Reading and re-inking historical German Kurrent script via ductus-model template fitting on scans.
→ Try it live at kurrentschrift.ink
The live site is organised into three areas — Schriftkunde (reference), Lesen (reading) and Schreiben (writing):
- Script overview (Schriftkunde) — a compact, fully sourced introduction to the German cursive scripts (Kurrent · Sütterlin · Offenbacher), their nibs, ink and letter quirks.
- Letter quiz — learn to read the alphabet one glyph at a time.
- Writing board (Tafel) — specimen rows on ruled lines, written stroke by stroke as a tracing model.
- Practice-sheet generator — printable worksheets with configurable ruling (Lineatur) and your own text.
- Live writer (Federprobe) — type any word or sentence and watch the synthesised Sütterlin ductus write it, connecting strokes included.
Everything below is the engine being built behind them.
Two problems live under "historical German handwriting":
- Read it — already solved. Transkribus ships public Kurrent models at ~5–7% CER. Integration, not research.
- Write it — so the output looks like ink, not a font. No off-the-shelf product exists. That's the project.
The hard part of writing is the crossing problem: from a still image of cursive script, the skeleton alone can't tell which stroke passed under which at a loop crossing. OpenType fonts fake the connections and never look inked; ML stroke-synthesis needs online pen-path data, but historical material is image-only.
The approach — analysis-by-synthesis with a ductus prior. The image supplies geometry and ink width; the ductus (stroke order and pen-lift points, known a priori for normed pre-1900 Kurrent) supplies the missing dimension. A canonical ductus template is fitted to the scan's skeleton and distance-transform profile. The library unit is (glyph, position, variant) — medial ſ and final s are different glyphs, not one s with a context switch — and transitions fall out of exit/entry tangents, so there is no bigram explosion.
The open research question is how tight to make the template: too rigid won't fit real hands; too loose and the crossing ambiguity returns. Finding that trade-off empirically is the project's research kernel.
In-progress MVP.
- Live on kurrentschrift.ink: the script overview (
/schriftkunde), the letter quiz and the writing board under Lesen, and the worksheet generator and the Federprobe live writer under Schreiben. - Built: the admin setup wizard, canonical glyph extraction, the template-to-instance fit routine, and server-side word composition — text → glyph sequence → placed glyphs with generated connecting strokes (
core/shaping.py+core/compose.py, served asGET …/write/word, scored against period word specimens by thetools/wordbenchbenchmark). - Next milestone: scan the author's own hand for a small lowercase alphabet (
a d e l n ſ s) plus a few words, fit the canonical ductus per glyph, aggregate, then re-render those words and at least one new word in the same hand. Four validation gates — stability, allograph separation, word rendering, animation playback — decide "kernel validated" vs. a fast, useful negative result. See the §8 MVP and the MVP roadmap.
Beyond the kernel, the roadmap sequences reading help (HTR), style analysis, hand comparison, and a possible open-data release — see architektur.md §10.
kurrentschrift/
├── core/ # Pure-Python compute + DB layer (extractor, template, Postgres models)
├── api/ # FastAPI service (thin routing over /core)
├── app/ # React 19 + Vite + MUI SPA — public site + admin behind /admin/*
├── alembic/ # Schema migrations
├── data/ # Sources, variants, samples (separate licensing — see below)
└── docs/ # Design rationale (German); references with technical specs
uv run alembic upgrade head # once, or when the schema changes
uv run uvicorn api.main:app --reload --port 8000 # Python backend on :8000
cd app && npm install && npm run dev # Vite dev server on :3000 (/api proxy)Open http://localhost:3000. The admin lives under /admin (Cloudflare Access in production). Point DATABASE_URL (see .env.example) at any empty PostgreSQL database — alembic upgrade head creates the schema and seeds the base styles. For local admin saves, set ADMIN_TOKEN=<x> on the API and a matching VITE_ADMIN_TOKEN=<x> in app/.env (without them, write requests return 401).
Design rationale lives in docs/ — start at docs/index.md. Highlights: the vision, the architecture reference (§1–§17), the MVP roadmap, and reference specs for HTR integration, animation rendering, style analysis, and the frontend stack. Internal docs are German (the domain is German); code, commits, and this README are English.
Code is MIT. Data is not. Each source under /data/sources/ carries its own license (see its SOURCE.md); corpora live outside git; NC-SA-derived materials never reach committed outputs. The first source is the Loth 1866 Kurrent table (Public Domain Mark 1.0, via Wikimedia Commons) — the geometry baseline for the MVP. Full index in DATA_PROVENANCE.md.
This is an in-progress MVP portfolio project — issues (including discussion of the approach) are welcome; external PRs are premature until the four MVP gates land. See the Contributing Guide for what's useful to send right now.
Open-core. The code is MIT — see LICENSE. The learned data this project produces — the curated glyph templates, the ductus, and the script statistics / trained reading models built in the admin — is not covered by MIT and stays reserved, even with the repo public; reuse by arrangement only. The public-domain sources under /data/ keep their own free license. If the methodology is useful to your work, see CITATION.cff.
Built by Markus Neusinger