assermosa/Tunisian-Arabic-Automatic-Speech-Recognition-ASR-
Domain:
natural language processingRecord type:
model
Creator:
ass
Host:
project combines multiple Tunisian speech datasets, applies audio augmentation techniques, and achieves a reduced Word Error Rate (WER) through targeted dialect-specific fine-tuning.
# 🎤 Tunisian Arabic Speech Recognition (ASR)
Fine-tuned SpeechT5 model for Automatic Speech Recognition of Tunisian Arabic dialect.
## 📋 Overview
This project presents a state-of-the-art Automatic Speech Recognition system specifically designed for the Tunisian Arabic dialect. Built on Microsoft's SpeechT5 architecture, the model has been fine-tuned on a carefully curated combination of three Tunisian Arabic datasets with advanced data augmentation techniques.
## ✨ Key Features
- **Specialized for Tunisian Dialect**: Fine-tuned specifically on Tunisian Arabic speech patterns
- **Multi-Dataset Training**: Combined 3 diverse datasets for robust performance
- **Advanced Augmentation**: White noise, background noise, and gain augmentation
- **Text Normalization**: Arabic text preprocessing and diacritics removal
- **High Accuracy**: Achieved **58.0% WER** on validation set after 1000 training steps
- **Production Ready**: Easy-to-use Gradio interface for inference
## 🎯 Model Performance
| Metric | Score |
|--------|-------|
| **Final WER (Normalized)** | 58.01% |
| **Final WER (Orthographic)** | 59.72% |
| **Training Steps** | 1,000 |
| **Best Checkpoint** | Step 900 |
### Training Progress
| Step | Training Loss | Validation Loss | WER (Normalized) |
|------|---------------|-----------------|------------------|
| 100 | 1.0428 | 0.8283 | 108.66% |
| 500 | 0.3311 | 0.2349 | 64.57% |
| 900 | 0.2364 | 0.1921 | **58.01%** |
| 1000 | 0.2704 | 0.1922 | 58.01% |
## 📊 Datasets
The model was trained on a combination of three high-quality Tunisian Arabic datasets:
1. **linagora/linto-dataset-audio-ar-tn (AmenyKH)**
- Subset: AmenyKH
- Samples: 3,094 (train) + 44 (test)
2. **Arbi-Houssem/Tunisian_dataset_STT-TTS15s_filtred1.0**
- Samples: 1,032 (train) + 71 (validation)
3. **Raniahossam33/tunisa_data**
- Samples: 1,235 (train)
**Total Dataset Size**: 5,361 samples
- …