@johnhenry/langdetect-lite

@johnhenry/langdetect-lite

Script and Latin-language detection with no model and no data files: a faithful port of laya-mlx laya_mlx/lang.py (from Laya, Apache-2.0). It answers one routing question: can an English-only checkpoint read this text?

Install

npm install @johnhenry/langdetect-lite
bun add @johnhenry/langdetect-lite
deno add jsr:@johnhenry/langdetect-lite

Zero runtime dependencies beyond @johnhenry/pyjson; Node ≥ 24, Bun ≥ 1.2, Deno and browsers.

import { analyse, isEnglish, detectScript, guessLatinLanguage } from "@johnhenry/langdetect-lite";

detectScript("客户被重复扣款要求退款");        // "han"
isEnglish("Please refund invoice 4411");         // true
isEnglish("Gătește-mi o rețetă de sarmale");     // false (Romanian diacritics)
guessLatinLanguage("Le client a ete facture deux fois et il demande un remboursement pour la facture"); // "fr"
analyse({ subject: "Refund", body: "Der Kunde wurde zweimal belastet und nicht erstattet" });
// { script: "latin", script_profile: { latin: 1 }, language: "de", is_english: false,
//   language_undecided: false, diacritic_rate: 0, non_latin_fraction: 0 }

API

All functions accept a state: a string, or any JSON-like value whose string leaves (object values and array items, up to depth 6) are joined with spaces. Keys are ignored.

Export	Python (laya_mlx.lang)
stateText(state, maxChars = 4000)	state_text
detectScript(text) → "latin" | "han" | ... | "unknown"	detect_script
scriptProfile(text) → { script: fraction }	script_profile
latinProfile(text) → { language, english_hits, diacritic_rate, looks_non_english }	latin_profile
guessLatinLanguage(text) → "en" | "fr" | "de" | "es" | "pt" | "it" | "nl" | "ro" | null	guess_latin_language
analyse(state) → Analysis	analyse
isEnglish(state)	is_english
SCRIPT_RANGES, STOPWORDS, NON_EN_DIACRITICS, NON_EN_DIACRITIC_RATE	module constants

snake_case aliases (state_text, detect_script, ...) are exported too. Result objects keep Python's snake_case keys so they compare equal to the Python output; every field of the fixture table (lang.json, 35 states) matches exactly, floats included (rounding uses pyRound from @johnhenry/pyjson, the only dependency).

Python semantics are reproduced where JavaScript differs: iteration and truncation by code point (not UTF-16 unit), str.isalpha() as Unicode \p{L}, and re [^\W\d_]+ as [\p{L}\p{Nl}\p{No}]+ (letters plus non-decimal numerics; combining marks split words, as in Python).

Limitations

• The Latin-language guess is a stopword/diacritic heuristic for 8 languages. Short input usually returns null (undecided) on purpose; Romanian without diacritics that contains an English function word reads as English. Pass an explicit language when you know it.
• Unicode tables come from the JavaScript engine, not from Python's unicodedata; the two can disagree on characters added in different Unicode versions.
• A state object's leaf order follows JavaScript key order (integer-like keys first), which only matters for the 4000-character truncation.

Family

Part of laya-js, Laya typed decisions in JavaScript on MLX, WebGPU and CPU — see its package map and results.

• @johnhenry/laya-router calls analyse() to choose between the English and multilingual checkpoints.
• Uses @johnhenry/pyjson for Python-identical number formatting in its reasons.

License

Apache-2.0. Ports logic from laya-mlx and Laya (both Apache-2.0); see NOTICE.