silma-tts-derja — Tunisian Derja speech synthesis
An F5-TTS fine-tune that speaks Tunisian Derja, the dialect Tunisians actually use in conversation.
Why this exists
Every off-the-shelf Arabic TTS produces Modern Standard Arabic — a formal register nobody speaks conversationally. A system that answers a nervous Tunisian user in MSA sounds like a news broadcast, which is close to the opposite of the intended effect.
No usable Derja voice existed, so this one was fine-tuned. It also handles the Derja/French code-switching that is normal in Tunisian speech, rather than suppressing it.
Intended use
Conversational applications aimed at Tunisian users: voice assistants, coaching and support systems, accessibility tools, educational software.
Built for and deployed in Solace, a real-time multimodal AI coaching system, where it runs locally and in-process so that user audio never leaves the machine.
Usage
Limitations
- The shipped reference clip is an English sample. F5-TTS conditions output prosody on a reference recording, so Derja prosody would improve measurably with a native Derja reference clip. This is the largest known weakness.
- Trained on a limited Derja corpus. Coverage of rare vocabulary and of regional variation within Tunisia is uneven.
- No formal MOS evaluation has been run; quality claims here are qualitative.
- Synthesis is CPU-bound in the deployment it was built for; a GPU roughly halves latency.
Training
Fine-tuned from SWivid/F5-TTS on Tunisian Derja speech.
Related work
| Model | Purpose |
|---|---|
Ghazouaniwala/emotions_speech | Speaker-independent speech emotion recognition — the other half of Solace's speech stack |
Ghazouaniwala/trocr-tunisian-arabic | Handwritten Tunisian Arabic recognition |
Citation
@misc{ghazouani2025silmattsderja,
author = {Ghazouani, Wala Eddine},
title = {silma-tts-derja: Tunisian Derja speech synthesis},
year = {2025},
url = {https://huggingface.co/Ghazouaniwala/silma-tts-derja}
}
Built by Wala Eddine Ghazouani · GitHub · LinkedIn