By name: fix-thai-pdf description: Fix Thai PDF text extraction issues caused by legacy UPC fonts (DilleniaUPC, IrisUPC, AngsanaNew, BrowalliaNew, etc.). Applies two fixes - (1) PUA CMap repair mapping Private Use Area U+F700-F71A back to correct Thai Unicode using the NECTEC/Microsoft WTT 2.0 standard mapping, and (2) micro-offset Td nudge after zero-width glyphs to fix position collisions that scramble character ordering. Use when user has Thai PDFs with garbled text extraction, missing tone marks, or scrambled diacritics.
Fix Thai PDF
Fix text extraction from Thai PDFs created with legacy Windows UPC fonts.
Problem
Thai PDFs created with Microsoft Office using DilleniaUPC, IrisUPC, AngsanaNew, BrowalliaNew, and related UPC font families have two systematic text extraction bugs:
PUA Encoding: Thai combining marks and certain consonant variants are stored as positional-variant glyphs in Private Use Area codepoints U+F700–F71A. These are NOT simple offsets of standard Thai — they are repositioned forms (shifted down, left, or down-left) used by the font's shaping engine for tall consonants (ป ฝ ฟ), descender-less consonants (ญ ฐ), and stacked mark combinations. Text extractors output the raw PUA codepoints, producing garbage.
Position Collisions: Thai combining marks (ั ิ ี ึ ื ุ ู ่ ้ ๊ ๋ ์ ํ) and repositioned consonant variants have zero advance width in the font. Text extractors sort characters by x-position, so zero-width marks collide with the next base character, scrambling the reading order (e.g.
ผู้ให้→ผใู้ห้).
These issues affect a massive number of Thai PDFs from the 1990s through today: government documents, academic papers, legal filings, anything created with Microsoft Office using these fonts.
Solution
Two PDF-level fixes applied with pikepdf (preserving exact visual rendering):
Fix 1: CMap PUA Repair
Rewrites ToUnicode CMap entries in Identity-H fonts, replacing PUA destinations (U+F700–F71A) with correct Thai Unicode. The fix parses all bfchar and bfrange sections, expands any ranges that target PUA codepoints into individual bfchar entries (since PUA→Thai mappings are non-contiguous), then rebuilds the entire CMap cleanly.
Fix 2: Micro-Offset (Td Nudge)
Inserts 0.01 0 Td after every zero-width glyph's Tj operation in Identity-H font content streams. This advances the text matrix by 0.01 text-space units (~0.05mm at 14pt — invisible to the eye) but breaks position ties so extractors sort characters correctly.
Usage
pip install pikepdf pdfplumber
python fix_thai_pdf.py input.pdf output.pdf
Dependencies
pikepdf— PDF content stream and CMap modificationpdfplumber— verification only (optional, not needed for the fix itself)
PUA Mapping Table (NECTEC/Microsoft WTT 2.0 Standard)
Source: Thai Script Shaping by Theppitak Karoonboonyanan (NECTEC).
These are positional variant glyphs — the same Thai character rendered at different vertical/horizontal positions depending on the consonant class. The PUA codepoint encodes position, not identity.
Descender-less consonants
| PUA | Thai | Char | Description |
|---|---|---|---|
| U+F700 | U+0E10 | ฐ | THO THAN without descender |
| U+F70F | U+0E0D | ญ | YO YING without descender |
Above-vowels shifted left (for ascending consonants ป ฝ ฟ ฬ)
| PUA | Thai | Char | Description |
|---|---|---|---|
| U+F701 | U+0E34 | ิ | Left SARA I |
| U+F702 | U+0E35 | ี | Left SARA II |
| U+F703 | U+0E36 | ึ | Left SARA UE |
| U+F704 | U+0E37 | ื | Left SARA UEE |
| U+F710 | U+0E31 | ั | Left MAI HAN AKAT |
| U+F711 | U+0E4D | ํ | Left NIKHAHIT |
| U+F712 | U+0E47 | ็ | Left MAITAIKHU |
Tone marks shifted down-left (for ascending consonants + above vowel)
| PUA | Thai | Char | Description |
|---|---|---|---|
| U+F705 | U+0E48 | ่ | Down-left MAI EK |
| U+F706 | U+0E49 | ้ | Down-left MAI THO |
| U+F707 | U+0E4A | ๊ | Down-left MAI TRI |
| U+F708 | U+0E4B | ๋ | Down-left MAI CHATTAWA |
| U+F709 | U+0E4C | ์ | Down-left THANTHAKHAT |
Tone marks shifted down (for tone without above vowel)
| PUA | Thai | Char | Description |
|---|---|---|---|
| U+F70A | U+0E48 | ่ | Low MAI EK |
| U+F70B | U+0E49 | ้ | Low MAI THO |
| U+F70C | U+0E4A | ๊ | Low MAI TRI |
| U+F70D | U+0E4B | ๋ | Low MAI CHATTAWA |
| U+F70E | U+0E4C | ์ | Low THANTHAKHAT |
Tone marks shifted left (for normal consonant + above vowel)
| PUA | Thai | Char | Description |
|---|---|---|---|
| U+F713 | U+0E48 | ่ | Left MAI EK |
| U+F714 | U+0E49 | ้ | Left MAI THO |
| U+F715 | U+0E4A | ๊ | Left MAI TRI |
| U+F716 | U+0E4B | ๋ | Left MAI CHATTAWA |
| U+F717 | U+0E4C | ์ | Left THANTHAKHAT |
Below-vowels shifted down (for strict descender consonants ฎ ฏ)
| PUA | Thai | Char | Description |
|---|---|---|---|
| U+F718 | U+0E38 | ุ | Low SARA U |
| U+F719 | U+0E39 | ู | Low SARA UU |
| U+F71A | U+0E3A | ฺ | Low PHINTHU |
Results on Test Document (พุทธธรรม index, 140 pages)
| Metric | Before | After |
|---|---|---|
| PUA codepoints (471K chars) | ~13,000 | 0 |
| Thai text readable | ❌ garbled | ✅ correct |
Example: มาสูชคติแหชง → มาสู่คติแห่ง (F70A was ช, now correctly ่)
Known Limitations
- Only fixes Identity-H encoded fonts (the main Thai text carriers). WinAnsi-encoded fonts in these PDFs typically carry only punctuation/numbers and do not use PUA.
- The PUA mapping table covers the 27 entries defined in the Microsoft Windows Thai PUA extension (WTT 2.0). Apple MacOS Thai fonts use a different PUA range (U+F880–F89E) which is not yet supported.
- Does not fix text-level issues (wrong character order already baked into extracted strings). This is a PDF-level fix that must be applied before extraction.
Status
v0.2 — Corrected PUA mapping. v0.1 used a wrong simple-offset mapping (F7xx→0Exx) which mapped all 27 PUA entries to wrong Thai characters. v0.2 uses the authoritative NECTEC/Microsoft WTT 2.0 mapping. Tested on พุทธธรรม ฉบับปรับขยาย index (140 pages, 471K chars, 0 PUA remaining after fix).
Contributions welcome — see the GitHub repo for issues and PRs.