Research paperVoice · Derja

Building a Tunisian voice model

Building a Tunisian voice model

Abstract

How we built the first speech synthesis (TTS) and speech recognition (STT) model for the Tunisian dialect: an 87.5-hour corpus from the North, Centre and South, a five-step data pipeline, and fine-tuning that cuts the error rate by nearly three.

87.5 hof Tunisian audio
41speakers, 3 regions
12.9%WER in quiet conditions
4.1 / 5MOS for synthesis

Introduction

Our model is built to understand and speak the Tunisian dialect, but above all it is your voice assistant. It adapts to the way you speak, to your regional accent and to your own expressions. It is not our place to tell you how to talk to your assistant.

This report sets out how IyedeX built the first TTS/STT voice model specialised in the Tunisian dialect, from an authentic corpus that includes:

  • openly licensed Tunisian radio and television archives;
  • legally available public recordings;
  • our own in-house voices;
  • the voices of collaborators from the North, the Centre and the South of Tunisia.

On top of that sit a complete pipeline for cleaning, annotating and preparing the audio, and the fine-tuning of the xTTS v2 model for a pronunciation that is genuinely Tunisian. The goal: natural synthesis and robust recognition, built into 3amAli, IyedeX’s assistant.

Our vision is 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 how Tunisian is really spoken. Here is the public version of its statistics.

Source Type Region Speakers Hours Licence Cleaning SNR (dB) Notes
Broadcast A Radio North 12 40.0 CC-BY 85% 18.5 Morning shows, debates
Broadcast B TV audio Centre 8 15.0 CC-BY-SA 80% 17.2 Reports
IyedeX recordings In-house North 1 5.0 Consent 95% 28.0 Amateur studio
Friends’ voices In-house 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 consent, and the rest from the field, where the noise is real.

pie showData title Where the recorded hours come from "Radio and television" : 55 "In-house voices" : 29 "Field" : 16

By region, the North accounts for 45 hours, the Centre for 15 and the South for 12.5. The remaining 20 hours mix speakers from several regions.

Preparing the data

Every hour of audio goes through the same five steps before it is used for training.

flowchart LR A[Normalisation] --> B[Cleaning and slicing] B --> C[Annotation] C --> D[Augmentation] D --> E[Final split]
  1. Normalisation. Conversion to mono 16 kHz or 24 kHz, 16-bit PCM, and consistent source files.
  2. Cleaning and slicing. Silences over 500 ms removed, audio cut into segments of 1 to 12 seconds, noise filtered automatically.
  3. Annotation. Manual and semi-automatic transcription, consistent spelling of the dialect, regional tags (North, Centre, South).
  4. Augmentation. Pitch shift of ±5%, time stretching, controlled background noise (markets, cars, music).
  5. Final split. 80% training, 10% validation, 10% test, with speakers kept strictly apart across the three.

The architecture

The model rests on two complementary parts: xTTS v2, fine-tuned for speech synthesis, and Wav2Vec2, fine-tuned for recognition.

Before fine-tuning After fine-tuning
Model Generic, multilingual Specialised in Tunisian
Pronunciation Approximate Authentic
Accent Not 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;
  • control over style and emotion.

Speech recognition (STT): fine-tuned Wav2Vec2

  • Wav2Vec2 as the base;
  • full fine-tuning on our corpus;
  • an error rate optimised for real-world conditions.

Doing a lot with a little

Our team worked on modest machines, without high-end GPUs. So we bet on engineering rather than raw compute.

Innovation does not depend on hardware. Our results show that the quality of the dataset and of the engineering matters more than raw computing power: we built a strong Tunisian voice model with limited resources.

Results

We evaluate the model with the 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, alongside phonetic error heatmaps.

Speech recognition

Model WER, quiet WER, noisy
Base Wav2Vec2 35.2% 58.7%
IyedeX, fine-tuned 12.9% 28.4%

In quiet conditions the error rate falls from 35.2% to 12.9%, nearly three times fewer errors. In noise, it is halved.

Speech synthesis

Version Mean 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, against 2.8 for the base model.

Safety and ethics

IyedeX applies strict rules to every voice in the corpus and every voice it produces:

  • explicit consent from every speaker;
  • encrypted storage of recordings;
  • a policy against voice cloning;
  • watermarking of synthetic voices;
  • compliance with Tunisian and international regulations.

Data protection and ethics are at the heart of what we do. Every voice in our corpus was collected with the explicit consent of its owner.

Conclusion

This is the first complete, professional pipeline for a Tunisian voice model. Despite limited hardware, our work on dialect data collection, geographic diversity, TTS and STT training, evaluation and ethical design lays the foundations of a Tunisian voice ecosystem that is reliable, natural and true to its culture.

This model is the cornerstone of 3amAli, of the IyedeX Voice API and of future integrations for businesses.

Cite this paper

IyedeX Research (2025). Building a Tunisian voice model. IyedeX. https://iyedex.tn/en/research/development-du-model-vocal

2 reads

All research