OCR language support matrix for scans
We tested the recognition error rate and replacement results for 17 scripts on real scanned documents. Simplified Chinese and English are primary languages; 8 scripts including Thai, Traditional Chinese and Japanese are usable but need checking; 8 are still experimental; Lao scans are not usable.
The error rates and replacement pass counts in this table were measured on 14 October 2026 with the previous PDF engine. Since 20 October 2026 the software uses its own engine, pdfr, and these results have not been re-measured with the current version.
The current version does not yet support searching, replacing or redacting by keyword in Arabic (including Persian and Urdu), Hebrew and other right-to-left scripts: if the text to find or the new text contains such characters, the job stops before anything is changed. The replacement results in the Arabic and Hebrew rows of the table are records of the previous version and do not apply to the current one.
Checked against code version b6246f1 (pdfr) on 2026-10-21
| Script | Recognizer | Source | Tested on | Character error rate on script lines | Accepted on real scans | Result |
|---|---|---|---|---|---|---|
| Chinese (Simplified) | PP-OCRv6 | Built in | Tested throughout development | — | — | Primary support |
| English and Latin-script languages | PP-OCRv6 | Built in | Tested throughout development | — | — | Primary support |
| Thai | PP-OCRv5 | Built in | Real scans | 0.4% | 1/4 | Usable, needs checking Usable with review on modern print; experimental on old print and low-resolution scans |
| Chinese (Traditional) | PP-OCRv6 | Built in | Real + simulated scans | 0.3–0.4% | 23/23 | Usable, needs checking Horizontal replacement tested on a simulated scan only (23/23); vertical pages with zhuyin are refused |
| Japanese | PP-OCRv6 | Built in | Real scans | 0.3% | 8/9 | Usable, needs checking Vertical text and furigana tested on 3 real scans (8/9) |
| Korean | PP-OCRv5 | Language pack | Real scans | 2.9% | 5/5 | Usable, needs checking 5/5 measured after removing the hidden text layer; weight and typeface may not match |
| Vietnamese | Tesseract 5 | Language pack | Real scans | 3.1% | 2/4 | Usable, needs checking |
| Hindi (Devanagari) | PP-OCRv5 | Language pack | Real scans | 2.5% | 4/7 | Usable, needs checking |
| Georgian | Tesseract 5 | Language pack | Real scans | 1.6% | 5/6 | Usable, needs checking |
| Armenian | Tesseract 5 | Language pack | Real scans | 2.8% | 0/1 | Usable, needs checking |
| Tamil | PP-OCRv5 | Language pack | Real scans | 2.3% | 3/6 | Experimental |
| Arabic | PP-OCRv5 + Tesseract | Language pack | Real scans | ≈8–12% | 2/4 | Experimental Search and replace not supported in the current version |
| Hebrew | Tesseract 5 | Language pack | Real scans | 9.1% | 1/1 | Experimental Search and replace not supported in the current version |
| Bengali | Tesseract 5 | Language pack | Real scans | 9.1% | 2/4 | Experimental |
| Khmer | Tesseract 5 | Language pack | Real scans | 2.8% | 1/9 | Experimental |
| Burmese | Tesseract 5 | Language pack | Real scans | ≈10% | 3/5 | Experimental |
| Sinhala | Tesseract 5 | Language pack | Real scans | 6.2% | 1/6 | Experimental |
| Amharic | Tesseract 5 | Language pack | Real scans | 2.5% | 3/6 | Experimental |
| Lao | Tesseract 5 | Language pack | Real scans | 2.0% | 0/3 | Not usable on scans |
Primary languages are tested continuously throughout development, so this round of language reports does not list separate figures for them.
Method: for each script we used real scanned documents (including old prints, government gazettes and low-resolution images) and native PDFs. Character error rate is measured on "script lines" (lines where the script makes up more than half the text, NFC-normalized, whitespace removed). Accepted replacements are the number visually accepted / the number replaced; accepted means the old text is fully erased, the new text broadly matches in size, color and baseline, and adjacent text is undamaged.
The following scripts cannot be recognized on scans. They are written right to left, and the current version also cannot search or replace them in text-layer PDFs yet: Syriac and Thaana (Maldives)
Limits that apply to all languages
- Horizontal single-line text slanted by up to about 3°. Vertical text: Japanese tested on real scans; Simplified Chinese vertical is supported in code but untested; vertical Traditional Chinese with zhuyin is refused; other scripts are horizontal only
- Handwriting, decorative lettering, images below about 150 DPI and very small text have low recognition rates
- Matches with confidence below 80% are marked "Needs checking"
- The original font of image text cannot be identified, so it is approximated with the selected font
Cite this page: EnjoyPDF OCR language support matrix for scans, 2026-10-14.