About this page
A chat that works in Tunisian Derja.
Baarcha builds speech and language models for Tunisian Arabic, and this page is where you talk to them. Type in Derja and the answer comes back in Derja. The model answering was trained on Tunisian text, not on Modern Standard Arabic asked to stand in for it.
Which model answers
Hannibal. It is a 27B model in the Barca H1 family, and in Derja it does chat, translation, diacritisation and writing in a style you hand it. It reads images and video as well. It does not hear: sound is Fennec's work. On a single stream it decodes about 111 tokens a second, which is why a long answer arrives while it is still being written.
Write the way you already write
Arabic script. Tunizi, which is Derja typed in Latin letters and numbers, as in a7kili 3la 9artaj. French. English. All of them in one sentence, which is how most people here write anyway. Hannibal reads every one of those.
The answer comes back in Derja, in Arabic script. That distinction is worth stating plainly: the model reads Tunizi, it does not write it, and this page does not pretend otherwise. So type Tunizi if that is your keyboard, and expect Arabic script back.
Talking to it, and hearing it back
The composer has a microphone. Press it and Fennec writes down what you said. It is a 1.7B speech model for Derja, at 23% word error and 9% character error on Tunisian speech, which is close enough that you can talk to it and be understood.
Under each answer there is a play control. Press it and Dido reads the answer out loud in a Tunisian voice. Dido can also copy a voice from one short recording, which is set up on its own page.
What it costs, and what we keep
Chatting is free while the models are in beta. It takes a Google account and one click. There is no card and no form to fill in. The account is there for two reasons: a browser that two people share should not show one of them the other's conversations, and a wrong answer needs a way back to us.
The conversations themselves stay in this browser rather than on our servers. What leaves it is what you choose to send: the message you want answered, and any reply you rate with a thumb, which reaches us with your verdict attached.
Questions people actually ask
Do I need an account to try the models?
For this chat, yes, and it is one click with Google. There is no card and no form. If you would rather not sign in at all, the playground runs the same three models without an account.
Can I write in Tunizi?
Yes. Tunizi is Derja typed in Latin letters and numbers, and Hannibal reads it: “a7kili 3la 9artaj” resolves the same way the sentence written in Arabic script does. The answer comes back in Arabic script. Reading Tunizi and writing it are two different abilities and only the first one is ours to claim.
What happens to French in the middle of a sentence?
It is read. Tunisians switch between Derja and French inside one sentence, so that is the ordinary case here rather than an exception. Fennec is trained for the French that Derja is code-switched with, and Hannibal reads a sentence that changes language halfway.
Where do my conversations go?
They stay in this browser. We do not keep them on a server. What does reach us is what you choose to send: your message goes to the model to be answered, and a reply you thumb up or down comes back to us with the verdict so we can read what went wrong.
Is it free?
The chat is free while the models are in beta. Building on the same models through the API is free during the beta as well, at 60 requests a minute.
Can Hannibal hear me?
No, and that is why there are three models rather than one. Hannibal reads and writes. Fennec does the hearing, at 23% word error and 9% character error on Tunisian speech. Dido does the speaking. The microphone and the play control in the chat are those two models at work.
The rest of it
- Hannibal, the model that answers here: Reads and writes Tunisian Arabic, and can look at an image or a video.
- Fennec, the one that listens, with the whole measurement table published on its page.
- Dido, the one that speaks: A Tunisian voice, not a Tunisian accent on someone else's.
- The playground runs all three with no account at all.
- The API docs, if you want these models inside something you are building.