Scanned book translation: what OCR accuracy alone can't deliver

Many publishers get stuck at the same point: legacy titles, rare works, and licensed files are scanned to PDF, yet still can't be translated. The bottleneck isn't character recognition—it's structure.

Body text, footnotes, captions, tables, and headers are jumbled together. A single misplaced footnote number in an academic book forces translators to backtrack constantly, making the process slower than manual entry.

We push multi-engine fusion above 99%, including blurry pages and handwriting fallback. But in publishing, accuracy is the floor. What makes text actually translatable is structure recovery: which content is body, which is notes, whether merged table cells survive, whether captions still map to references.

Structured output must be editable Word/TXT—not an image-like PDF that only looks faithful. Only then can translation agents, termbases, and QA systems take over directly.

Every OCR pipeline has blind spots: water damage, gutter shadows, worn type. Burying them inside a '99% accurate' claim just shifts risk to the editorial team.

We flag low-confidence characters and regions separately, shipping a confidence list. Reviewing only the suspicious 1% is an order of magnitude faster than a full read-through, and avoids errors that 'look right'.

A 15-volume academic series, 8.9M characters, delivered in 10 weeks: OCR, terminology harmonization, translation, and full review. The win wasn't raw speed—it was fewer do-overs. Structure right the first time, terms locked once, editors spending time only on real questions.