FastConformer-Arabic-SADA-Finetune-SAREI
Arabic Automatic Speech Recognition (ASR) model fine-tuned using NVIDIA NeMo FastConformer on the SADA Arabic dialect dataset.
This model was developed as part of the SAREI Intelligent Unified Medical Emergency Platform project to support Arabic emergency call transcription and EMS triage assistance.
Project Repository:
https://github.com/vabdull/SAREI-AI-Assisted-Voice-Based-Triage-System-for-EMS
Model Details
- Base Architecture: NVIDIA NeMo FastConformer
- Framework: NVIDIA NeMo
- Language: Arabic
- Task: Automatic Speech Recognition (ASR)
- Dataset: SADA Arabic Speech Dataset
- Decoders Evaluated:
- RNNT
- CTC
The best-performing final decoder was RNNT.
Dataset
SADA Arabic Speech Dataset
The model was fine-tuned using the SADA Arabic Speech Dataset, which contains Arabic dialect speech samples including:
- Najdi dialect
- Hijazi dialect
- Khaleeji dialect
The dataset was used to improve Arabic dialect speech recognition performance for real-world conversational and emergency-response scenarios.
Dataset Usage
The dataset was used for:
- ASR fine-tuning
- Validation
- Testing
- Dialect evaluation
- RNNT and CTC decoder comparison
ASR Model Performance
Final RNNT Accuracy
- Validation Word Accuracy: 72.96%
- Validation Character Accuracy: 88.53%
- Test Word Accuracy: 68.99%
- Test Character Accuracy: 86.60%
Overall Evaluation Results
| Split | Decoder | WER | CER |
|---|---|---|---|
| Validation | RNNT | 27.04% | 11.47% |
| Test | RNNT | 31.01% | 13.40% |
| Validation | CTC | 29.69% | 11.53% |
| Test | CTC | 34.11% | 13.63% |
Test Performance by Dialect
| Dialect | RNNT WER | RNNT CER | CTC WER | CTC CER |
|---|---|---|---|---|
| Najdi | 30.06% | 13.02% | 33.38% | 13.34% |
| Hijazi | 28.86% | 12.46% | 31.02% | 12.61% |
| Khaleeji | 33.94% | 14.69% | 37.45% | 14.85% |
Intended Use
This model is primarily intended for integration within the:
SAREI Intelligent Unified Medical Emergency Platform
The model is designed to support Emergency Medical Services (EMS) by providing:
- Real-time Arabic emergency call transcription
- AI-assisted dispatcher support
- Live speech-to-text conversion
- Emergency keyword detection
- Faster emergency response workflows
- Arabic dialect speech recognition in EMS environments
The system is intended to assist responders and dispatchers, not replace human decision-making.
This model may also be used for:
- Arabic ASR research
- Academic projects
- Dialect speech recognition experimentation
- Healthcare AI applications
Notes
- The model was fine-tuned on Arabic dialect speech.
- Performance may vary depending on:
- background noise
- microphone quality
- speaker accent
- emergency/stress conditions
Project Context
This ASR model is part of the:
SAREI Intelligent Unified Medical Emergency Platform
An AI-assisted emergency response system designed to support EMS dispatchers through:
- live speech transcription
- keyword detection
- AI-assisted triage
- emergency case reporting
Related Project Repository
🔗 https://github.com/vabdull/SAREI-AI-Assisted-Voice-Based-Triage-System-for-EMS
The repository contains the complete implementation of the AI-assisted EMS platform, including:
- Dispatcher portal
- Ambulance portal
- Hospital portal
- Live ASR integration
- Real-time transcription pipeline
- Emergency triage workflow
- AI-assisted responder support
Authors
Developed by Abdullah Alotaibi, Abdulmalik Alotaibi, Mohammed Aljabri. Supervised By Dr. Ismail Keshta, Dr. Mohammed Al Gabri
License
Apache-2.0
- Downloads last month
- 7