Multilanguage On-Device Translator

Links: live demo (lite edition) and GitHub repository

Lingua translates between 61 Indian and world languages, built to run on your own hardware first and to fall back to free services only when it has to. Four things separate it from a thin wrapper around a translation API.

It reads romanized Indic text

mera naam rahul hai is Hindi typed in Latin letters. Lingua detects that, converts it to मेरा नाम राहुल है, shows you what it understood, and translates from there. Fixing the script first is not cosmetic, and measured against Google’s own endpoint it is strictly better:

That last row is the exact example Google’s own UI returns for sl=as&tl=en.

It prefers local weights

IndicTrans2 (MIT) for English to and from 22 Indian languages and Indic to Indic, NLLB-200 for everything else, and a two-hop English pivot for Indic to foreign, which measurably beats NLLB direct. All run on a 4 GB laptop GPU or CPU only. Every request walks a tier ladder and the response records which rung answered:

Tier 0  LOCAL      IndicTrans2 · pivot_en · NLLB-200   open weights, on your machine
Tier 1  FREE_WEB   Google gtx · Google clients5        keyless, no account
Tier 2  FREE_API   MyMemory · LibreTranslate           free tiers, documented
Tier 3  FREEMIUM   only if you configure a key         off by default

The lowest tier that can serve the pair wins. Capping it per request (max_tier: "LOCAL") gives a strictly offline translation or an error.

Telling apart languages that share a script

Devanagari carries Hindi, Marathi, Nepali, Sanskrit, Maithili, Bodo, Dogri and Konkani; Bengali script carries Bengali and Assamese; Arabic script carries Urdu, Sindhi and Kashmiri. Assuming the most common one is not a rounding error, it changes meaning. माझे नाव राहुल आहे read as Hindi becomes “My boat is Rahul” and read as Marathi becomes “My name is Rahul”, because नाव is boat in Hindi and name in Marathi.

Disambiguation uses two signals. Signature characters are decisive: Assamese writes ৰ and ৱ where Bengali writes র and ব, and Sindhi and Kashmiri each have letters Urdu never uses. Otherwise it falls back to function-word lexicons, copulas and pronouns, because they appear in nearly every sentence and rarely cross over. It reports honest confidence and defers to a provider that can detect for itself below a threshold. 11 of 11 languages are now identified correctly, offline, with no extra dependency; before this, 5 of 6 Devanagari languages were silently read as Hindi.

It checks its own output

Small models garble proper nouns across scripts. NLLB-600M renders “Rahul” into Japanese as ラहुल, Devanagari spliced into katakana, and does it even when handed clean English input, so no amount of routing fixes it. A sanity check catches it: a translation whose target script is Japanese should not contain Devanagari. Latin is always allowed, since names, acronyms and loanwords appear in Latin letters in every language, Japanese may mix kana with Han, Korean may mix Hangul with Han, and anything under 2% of characters is treated as incidental. A wrong-script result falls through to an engine that gets it right rather than being served.

It can also estimate its own accuracy. Ask for a check and the translation is fed back through a different engine and scored with chrF, an F-score over character n-grams chosen because it needs no tokeniser, which matters when the languages include Japanese and Tamil. best_of: 3 runs three engines and keeps the highest scorer, with one grader scoring every candidate, because grading each with a different engine is not a comparison.

Coverage, measured rather than assumed

Speech-to-text chains two engines. Whisper is the only one that can identify an unknown language, so it always runs first on auto-detect; IndicConformer-600M covers all 22 scheduled languages, including the eight Whisper cannot transcribe at all, and is more accurate on the rest. Text-to-speech chains three, with MMS-TTS running fully offline and providing the only voice that exists for Assamese, Odia and Maithili. Eight languages have no voice anywhere public, which is a gap in what has been built in the world rather than in the app.

Things that were in the way

Google’s free endpoint fingerprints the TLS handshake, not the headers. It returns HTTP 429 to httpx and requests even with browser-identical headers, while curl with the same headers gets 200. Header spoofing does not fix it, which is the single thing most tutorials get wrong. The fix is curl_cffi, which presents a real browser handshake and has wheels for both win_amd64 and manylinux aarch64.

transformers 5 breaks IndicTrans2 at import. It moved PreTrainedTokenizerBase out of transformers.tokenization_utils, and both IndicTransToolkit and IndicTrans2’s own modelling file still import it from there, so the failure happens before any download. The requirements pin transformers>=4.44,<5.

The engine knows nothing about HTTP: the full edition and the network-only lite edition are both thin shells over the same Router, and adding a provider is one class and one line.