# PDFSplat: make scanned text editable

PDFSplat can recognize printed English and Spanish text inside scanned PDFs,
photos, screenshots, and image regions. Recognition runs on your device with
bundled Tesseract.js 7 and language models. No document upload, API key, account,
or model download is required.

## Sensitive records and real redaction

**OCR covers are visual replacements, not secure redaction. Original text and images remain in the PDF and may be recovered. Do not use covers to hide FERPA-protected or other sensitive records before sharing.** This warning appears prominently in the OCR dialog and above the document when visual covers are present in the current editing session. Reopening an exported PDF does not restore the annotation history or reliably identify earlier covers.

Use your district's approved redaction workflow:

1. Keep the original in approved restricted storage; work on a copy.
2. In a real redaction tool, mark sensitive text **and images**, then **apply** the redactions. A colored rectangle or an un-applied mark is not removal.
3. Remove hidden information, including metadata, comments, attachments, hidden layers and embedded originals as supported by the tool.
4. Save a new copy, reopen it in an independent viewer, and check searching/copying text, image content and hidden information. Follow district review and release procedures before sharing.

[Adobe's apply-and-sanitize workflow](https://helpx.adobe.com/acrobat/desktop/protect-documents/redact-pdfs/redact.html) is an example; use the tool approved by your organization. [U.S. Department of Education de-identification guidance](https://studentprivacy.ed.gov/resources/data-de-identification-overview-basic-terms) provides context for education records. OCR, visual covers, PDF passwords and PDFSplat's rasterized sanitize option are not certified redaction or a guarantee of FERPA compliance. Rasterizing may leave sensitive information visibly present. OCR does not automatically produce a tagged, accessible PDF.

## Recognize, review, and apply

1. Open a PDF or image in [PDFSplat](../solutions/pdfsplat/). For camera scans,
   choose **Tools → Scan to PDF**, correct the corners, and **Add scans to editor**.
2. Rotate pages upright before recognition. Set fine rotation to **0°**; use a
   straightened source image or the scanner's perspective correction if needed.
3. Choose **OCR** (Recognize text) in the top toolbar. Select the current page,
   selected pages, or all pages, then English, Spanish, or both languages.
4. Choose **Recognize text**. Review the original-page previews and every detected
   line. Correct spelling, accents, punctuation, and reading order; uncheck anything
   you do not want. Confidence below 75% is highlighted; below 50% starts unchecked.
5. Choose **Apply selected text**. PDFSplat adds editable text boxes with visual
   covers over the original lettering. Click a box to type or use **Selected text**
   in the sidebar. Drag its move handle to reposition it; the original cover stays
   anchored. Text size, color, formatting, and resizing remain available.
6. **Save as → PDF** creates a new PDF with selectable text. Applied OCR also feeds
   text-based document/ebook exports. **Download recognized text** in the review
   panel saves checked lines as TXT without changing the PDF.

Applying a batch is one Undo step. Closing a review makes no changes. Stop
terminates the recognition worker, including during model loading; completed
page previews remain available. Recognition skips existing selectable text and
regions outside the current crop. It reads the visible composition, including
inserted images and visual covers, rather than uncovering masked image pixels.
Running OCR again skips text already applied.

## Offline and self-hosted use

Use the v3.1.49 (or later) PDFSplat, Tools, full DrawSplat, or complete district-suite ZIP from
[Downloads](../pages/download.html#pdfsplat-ocr). Extract the whole ZIP and retain
`solutions/pdfsplat/vendor/ocr/` plus the shared vendor/assets folders. Open
**START-HERE.html** and use the included local launcher (Python 3), or serve the
files on an HTTPS static host. OCR workers need HTTP/HTTPS and do not run from a
`file://` page. Ordinary PDF editing still supports direct-file opening in tested
desktop Chrome and Firefox. Local OCR needs no internet connection.

The worker and language data are discarded when recognition finishes or stops.
Models are bundled locally and are not cached in IndexedDB. Documents and review
results stay in memory; save a result before closing the app. Optional Google
Drive saving remains a separate, explicitly authorized action.

