Euspell

Convert English text into euspell — a reformed, more regular spelling — everywhere you read and write, entirely on your device. No account, no server, no telemetry: the lexicon ships inside each product and your text never leaves it. (One exception, and only while you use it: dictation hands the microphone to the browser's own speech recognition, which on Chrome transcribes in the cloud. See docs/dictation.md.)

The flagship is a browser extension (Chrome, Firefox, and Safari) that rewrites web pages and PDFs in place, but the same conversion engine drives a family of integrations: a bundled PDF viewer, a LibreOffice extension, a Microsoft Word add-in, a Google Docs script, an Apple Pages converter, downloadable dictionaries for other spell-checkers, and the sibling Eupub EPUB/PDF reader. One engine, many surfaces — none of them re-implement the reform.

Installing it: the browser extension is covered by docs/installing.md; the four word-processor tools by docs/installing-addins.md.

What it runs on

How the reform works

Euspell is not a find-and-replace. Each word is looked up in a large English lexicon; most words are unchanged, and the reformed minority each carry their new spelling plus part-of-speech tags. The tags matter because English is full of homographs whose spelling should reform differently by sense:

• records is NN2|VVZ — the noun ("the records") and the verb ("she recordz") take different euspellings, chosen based on the surrounding words using SVM.
• wind, bow, read, lead, bass, beloved, … each split by meaning.

Each entry also carries a three-digit encoding saying how many spellings the word has and what kind of ambiguity, if any, has to be resolved to choose between them — see docs/encoding.md for the full scheme.

So the engine (src/content/converter.js + dom-walker.js) walks a document's text nodes, tokenizes and tags them (src/content/tagger.js), and resolves the ambiguous cases with two layers of disambiguation in src/disambig/: a part-of-speech SVM/rule set (pos.js, vvz-svm.js) and per-word semantic rules (semantic/), trained against the corpora under disambig/. Only text node values change — never structure — so links, layout, selection, and search survive the rewrite.

Build & load the extension

npm install
npm run build        # compiles the lexicon, copies PDF.js, bundles the extension → dist/

Then load it unpacked:

• Chrome: chrome://extensions → enable Developer mode → Load unpacked → pick this folder (it reads manifest.json and dist/).
• Firefox: npm run build:firefox, then npm run run:firefox (web-ext) or load build/firefox/ via about:debugging.
• Safari (macOS): npm run build:safari, then open safari/Euspell.xcodeproj in Xcode and Run. See safari/README.md for enabling it in Safari.

dist/ is generated (git-ignored); rebuild after changing anything under src/ or the lexicon data.

Integrations & dictionary exports

The same lexicon feeds a set of generators (build/), each producing a drop-in artifact for another tool:

npm run gen:lo        # LibreOffice extension (.oxt)
npm run gen:word      # Word Office.js add-in (engine + data + icons)
npm run gen:oxt       # Hunspell dictionary packaged as .oxt
npm run gen:harper    # Harper grammar-checker metadata
npm run gen:pls       # TTS pronunciation lexicon (.pls)
npm run gen:pos       # part-of-speech lexicon (for external grammar checkers)

Outputs land in dict/. A custom Word dictionary (dict/euspell-word.dic, the reformed spellings, so Word stops flagging them) comes from npm run gen:word-dict.

Project layout

Tests

npm test        # node --test over tests/**/*.test.js

License & privacy

GPL-3.0-or-later. Copyright (C) 2026 Kamran Ossia. The full licence text is in LICENSE — verbatim, so that a copy travels with every distribution rather than only being pointed at.

The lexicon data is licensed separately. The files under data/*.csv are Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0), not GPL. The reform is meant to be reusable by anyone building on it, including outside a GPL project; the code that applies it is not.

All conversion runs locally — no page text, no telemetry, and no network requests leave the device.

Dictation is the single exception, and worth stating plainly because the claim above is otherwise absolute: it calls the browser's Web Speech API, and Chrome implements that by streaming the audio to Google's speech service. The extension never receives the audio and never sees the transcript before the browser returns it; the euspell conversion still happens locally. Nothing is captured unless you start dictation.