arabizikit

Transliterate Arabizi (Romanized Arabic) to Arabic script in the browser or Node, with dialect tagging and ranked candidates. Zero dependencies.

Arabizi is how millions of Arabs text: "ana 3ayz 2akol" means "أنا عايز آكل". This package turns that back into Arabic script, handles the ambiguity honestly (2 can be hamza or qaf, ay can be two words or one), tags the dialect, and returns ranked candidates instead of one guess.

const { transliterate } = require("arabizikit");

const res = transliterate("ana 3ayz 2akol", { topK: 3, withDialect: true });
console.log(res.text);            // أنا عايز آكل
console.log(res.candidates);      // [{ ar: "أنا عايز آكل", score: 0 }, ...]
console.log(res.dialect.dialect); // egyptian

API

• transliterate(text, options?) returns { text, candidates, dialect, evidence }
  • options.topK number of ranked candidates (default 3)
  • options.withDialect include the dialect guess (default false)
  • options.dialectHint assume a convention: egyptian, levantine, gulf, maghrebi
• tokenize(text) split into space, latin, and other tokens
• processWord(word, dialectHint?) the ranked readings for one word

Browser

The package ships a single file, arabizikit.js, which attaches window.ArabiziKit. Load it with a script tag, or bundle it with any module system.

<script src="arabizikit.js"></script>
<script>
  const res = ArabiziKit.transliterate("shlonak ya 5al");
  console.log(res.text);
</script>

Try the live demo: https://arabizi-kit.vercel.app

Notes

The tables are the single source of truth: this bundle is generated from the Python package arabizikit (github.com/rb2625/arabizi-kit), which adds a learned layer (dialect classifier, word reading table, language-model reranking) and an LLM-assisted mode on top of the same engine.

MIT license.