Français
Tunisie
A historic step for open‑source AI in Arabic and Tunisian dialect!
Today, at LINAGORA, we are crossing a threshold that no one has dared to cross before:
a LLM that speaks, understands, and answers in Tunisian dialect, both in text and voice. A model built for Derja, with Derja, and for those who speak it.
What we are showcasing today:
- Chat Mode: ask a question in Tunisian and receive a textual response.
- Speak Mode: ask the question by voice, the model recognises the speech (ASR), generates the answer and renders it in speech‑to‑speech.
- Demo performed with the question: “What do you know about Tunisian culture?
شنوه تعرف على الثقافة التونسية ؟شنوه تعرف على الثقافة التونسية ؟
→ transcription + LLM answer + spoken answer.
Why it matters
- First speech‑to‑speech model dedicated to Tunisian dialect.
- Based on a complete pipeline: wav2vec ASR, Labess‑7B‑Chat LLM, Tunisian TTS.
- Labess‑7B‑Chat: a LLM trained specifically on Derja, launched 5 months ago.
- Implementation of a dedicated MMLU to evaluate the model in dialect.
- More versions are coming: larger data sets, higher vocal quality, reduced latency.
A concrete proof of open innovation
Let’s be honest: today’s major models do not speak your language. Literally. For reference, the current large LLMs such as LLAMA V2 are composed of:
- 90 % English,
- 0.17 % German,
- 0.16 % French,
- 0.13 % Spanish
In other words, the rest of the world carries very little weight.
At LINAGORA, we decided that a useful AI must speak the language of the people who use it – not the other way around. That is how we build a more equitable, freer, more human AI.
And now?
We continue collecting Tunisian data and improving the model.
Help us build tomorrow’s digital world! This AI is here for you, thanks to you.
What would you like it to learn or improve first?
Thanks to our collaborators: Wajdi Ghezaiel, Jean‑Pierre LORRE, Hedi Naouara, PhD Sarah ZRIBI LAKHOUA
