Introduction
Our model is optimized to understand and speak the Tunisian dialect, but it is first and foremost your voice assistant. It adapts to your way of speaking, your regional accent, and your expressions. It’s not up to us to tell you how to talk to your assistant.
This report presents IyedeX’s approach to building the first TTS/STT voice model specialized in the Tunisian dialect, based on an authentic corpus:
- open-licensed Tunisian radio and television archives;
- legally available public recordings;
- our own internal voices;
- voices of collaborators from the North, Center, and South of Tunisia.
This is complemented by a complete pipeline for audio cleaning, annotation, and preparation, followed by fine-tuning the xTTS v2 model to achieve truly Tunisian pronunciation. The objective: natural synthesis and robust recognition, integrated into 3amAli, IyedeX’s assistant.
Our vision is to create a voice model that captures the authenticity and richness of the Tunisian dialect in all its regional diversity.
The Corpus
The corpus was designed to reflect the real diversity of the Tunisian dialect. Here is the public version of its statistics.
| Source | Type | Region | Speakers | Hours | License | Cleaning | SNR (dB) | Notes |
|---|---|---|---|---|---|---|---|---|
| Broadcast A | Radio | North | 12 | 40.0 | CC-BY | 85 % | 18.5 | Morning shows, debates |
| Broadcast B | TV Audio | Center | 8 | 15.0 | CC-BY-SA | 80 % | 17.2 | Reports |
| IyedeX Recordings | Personal | North | 1 | 5.0 | Consent | 95 % | 28.0 | Amateur studio |
| Friends’ Voices | Personal | Mixed | 14 | 20.0 | Consent | 90 % | 20.1 | North / South variety |
| Field Recordings | Field | South | 6 | 12.5 | Consent | 70 % | 12.2 | Noisy environments |
| Total | 41 | 87.5 |
More than half of the hours come from radio and television, nearly a third from voices recorded with their consent, and the rest from the field, where noise is real.
By region, the North accounts for 45 hours, the Center for 15 hours, and the South for 12.5 hours. The remaining 20 hours mix speakers from several regions.
Data Preparation
Each hour of audio goes through the same five steps before being used for training.
- Normalization. Conversion to mono 16 kHz or 24 kHz, 16-bit PCM, and standardization of source files.
- Cleaning and Segmentation. Removal of silences longer than 500 ms, segmentation into 1 to 12-second segments, automatic noise filtering.
- Annotation. Manual and semi-automatic transcription, standardized dialect spelling, regional tags (North, Center, South).
- Augmentation. Pitch variation of ±5%, time stretching, controlled noise addition (markets, cars, music).
- Final Split. 80% training, 10% validation, 10% testing, with strict speaker separation between the three.
The Architecture
The model relies on two complementary components: xTTS v2, fine-tuned for speech synthesis, and Wav2Vec2, fine-tuned for recognition.
| Before fine-tuning | After fine-tuning | |
|---|---|---|
| Model | Generic, multilingual | Tunisian specialized |
| Pronunciation | Approximate | Authentic |
| Accent | Non-Tunisian | Natural Tunisian |
| Local expressions | Poorly understood | Understood |
Speech Synthesis (TTS): Fine-tuned xTTS v2
- a multilingual model adapted to Tunisian;
- learning Tunisian prosody;
- multiple speakers;
- style and emotion control.
Speech Recognition (STT): Fine-tuned Wav2Vec2
- Wav2Vec2 as a base;
- comprehensive fine-tuning on our corpus;
- an error rate optimized for real-world conditions.
Doing a Lot with Little
Our team worked on modest machines, without high-end GPUs. We therefore focused on engineering rather than computational power.
Innovation does not depend on hardware power. Our results show that the quality of the dataset and engineering matters more than raw power: we built a high-performing Tunisian voice model with limited resources.
Results
We evaluate the model using standard international metrics: WER (word error rate) and CER (character error rate) for recognition, MOS (mean opinion score) for synthesis, and, optionally, STOI and PESQ, complemented by phonetic error maps.
Speech Recognition
| Model | WER, quiet environment | WER, noisy environment |
|---|---|---|
| Base Wav2Vec2 | 35.2 % | 58.7 % |
| IyedeX, fine-tuned | 12.9 % | 28.4 % |
In quiet environments, the error rate drops from 35.2% to 12.9%, nearly three times fewer errors. In noisy environments, it is halved.
Speech Synthesis
| Version | Average MOS | 95% Confidence Interval |
|---|---|---|
| Base xTTS | 2.8 / 5 | ±0.15 |
| IyedeX xTTS, fine-tuned | 4.1 / 5 | ±0.12 |
Listeners rate the fine-tuned voice 4.1 out of 5, compared to 2.8 for the base model.
Security and Ethics
IyedeX applies strict rules to every voice in the corpus and every voice produced:
- explicit consent from each speaker;
- encrypted storage of recordings;
- a policy against voice cloning;
- a watermark on synthetic voices;
- compliance with Tunisian and international regulations.
Data protection and ethical compliance are at the heart of our approach. Every voice in our corpus was collected with the explicit consent of its owner.
Conclusion
We present here the first complete professional pipeline for a Tunisian voice model. Despite limited hardware resources, our work on dialectal collection, geographical diversity, TTS and STT training, evaluation, and ethical design lays the foundations for a reliable, natural, and culturally faithful Tunisian voice ecosystem.
This model is the cornerstone of 3amAli, the IyedeX Voice API, and future enterprise integrations.
Comments 3
cest magnifique
nice
good + add transleter