Whisper Darija v4
Whisper Darija v4 is a fine-tuned version of ayoubelfallah1/whisper-darija-v3 trained on an augmented Moroccan Darija speech dataset dedicated to flight search and reservation.
Compared to the previous version, this model significantly improves transcription accuracy and is designed to understand spoken Moroccan Darija queries related to air travel.
Model Description
The model is based on OpenAI Whisper and has been fine-tuned for Automatic Speech Recognition (ASR) in Moroccan Darija.
It recognizes natural spoken requests such as:
- Flight booking
- Flight search
- Departure city
- Destination city
- Travel dates
- Number of passengers
- Airline-related requests
The model is intended to be integrated into intelligent travel assistants and airline reservation systems.
Base Model
- Base model: ayoubelfallah1/whisper-darija-v3
Dataset
The model was trained on a custom Moroccan Darija speech dataset focused on the flight reservation domain.
The dataset was augmented using several data augmentation techniques to increase diversity and improve generalization.
Dataset characteristics:
- Original recordings
- Augmented audio samples
- Flight booking conversations
- Departure and destination cities
- Dates and passenger information
Training Hyperparameters
| Parameter | Value |
|---|---|
| Learning Rate | 5e-6 |
| Train Batch Size | 8 |
| Eval Batch Size | 8 |
| Gradient Accumulation Steps | 2 |
| Effective Batch Size | 16 |
| Number of Epochs | 2.7 |
| Warmup Steps | 100 |
| Weight Decay | 0.01 |
| FP16 | True |
| Gradient Checkpointing | True |
| Optimizer | AdamW |
| LR Scheduler | Linear |
Training Results
| Step | Training Loss | Validation Loss | WER |
|---|---|---|---|
| 100 | 1.4161 | 0.2382 | 23.19 |
| 200 | 0.3389 | 0.0932 | 11.76 |
| 300 | 0.1494 | 0.0555 | 8.20 |
Final Performance
The best checkpoint achieved:
- Validation Loss: 0.0555
- Word Error Rate (WER): 8.20%
This represents a substantial improvement over the previous version (v3), making the model more accurate for Moroccan Darija speech recognition in the flight booking domain.
Intended Uses
This model can be used for:
- Automatic Speech Recognition (ASR)
- Moroccan Darija transcription
- Voice assistants
- Airline reservation systems
- Conversational AI
- Smart travel assistants
Limitations
Although the model performs well on flight-related speech, performance may decrease for:
- General conversations
- Medical vocabulary
- Legal terminology
- Strong regional accents
- Very noisy recordings
Example
Input:
"بغيت نحجز رحلة من كازا لباريس نهار الجمعة."
Output:
"بغيت نحجز رحلة من كازا لباريس نهار الجمعة"
Framework Versions
- Transformers
- PyTorch
- Datasets
- Tokenizers
Author
Ayoub El Fallah
This model was developed as part of a graduation project focused on building an intelligent Moroccan Darija voice assistant for flight search and booking.