Little written data
Dialects are spoken far more than they are written, so the text large models learn from barely includes them.
Technology
Built for the languages big models skip.
Most language models learn from standard, edited text. Nobody talks like that. Linguini runs on Lebanese.ai’s patent-pending technology: one engine that understands a dialect however it’s typed or spoken, and answers the same way, even in three languages in one sentence.
Patent pendingUS provisional application 64/021,845
Move your cursor to turn the engine
The problem
Why dialects break ordinary AI.
Dialects are low-resource languages. Four things make them hard, and Linguini is built around all four.
No standard spelling
The same word is spelled many ways, in two scripts, with numbers standing in for sounds.
Languages mixed mid-sentence
Arabic, English and French switch inside one sentence. Most models break on the switch.
Speech first
Much of the dialect lives in voice notes and calls, not in typed text.
Step by step
From “shu 3am ta3mel?” to four answers.
Watch the approved demo sentence go through the engine. The Arabizi is read word by word, its numbers become the sounds they stand for, and the meaning comes back in four forms.
- Normalisation maps every spelling, script and sound to one form
- Dialect memory applies how the dialect is really used
- Generation answers in dialect, Modern Standard Arabic, English or French
1 · Typed in Arabizi
shu3amta3mel?
2 · Numbers read as sounds
3 → ع2 → ء5 → خ7 → ح
3 · One form the model reasons over
شوعمتعمل؟
4 · Answers in the form you need
Lebaneseشو عم تعمل؟
Modern Standard Arabicماذا تفعل؟
EnglishWhat are you doing?
FrenchQu’est-ce que tu fais ?
How it works
Four steps from what was said to what was meant.
- 1
Dialect input
Text or audio, exactly as people type it or say it: Arabic script, Arabizi, a mix, or speech.
shu3amta3mel - 2
Normalisation
Spellings, scripts and sounds are mapped to one form the model can reason over.
3am→عم - 3
Dialect memory
The model applies what it knows about how the dialect is really used, so answers stay in the user’s dialect.
dialectcontext - 4
Output
The answer comes back in dialect, Modern Standard Arabic, English or French, or all three languages in one sentence, as text or voice.
شو عم تعمل؟
What the engine does
Capabilities, not promises.
Everything here works today in Lebanese, the first live dialect. Other dialects are on the roadmap.
Two scripts, one language
Arabizi and Arabic script are read as the same words, so you never have to guess which one a user will type.
Three languages in one sentence
It follows Arabic, English and French as they switch mid-sentence, and it can answer the same way.
Text and audio
It understands spoken dialect and answers out loud, in dialect, as well as in writing.
Dialect memory
Knowledge of how the dialect is really used keeps answers natural, not translated.
Four outputs from one input
The same message can come back in dialect, Modern Standard Arabic, English or French.
One engine, many dialects
Built so the next dialect runs on the same engine. Gulf, Egyptian and Moroccan are on the roadmap.
Hear it
Real recordings, real dialect.
Two recordings from a voice assistant built on the engine, taking a shop order. Listen to it switch between Lebanese and English inside one sentence, the way people here talk.
Recording 1 · confirming an order
Okay، بدّك ٣ تي شيرتات size medium، بس عم لخّصلك طلبك وما رح أعمل طلبية.
“Okay, you want 3 T-shirts, size medium. I’m just summarising your order, I won’t place it yet.”
Recording 2 · checking before ordering
شفت إنّك بدّك ٣ تي شيرتات... المقاس medium، وما عملت أي طلبية.
“I see you want 3 T-shirts… size medium, and I haven’t placed any order.”
Intellectual property
Invented here. Patent pending.
The engine is Lebanese.ai’s patent-pending technology, invented and built by Code-X in Beirut by people who speak the dialect every day.
US provisional 64/021,845Filed with the USPTO
Backed byLAU Innovate (FMIC SPARK), the Google AI program, and a grant from Anthropic
Built by Code-XBeirut, Lebanon
Put the engine to work.
One API for text and audio, in developer preview.