## Accuracy, layout, and limits

- This is printed-text recognition, not handwriting or mathematical notation
  reconstruction. Clear, upright scans with simple backgrounds work best.
- Fonts, colors, spacing, columns, and covers are approximate. Complex artwork,
  low-resolution images, gradients, and rotated text require careful review.
- Output uses standard PDF fonts supporting English and Spanish accents. If a
  reviewed line contains an unsupported symbol, Apply asks you to correct or
  uncheck it before making any changes. Very long corrections must fit the box.
- Each batch accepts up to **20 pages and 2,000 new lines**. Page rendering is
  bounded to **8 million pixels** and a **3,200-pixel longest edge**. Recognize
  smaller batches on devices with limited memory.
- Set orientation before OCR. Changing page rotation after applying OCR is blocked
  to protect placement; Undo the OCR batch, rotate, and recognize again. Fine
  rotation must be 0° when recognizing.
- Covers preserve the original image underneath. They are **visual replacement,
  not secure redaction**. Scanned text that you leave unchecked remains an image.
- Saving PDF writes text into the output. Reopening it uses PDFSplat's existing
  click-to-edit replacement workflow; the in-memory annotation history is not
  saved as a restorable project. Sanitized image-only export removes selectability.

## Lower-memory mode and accessibility testing

Select **Lower memory use** before recognizing. It caps batches at **5 pages / 500 lines**, page rendering at **2 million pixels / 2,200-pixel longest edge**, and smaller review previews. It is initially selected when the browser reports at most 4 GB memory or at most two CPU cores; browsers may omit or approximate these hints. Users can change the setting. These hints are not saved or sent as telemetry. Lower resolution can reduce recognition quality; try one page and review carefully.

The dialog exposes its safety and accuracy descriptions, review confidence, busy state and language to assistive technology. Spanish review/edit boxes declare `lang="es"`. Stop returns focus to the retry button; recognition progress announcements are limited to useful increments. Automated checks cover keyboard navigation, cancellation, narrow screens, real Spanish OCR/accented exports, offline requests and the actual lower-resolution worker input. Mocked 2 GB/two-core browser hints are **not** a physical low-memory test. They do not demonstrate ChromeOS performance or spoken Spanish pronunciation.

Actual Linux Orca/Chrome testing on a private virtual display and session bus announced the redaction warning, workflow, language controls, Stop, a recognized Spanish line and its confidence. Speech was captured silently from Orca; audible pronunciation and braille were not assessed. A separate process group capped at **1 GiB RAM, zero swap and one CPU** completed a synthetic Spanish page with accents and no external requests; peak memory was **731,795,456 bytes**. This is a real process limit, not physical Chromebook testing. [Recorded environment and observations](../data/pdfsplat-ocr-validation-2026-10-31.json).

The physical Chromebook/ChromeVox, NVDA/Firefox, VoiceOver/Safari and low-memory-device matrix remains open in [issue #66](https://github.com/mguhlin/drawsplat/issues/66). Complete those checks with representative district documents before claiming assistive-technology or device coverage; use synthetic records for testing. Reading order, headings, tables, columns and exported PDF tags require separate review.

## Dependencies and verification

Tesseract.js 7.0.0, Tesseract.js-core 7.0.0, and the English/Spanish best-integer
models from `@tesseract.js-data/*` 1.0.0 are bundled under Apache-2.0. Provenance,
licenses, package integrity, and file hashes are in
[`vendor/ocr/`](../solutions/pdfsplat/vendor/ocr/README.md).

Real-engine Chrome and Firefox checks cover English/Spanish scans, mixed native
text and scans, saved selectable text, corrected/excluded lines, atomic Apply,
Undo/Redo, duplicate detection, cancellation, missing-model recovery, movement
with anchored covers, and phone review. No external recognition requests are
made. Physical phone performance and complex layouts still require review.

In v3.1.48, thumbnails reflect applied text edits and Undo/Redo. OCR estimates background color from around a detected line, improving light text on dark covers. Complex backgrounds remain approximate. To use this sampling on an earlier result, reopen the original PDF and recognize it again. [Toolbar and ODT guide](pdfsplat-odt-toolbar.md).