The problem
Arabic is not one language, and every assistant treats it like one.
Arabic speakers grow up in two languages at once. There is the formal one, Modern Standard Arabic, which is what gets written, taught and broadcast. And there is the one people are actually raised in: Egyptian, Saudi, Lebanese, Jordanian and dozens more, each with its own grammar, humour and rhythm.
Nobody speaks the formal one at home. When an assistant answers in it, it is technically correct and socially wrong, in the way a stranger reading from a script is correct. That gap is why voice AI has barely touched a market this size.
One greeting, five ways
The scale of it
0 people speak Arabic. It is among the most spoken languages on earth, and one of the least served by conversational AI.
Arabic is official
the app today
Nobody's mother tongue
way of messaging
The solution
Pick who you are talking to. Not which model.
Every other assistant opens by asking which engine you want. Lyceum opens by asking which voice you want to hear. You choose a companion with a name, a dialect and a real voice, and then you talk. It talks back, out loud, the way you would actually be answered.
Dialect is the product
Six companions, each locked to a dialect and a voice. Not a translation layer bolted onto an English assistant, and not one Arabic setting standing in for a region of very different places.
Voice, not a text box
A full realtime call you can interrupt mid-sentence, with live captions. In a region raised on voice notes, speaking is the interface and typing is the fallback.
It does the work too
It reads documents, watches video, searches, and generates images and video, all inside the same conversation. The companion becomes the way into everything else.
Why now
This was not buildable two years ago.
Dialect-quality Arabic speech, in both directions, only became good enough very recently, and the cost of a spoken turn fell sharply at the same time. That combination is what turns an interesting demo into something a person can afford to use every day.
We build entirely on Google: Gemini for reasoning and for understanding images, documents and video, Google speech for Arabic recognition and synthesis, Imagen and Veo for generated media, and Google Cloud underneath all of it.
Where it stands
Built, running, and in testers' hands.
Not a deck. The app runs on real phones against a live production backend, and everything described above works today.
on Google Cloud
testers now
public launch
not a localization pass