New Ithkuil script converter

Encode · text → script

Renders as you type. No server, no round trip — the layout engine runs client-side.

compact layout (collision kerning)

Compact layout needs SVG hit-testing (isPointInStroke), which browsers have and Node does not — the repo ships a geometric shim so the same code path works headlessly.

Decode · script image → text

The reverse direction is the part no other tool does — and the reason this page can't run it. It needs ~15 MB of cached glyph templates and five TensorFlow CNNs, so it stays on the Node side. Here is what it produces.

Four panels: romanized text, rendered script, segmentation overlay, decoded text — One full round trip: text → script → segmented characters (magenta = character, green = base, blue/orange = super/underposed diacritics) → text recovered by the decoder.

Benchmark	What it measures	Result
Real-lexicon round trip	4,387 real roots, sampled by length and weighted by frequency	92.6%
…the same, on template matching alone	a fresh clone, before the optional CNNs are trained	42.6%
Alphabetic register	phonetically spelt words, character level	100%
Printed & re-scanned	paper sheets recaptured on a flatbed and two phones	78.1%

Numbers are reproducible from the repo: npm run lexicon-test, npm run alphabetic, npm run scan-test.

How it works

The forward renderer doubles as the labelling machine: every glyph class is rendered and augmented, which is where the training and template data for the reverse pipeline comes from.

Forward is a solved problem and is reused from @zsnout/ithkuil (MIT). Reverse — segmentation, character typing, per-slot classification and reassembly into a valid word — is original work in this repository.