Printed text & born-digital PDFs
Embedded PDF text (born-digital) reads at 100% accuracy with no model involved. Printed scans run through Tesseract for Latin scripts and PaddleOCR for CJK/Indic.
Languages
35+ scripts. Real coverage.
Best-in-class engine per language.
We route each script to the engine that actually handles it well — PaddleOCR for printed CJK, Surya for Indic + Arabic, Tesseract for Latin, vision-LLM for handwriting. Three maturity tiers (stable / beta / experimental), per-language accuracy notes, honest caveats. The full matrix below.
Beyond language
We also read what others miss.
Language coverage is only half the story. The other half is document shape and quality: handwriting, complex tables, degraded scans. Here's what the engine handles regardless of language.
Handwriting — three tiers
Handwritten English (Latin/Cyrillic/Greek) uses a fast vision model at 5 cr/page. Handwritten non-English (CJK/Indic/Arabic) routes to the script's best-in-class model — Gemini Flash for CJK, GPT-4o for Arabic/Hebrew. Handwritten mixed-language (forms with 2+ scripts) runs a dual-engine pipeline. Each tier priced separately.
Multi-column tables
Single-page tables extracted with bounding-polygon awareness. Multi-page tables (continuation rows across pages) are on the roadmap — tracked in KNOWN_ISSUES.
Degraded scans
Low-resolution photos, faxed documents, stained pages. The cascade escalates to vision-LLM automatically when traditional OCR confidence drops below 0.65.
Tier · Stable
Stable support
Production-ready. Validated against a comprehensive fixture set. Use these in business-critical workflows with confidence.
English
en- Script
- Latin
- Engine
- Tesseract
- Accuracy
- 99%+ on born-digital PDFs; 96–98% on clean scans.
Caveats
Degraded scans escalate to vision-LLM (Tier 3).
Spanish
es- Script
- Latin
- Engine
- Tesseract
- Accuracy
- 99%+ on born-digital; 96–98% clean scans.
Caveats
Diacritic-handling validated; bilingual EN/ES docs use dominant-script routing.
French
fr- Script
- Latin
- Engine
- Tesseract
- Accuracy
- 97–99% on born-digital; 95–97% clean scans.
Caveats
Diacritic-heavy text fully validated.
German
de- Script
- Latin
- Engine
- Tesseract
- Accuracy
- 97–99% on born-digital; 95–97% clean scans.
Caveats
Umlauts + ß handled; compound nouns parsed correctly.
Italian
it- Script
- Latin
- Engine
- Tesseract
- Accuracy
- 97–99% on born-digital.
Caveats
Accent handling validated.
Dutch
nl- Script
- Latin
- Engine
- Tesseract
- Accuracy
- 97–99% on born-digital.
Caveats
Compound-word handling validated.
Portuguese
pt- Script
- Latin
- Engine
- Tesseract
- Accuracy
- 95–97% on clean printed text.
Caveats
BR and PT variants both supported; currency/date formats locale-aware.
Chinese (Simplified)
zh-Hans- Script
- Han
- Engine
- PaddleOCR
- Accuracy
- 94–96% on clean printed text.
Caveats
Printed-only at Tier 2A; handwritten CJK routes to Gemini (HW-NE mode).
Chinese (Traditional)
zh-Hant- Script
- Han
- Engine
- PaddleOCR
- Accuracy
- 92–95% on clean printed text.
Caveats
Some less-frequent traditional characters benefit from HW-NE mode.
Japanese
ja- Script
- Han + Hiragana + Katakana
- Engine
- PaddleOCR
- Accuracy
- 93–96% on clean printed text.
Caveats
Handwritten Japanese (HW-NE mode) uses Gemini Flash — NLS 0.899 benchmark.
Korean
ko- Script
- Hangul
- Engine
- PaddleOCR
- Accuracy
- 92–95% on clean printed text.
Caveats
Mixed Hangul/Hanja documents (formal text) handled.
Hindi
hi- Script
- Devanagari
- Engine
- Surya
- Accuracy
- 90–94% on clean printed text.
Caveats
Handwritten Hindi: HW-NE mode (Qwen3-VL).
Marathi
mr- Script
- Devanagari
- Engine
- Surya
- Accuracy
- 88–92% on clean printed text.
Caveats
Shared Devanagari model with Hindi.
Tamil
ta- Script
- Tamil
- Engine
- Surya
- Accuracy
- 88–93% on clean printed text.
Caveats
Complex consonant clusters fully validated.
Telugu
te- Script
- Telugu
- Engine
- Surya
- Accuracy
- 88–93% on clean printed text.
Caveats
Compound vowel signs handled correctly.
Bengali
bn- Script
- Bengali
- Engine
- Surya
- Accuracy
- 88–92% on clean printed text.
Caveats
Conjuncts and ligatures handled.
Punjabi
pa- Script
- Gurmukhi
- Engine
- Surya
- Accuracy
- 88–92% on clean printed text.
Caveats
Gurmukhi-specific model; eMunshi corpus has validated 1M+ pages.
