Genealogy Document Scanning & Handwriting OCR (HTR)

The handwritten-text-recognition (HTR) leap of 2023–2026 changed genealogy. You can now feed a 17th-century parish register to a model and get readable text. For storing and tagging the resulting scans use Self-Hosted Document Management (Paperless-ngx); for printed-text OCR see OCR & Vision; for translating the recognized text see Genealogy Translation & Paleography; for restoring photographed pages see Genealogy Photo Restoration AI.

Handwritten Text Recognition (HTR) — the 2024–2026 revolution

• ★ Transkribus — paid + generous free tier (500 free credits ~= 500 pages/month historically). University of Innsbruck spin-off; the dominant HTR platform. Pre-trained models for German Kurrent, English secretary hand, Latin, Spanish, French, Russian, Hebrew, Arabic, etc. Can train your own model on 50–100 transcribed pages of a difficult hand. Lifted "stuck reading parish registers" from a year-of-Latin-paleography skill to "upload, click." Used by national archives across Europe. Web app + desktop client.
  • https://transkribus.eu/
• ★ eScriptorium — AGPL, FOSS Django app; the open-source equivalent. Built on Kraken HTR engine. Self-host or use the public hosted instances (universities run free ones). Pretrained models on Hugging Face / shared via the eScriptorium hub. Less polished than Transkribus but no per-page fees.
  • https://gitlab.com/scripta/escriptorium
• Kraken — BSD; the underlying OCR/HTR engine; CLI use for pipelines. Pairs with eScriptorium.
• READ-COOP / Read & Search — Transkribus's cooperative model; some content free for searching transcribed European archives.
• CITlab HTR+ — research-grade; mostly subsumed by Transkribus.
• Calfa Vision — paid; oriental scripts (Armenian, Syriac, Arabic) HTR specialist.

Printed-text OCR (typed records, indexes, gazetteers)

• ★ Tesseract — Apache 2.0; the FOSS workhorse. v5.x with LSTM models is competent on clean print. Weak on cursive / damaged scans. Languages installable per-locale.
• OCRmyPDF — MIT; wraps Tesseract to add a searchable text layer to scanned PDFs in place. Pair with Paperless-ngx.
• PaddleOCR — Apache 2.0; great on multilingual including CJK; useful for Asian record-set indexing.
• Surya — newer FOSS; document AI including layout + reading order on top of OCR. See OCR & Vision.
• ABBYY FineReader — paid Windows/Mac (~$200); historically the best commercial OCR for old printed books. Strong on layout reconstruction.
• Adobe Scan / Adobe Acrobat OCR — included with Acrobat Pro subscription; convenient for one-off scans.
• CamScanner, Microsoft Lens — free + paid mobile scanners; good auto-deskew, OCR included.

LLM-as-OCR (the 2024–2026 sleeper)

• Claude / GPT-4o / Gemini Flash with vision — paid per-token; surprisingly good on damaged or unusual hands when given context ("This is a 1750 Latin Catholic parish register from Bavaria"). Cheaper than Transkribus credits at low volumes; quality variable on long pages. Build a workflow: split into pages → vision API → review.
• Local LLM vision models (LLaVA, Qwen-VL, InternVL, MiniCPM-V) — see Self-Hosted AI / LLM. Privacy-respecting; quality below cloud frontier models in 2026 but improving.

Mobile scanning

• ★ Microsoft Lens — free; auto-deskew, multi-page PDF, lightweight OCR, exports to OneDrive / OneNote / PDF.
• ★ Adobe Scan — free + paid; better OCR; ad-driven free tier.
• Genius Scan — paid (~$8 one-time); no subscription. Workflow-friendly.
• Scanner Pro (Readdle) — paid iOS; long-running; OCR built-in.
• Apple Notes / Files document scanner — free, built-in iOS / iPadOS; surprisingly good auto-crop.
• Google Drive scanner — free; built into Drive Android app.
• CamScanner — free + paid; functional but ads heavy and historical privacy concerns.

Flatbed / archival scanning

• Epson Perfection V600 / V850 — paid hardware; archival flatbed for fragile documents and photo originals.
• Fujitsu ScanSnap iX1600 — paid; production document scanner for inherited family papers.
• CZUR / Fujitsu ScanSnap SV600 — paid overhead book scanners; preserve binding of fragile heirlooms.
• DIY rig — DSLR + copy stand + Lightroom + Photoshop tethered. Highest quality for irreplaceable items.

Storage / archive workflow

• ★ Paperless-ngx as the family DMS — see Self-Hosted Document Management. Set up tags: vital-record:birth, vital-record:marriage, correspondent:Maternal-Grandmother, event:1880-Census. Paperless OCRs incoming via Tesseract; for handwritten material, run Transkribus first and pre-pend the transcription as text.
• Folder convention — Archive/{surname}/{individual-key}/{record-type-YYYY-MM-DD}.pdf. Paperless tags overlay; the folder structure is the durable backup.
• Link DMS to your tree — webtrees and Gramps Web both support media URLs. Paperless emits stable per-document URLs; embed in source citations.

What's changing in 2024–2026

• Transkribus public models for major European archive scripts pushing transcription accuracy past 95% on legible hands.
• FamilySearch Full-Text Search rolled HTR-based search across handwritten US deeds and probates — searchable for the first time, transformative for "find this name in any deed book."
• Antenati (Italy) expansion — new HTR-indexed parish registers added monthly; see Genealogy By Country (Europe).
• LLM vision OCR catching up to Transkribus for general-purpose hands; still behind on extreme cursive but useful for $0.01/page on Anthropic / OpenAI / Google.
• eScriptorium maturing as the FOSS escape hatch; running on a Hetzner box with a $5 GPU rental gets useful throughput.

Pick this if…

• Best HTR, low volume: Transkribus free tier.
• High-volume HTR, control: Transkribus subscription or self-hosted eScriptorium.
• Printed-text OCR, free: Tesseract via OCRmyPDF, indexed in Paperless-ngx.
• One-off page from a tricky hand: Claude / GPT-4o vision API.
• Mobile scanning while at the courthouse / archive: Microsoft Lens or Adobe Scan.
• Family papers archive at home: ScanSnap iX1600 → Paperless-ngx consume folder.
• Fragile heirloom album / Bible: flatbed (V600) or overhead book scanner.