indic-pdf-check

Can a screen reader, a search engine or an AI read the text of this PDF?

Many PDFs in Indian and other South/Southeast Asian scripts look perfect on screen but carry a broken text layer: copy-paste, search, screen readers and AI pipelines get garbled words. Microsoft Word's "Save as PDF" is a common cause (see the research report). Others are scans with no text at all. This tool reads a PDF's own text layer and tells you which case you have, in a second, on your own computer.

Next steps after BROKEN: recover the text with Unshape; verify that extracted text matches what people see with SeenRead.

pip install indic-pdf-check
indic-pdf-check circular.pdf notice.pdf

circular.pdf: BROKEN             script=Malayalam letters=1994 broken=29.4% tagged=True creator='Microsoft® Word for Microsoft '
notice.pdf:   NO_TEXT            script=None letters=0 broken=0.0% tagged=False creator='Canon '
-- 2 PDFs: BROKEN 1, NO_TEXT 1

Pages whose own text has a high share of impossible spellings are listed as suspicious_pages (with --json), so damage confined to some pages or to a second script is visible.

Also reported: the script, the share of broken words, whether the PDF is tagged for accessibility, and the program that made it. --json gives machine-readable output; the exit code is 1 if any file is NO_TEXT or BROKEN (useful in CI).

Scripts: Devanagari (Hindi, Marathi, Nepali...), Bengali/Assamese, Gurmukhi, Gujarati, Odia, Tamil, Telugu, Kannada, Malayalam, Sinhala, Thai, Lao, Khmer, Myanmar, Thaana (Divehi).

Python:

from indic_pdf_check import check, check_text
check("notice.pdf")["verdict"]          # 'BROKEN' | 'OK' | 'NO_TEXT' | 'INSUFFICIENT_TEXT' | 'UNSUPPORTED_SCRIPT'

How well it works (tested before release)

• Word for Microsoft 365 PDFs with known-broken text in 14 scripts: 12 flagged BROKEN; Kannada was passed as OK (only 4% impossible spellings although many words are wrong - see limits).
• Clean source text in the same 14 scripts: 0 false alarms.
• Microsoft Edge PDFs of the same text are flagged in 3 of 4 scripts: they really do contain doubled vowel signs, which a screen reader also reads.

Limits

• It finds impossible spellings. Garbled text can still be spelled legally (e.g. a wrong consonant), so OK means "no sign of breakage", not "verified correct".
• Scripts not listed (Arabic, Hebrew, Latin, Chinese...) come out as UNSUPPORTED_SCRIPT (version 0.1.0 said NO_TEXT).
• Version 0.2.0 kept every verdict of the release tests above unchanged; it only splits the old NO_TEXT into three clearer cases and adds per-script and per-page checks.
• It reads only the text layer; it does not look at how the page renders.

Apache-2.0 · Novain Shibu · github.com/novain-shibu

It uses PyMuPDF (installed separately), which is licensed under AGPL-3.0 or a commercial licence from Artifex; using the two together is subject to PyMuPDF's licence.