for cross-reference, never relies on OCR output alone.

What makes a scan genuinely unusable

A handful of patterns show up repeatedly:

  • A photo taken at a steep angle, with text compressed or distorted toward one edge
  • Low light or a shadow across part of the page, especially common with phone photos of a document held flat by hand
  • A stamp or seal placed directly over the text it's meant to accompany, rather than in the margin
  • Multi-generation photocopies, where each copy-of-a-copy loses contrast until the text is barely legible
  • A resolution under roughly 200 DPI, where individual characters start blurring into each other

None of these make a document untranslatable, but they mean a human has to read and transcribe by eye instead of leaning on OCR, which adds time. If you have any choice in how the scan is produced, a flatbed scan at 300 DPI beats a phone photo every time, and it's usually free to redo before you send the file.

Layout reconstruction: why the output isn't just translated text

For most certified, sworn, or notarized work, the deliverable isn't a clean paragraph of translated text — it's a document that mirrors the original's structure: same headings in the same order, tables reproduced as tables, stamps and signatures noted in the position they appear on the source. This matters practically, not just cosmetically. An officer or registrar reviewing a translated birth certificate alongside the original wants to find the same information in the same place on both documents, without hunting for it. A translation that reflows everything into plain running text is harder to verify and more likely to get a second look it didn't need.

Stamps, seals, and illegible marks get an explicit annotation — something like "[official seal, partially illegible]" — rather than being silently dropped, because omitting them (even when they add no informational content) is a common reason official filings get bounced back for correction.

Why formatting drives cost more than word count sometimes does

A one-page form with a table, checkboxes, and a letterhead can take longer to translate than three pages of plain running prose, because someone has to manually rebuild the table structure and spacing rather than just render the words. Multi-column layouts, forms with fixed-width fields, and documents with mixed languages in different sections (a bilingual certificate, for instance) all add reconstruction time that a pure per-word estimate doesn't capture on its own. This is worth flagging to a vendor upfront rather than being surprised by it in the quote.

A different case entirely: video and subtitle files

Requests for video translation quotes work on a different axis altogether — per minute of runtime, not per word — because the deliverable is usually a subtitle file or a dubbed track, and pricing depends on whether a transcript or existing subtitle file (SRT/VTT) already exists. If it doesn't, transcription has to happen first, which adds a step and a cost line before translation even starts. Sending an existing transcript, even a rough one, is the single fastest way to get an accurate video quote instead of a placeholder estimate based on runtime alone.

What to send for the fastest, most accurate quote

  • The original file, if you have it — a native PDF or Word export beats a photo of a printed copy every time, and it's free to ask whoever issued the document for a digital copy before resorting to a photo.
  • The highest-resolution scan available, if the original truly doesn't exist digitally — 300 DPI flatbed over a phone photo, whenever you have the choice.
  • A note on the target format you need — a stamped, formatted PDF for official submission versus an editable Word file for internal use are different deliverables with different production time.
  • Any context on stamps, handwriting, or damage — telling the translator upfront what's illegible or unusual saves a round of clarifying questions mid-job.

Getting a quote on an unusual file

Upload whatever you actually have — PDF, scan, photo, or video file — and we'll tell you upfront if OCR quality is going to be an issue rather than discovering it mid-job. Documents heading to an official authority typically need certified, sworn, or notarized formatting; general documents just need standard translation. Start your order for an instant estimate followed by a human-confirmed fixed price within about two business hours, with nothing charged until you approve it.