Arabic
ar- Script
- Arabic (RTL)
- Engine
- Surya
- Accuracy
- 89–93% on clean printed RTL text.
Caveats
Handwritten Arabic routes to GPT-4o (HW-NE mode); cursive ligatures are the hardest non-CJK problem.
Tier · Beta
Beta support
Works on clean documents. Real-world accuracy depends on your document quality. Validate on a sample before committing to volume.
Vietnamese
vi- Script
- Latin (diacritic-heavy)
- Engine
- Surya
- Accuracy
- 93–96% on clean printed text.
Caveats
Stacked tone marks; validate on your scan quality.
Tagalog
tl- Script
- Latin
- Engine
- Tesseract
- Accuracy
- 94–96% on clean printed text.
Caveats
Spanish loanwords and abbreviations handled correctly.
Nepali
ne- Script
- Devanagari
- Engine
- Surya
- Accuracy
- 85–90% on clean printed text.
Caveats
Validate on real corpus before production deployment.
Kannada
kn- Script
- Kannada
- Engine
- Surya
- Accuracy
- 85–90% on clean printed text.
Caveats
Less-common conjuncts may need HW-NE escalation.
Malayalam
ml- Script
- Malayalam
- Engine
- Surya
- Accuracy
- 85–90% on clean printed text.
Caveats
Atomic-character changes since 1971 reform — older documents may need HW-NE.
Odia
or- Script
- Bengali (Odia variant)
- Engine
- Surya
- Accuracy
- 83–88% on clean printed text.
Caveats
Shared model with Bengali; validate per use case.
Gujarati
gu- Script
- Gujarati
- Engine
- Surya
- Accuracy
- 85–90% on clean printed text.
Caveats
Validate on stamped/scanned government forms.
Persian
fa- Script
- Arabic (RTL)
- Engine
- Surya
- Accuracy
- 86–91% on clean printed RTL text.
Caveats
Shared model with Arabic + Persian-specific chars validated.
Urdu
ur- Script
- Arabic (Nastaliq)
- Engine
- Surya
- Accuracy
- 83–88% on clean printed Nastaliq text.
Caveats
Nastaliq is harder than naskh; cursive variants benefit from HW-NE.
Hebrew
he- Script
- Hebrew (RTL)
- Engine
- Surya
- Accuracy
- 85–90% on clean printed RTL text.
Caveats
Handwritten Hebrew → GPT-4o (no open-source handwriting model exists).
Thai
th- Script
- Thai
- Engine
- Surya
- Accuracy
- 85–90% on clean printed text.
Caveats
No word boundaries in Thai; OCR confidence calibration varies.
Russian
ru- Script
- Cyrillic
- Engine
- Tesseract
- Accuracy
- 95–97% on clean printed Cyrillic.
Caveats
Handwriting → Qwen3-VL.
Ukrainian
uk- Script
- Cyrillic
- Engine
- Tesseract
- Accuracy
- 94–96% on clean printed Cyrillic.
Caveats
Ukrainian-specific characters fully supported.
Bulgarian
bg- Script
- Cyrillic
- Engine
- Tesseract
- Accuracy
- 93–96% on clean printed Cyrillic.
Caveats
Validate on real corpus for production.
Greek
el- Script
- Greek
- Engine
- Tesseract
- Accuracy
- 93–96% on clean printed Greek.
Caveats
Modern Greek validated; polytonic ancient Greek experimental.
Tier · Experimental
Experimental support
Active development. Accuracy varies. The hardest scripts most tools fail on entirely — we ship them honestly labeled instead of overpromising.
Khmer
km- Script
- Khmer
- Engine
- Surya
- Accuracy
- 75–85% on clean printed text.
Caveats
Handwritten Khmer always flags for human review — no production-grade model exists.
Serbian
sr- Script
- Cyrillic
- Engine
- Tesseract
- Accuracy
- 90–93% on clean printed Cyrillic.
Caveats
Latin Serbian also supported via Latin pipeline.
How we set the tiers
Honest beats optimistic.
Stablemeans we have run thousands of synthetic and anonymized real-world documents through it, validated extraction accuracy against ground truth, and shipped it as production-ready. We'd use it ourselves in a regulated workflow.
Betameans the OCR engine and routing work, but the maturity isn't backed by exhaustive validation. We've tested it on clean documents. Your scan quality, document layout, and font choice might surface failures we haven't seen. Validate on a sample first.
Experimentalmeans it works enough to be useful, but accuracy varies meaningfully across documents. Indic scripts are the hardest cases in OCR — most tools refuse to ship them at all, or ship them with misleading "supported" labels. We ship them with honest expectations instead.
As accuracy improves, languages get promoted. Promotions are documented in the CHANGELOG with the validation work that supported them.
Language not on this list?
Inspire AI Lab has run extraction at scale on a 230M-document multilingual legal corpus. If you have a custom language requirement — fine-tuning, new script support, dialect handling — we can scope a custom build against your real corpus.