| | |

🇹🇳 Tunisian STT — Automatic Speech Recognition

Tunisian STT is a high-quality Automatic Speech Recognition (ASR) system dedicated to the Tunisian Arabic dialect. The model is designed for real-world usage and supports code-switching, enabling robust recognition of Tunisian Arabic mixed with French and English terms.

This repository provides production-ready speech-to-text models along with a language model (ARPA format), making it suitable for research, prototyping, and deployment across multiple platforms.

✨ Key Features

🎙️ Native support for Tunisian Arabic dialect
🌍 Robust handling of code-switching (Arabic–French–English)
⚡ Optimized variants for mobile and embedded devices
🧠 Powered by Kaldi TDNN architecture
📦 Ready-to-use with Vosk

📦 Available Models

This repository includes two acoustic model variants:

STT_Tun_Model The full-scale, high-accuracy model suitable for servers and desktop environments.
android-model A lightweight version with a simplified decoding graph, optimized for Android, Raspberry Pi, and edge devices.

Both models are shipped with a language model (ARPA) for decoding.

🧠 Model Overview

Model Type: TDNN (Time-Delay Neural Network)
Framework: Kaldi
Primary Language: Tunisian Arabic
Task: Automatic Speech Recognition (ASR)

📊 Performance Benchmarks

The table below reports Character Error Rate (CER) and Word Error Rate (WER) across multiple Tunisian speech datasets:

Dataset	CER	WER
OneStory	1.53%	4.47%
YouTube_TN_Scraped_V1	25.39%	37.51%
TARIC	10.60%	16.06%
TunSwitchTO	11.13%	22.54%
TunSwitchCS	17.72%	20.51%
ApprendreLeTunisien	11.81%	23.27%

Note: Performance varies depending on recording quality, speaker accent, and code-switching density.

🏗️ Training Pipeline

The model was trained using the official Kaldi-based training pipeline available here:

➡️ Training Code: ASR_train_kaldi_tunisian

🗂️ Training Datasets

The acoustic model was trained on a diverse collection of Tunisian Arabic speech datasets:

TARIC Real-world speech recordings collected in Tunisian train stations.
LinTO Dataset – Audio for Tunisian Arabic A curated corpus of annotated Tunisian dialect speech sourced from Hugging Face, YouTube, and web data.
LinTO Dataset – Augmented Version An enhanced dataset with data augmentation techniques including noise robustness and voice conversion.

🚀 Getting Started

1️⃣ Download the Model

Option A — Hugging Face Website

Visit the model page: 👉 https://huggingface.co/Sali7a8603/Tunisian_STT
Click Download ZIP to retrieve the desired model.

Option B — Command Line (Recommended)

sudo apt-get install curl

curl -L https://huggingface.co/Sali7a8603/Tunisian_STT/resolve/main/STT_Tun_Model.zip \
     --output STT_Tun_Model.zip

Replace STT_Tun_Model.zip with android-model.zip if needed.

2️⃣ Extract the Model

mkdir tunisian_stt_model
unzip STT_Tun_Model.zip -d tunisian_stt_model

3️⃣ Run Test (Python)

Install Dependencies

pip install vosk

Run Speech-to-Text

python testing.py <path/to/model> <path/to/audio.wav>

The output will be the transcribed text in Tunisian Arabic, with support for embedded French and English terms.

📜 License

This project is released under the Apache 2.0 License, allowing both academic and commercial usage.

🤝 Contributions & Feedback

Contributions, issues, and feature requests are welcome. If you use this model in your research or product, a citation or acknowledgment would be appreciated.

Downloads last month: -; Downloads are not tracked for this model. How to track

Sali7a8603
/

Tunisian_STT