A scanned PDF is usually a set of images wrapped in a PDF. Sometimes only a few pages are scans, while the rest still have a real text layer. Treating the whole file as an image would throw away text you already have; treating it as text would ignore the pages that need recognition.
AudiobookCraft handles this page by page. It reads the existing text first, recognizes only the pages that need it, and shows you the pages whose result is not reliable.
How page-by-page OCR works
- The document checker reads every page’s text layer.
- Pages with usable text keep their original words and reading order.
- Pages with very little text and a scanned image are rendered and processed with local OCR in English and Simplified Chinese.
- Recognized pages are merged back at the same page numbers.
- Pages whose OCR is weak, or whose body text is missing, become structured page notes instead of silent successes.
The upload size, page count and character limits are the same for scanned and text PDFs: up to 20 MiB, 100 pages and 200,000 extracted characters. See current limits.
Mixed PDFs keep the text you have
In a mixed PDF, selectable text pages are never replaced by OCR. Only the pages that need recognition are rendered, which keeps existing punctuation, names and numbers intact and avoids a full re-scan.
If a page that needs OCR fails, the document stays blocked rather than producing an audiobook with a hidden gap. You can exclude the unreadable page if the remaining pages still tell a coherent document, or fix the source and upload again.
Page review: see what will not be read
The review page lists structured notes for each page:
- visual content that is present but not described;
- footnotes excluded from narration;
- table-like or ruled layouts where numbers may be interleaved;
- low OCR confidence, low scan resolution or skewed text;
- missing body text;
- reading order that could not be established reliably.
Warnings about quality or visual limits can be confirmed. Missing text and unsafe reading order cannot be confirmed away: those pages must be excluded or the source corrected. Choices are saved with the text version, so a refresh keeps them and older tasks keep their original snapshot.
Two-column and complex layouts
Simple two-column pages are reordered left column, then right column, with full-width headings and blocks inserted at their vertical position. Asymmetric columns and same-page mixes are handled when the geometry is clear.
Three or more columns, narrow sidebars and rotated body text are flagged as unreliable rather than guessed. Table cells and code listings are best effort; the tool does not claim to understand a chart or to preserve every table’s cell order.
How accurate is OCR?
Accuracy depends heavily on the scan. On our fixed clear-scan samples — 300 DPI printed pages in English and Simplified Chinese — measured character error rates were 0.00% for English and 4.49% for Chinese on October 18, 2026.
Those numbers are not a guarantee for a phone photo, a faded photocopy or a skewed page. Blurred, tilted, low-resolution or noisy scans are the ones most likely to be flagged. A quality warning is useful information, not a promise of failure; preview the document when accuracy matters.
On the resource side, we also test a 100-page scan against our target 2 vCPU / 2 GiB environment. That check is about completion and memory, not about the quality of any particular scan.
What it cannot do
- Describe images, charts or diagrams.
- Translate a document when you choose a narration language.
- Guarantee a reliable reading order for every complex layout.
- Recover text that is genuinely unreadable in the scan.
- Confirm missing body text into existence.
Ready to listen?
Upload a scan and review the notes before committing to a full audiobook. Convert a scanned PDF free →