Research
Three models. Three ways to work in Tounsi.
Which model should you try? A practical introduction to Hannibal for language, Fennec for transcription and Dido for speech, with a playground for each.
Hannibal: start with words
Hannibal is the language model in the family. Use it for conversation, writing, translation and working through an idea in Tunisian Derja. It understands Tunizi as well as Arabic-script Derja and French mixed into the conversation.
Try a small task you can judge: ask for a customer message, a simpler explanation or a rewrite in another tone. In the playground’s chat mode, you can change the prompt and inspect the request behind the result.
Fennec: turn a recording into text
Fennec handles transcription: spoken Tunisian comes in, written text comes out. It is designed for Tunisian speech and French code-switching. Transcription is Fennec’s job, separate from Hannibal’s language work.
Start with a short recording whose words you know. Read the transcript while listening to the original, paying attention to names, numbers and words that cross languages. Our published transcription benchmark includes the evaluation setup and spelling analysis so you can interpret the results in context.
Dido: give the words a voice
Dido takes text and produces Tunisian speech. Try a short sentence in text-to-speech mode, choose an available voice and listen to how the wording comes through. The voice choices come from the current service rather than a fixed list in this article.
Use the same sentence when comparing voices. Listen for pronunciation, rhythm and whether important words remain clear. A sample you like is a starting point; listen to the full text you plan to use before sharing it.
Try it, inspect it, then connect it
The playground gives you a place to test these capabilities before building an integration. Check the displayed availability and mode: a page marked Demo uses recorded examples rather than generating a fresh model result.
When you are ready to integrate, the documentation covers chat, transcription and speech endpoints. API usage draws from a separate API credit balance; a chat subscription does not include that credit. Start with a small test that reflects your real task, then expand from what you learn.