Whisper Small Finetuned on Surah Fatiha
This model is a fine-tuned version of openai/whisper-small, transcribing Surah Fatiha, the first chapter of the Quran. It has been trained using The Truth 2.0 - Surah Fatiha dataset and achieves excellent results with a Word Error Rate (WER) of 0.0, indicating perfect transcription on the evaluation set.
Model Description
Whisper Small is a transformer-based automatic speech recognition (ASR) model developed by OpenAI. By fine-tuning it on the Surah Fatiha dataset, this model becomes highly accurate in transcribing Quranic recitation. It is designed to assist in religious, educational, and research-oriented tasks that require precise Quranic transcription.
Performance Metrics
On the evaluation set, the model achieved:
- Loss: 0.0088
- Word Error Rate (WER): 0.0
These metrics showcase the model's exceptional performance and reliability in transcribing Surah Fatiha audio.
Training Results
The following table summarizes the training process and results:
| Training Loss | Epoch | Step | Validation Loss | WER |
|---|---|---|---|---|
| No log | 0.5556 | 10 | 1.1057 | 96.2766 |
| No log | 1.1111 | 20 | 0.3582 | 29.7872 |
| 0.6771 | 1.6667 | 30 | 0.1882 | 23.4043 |
| 0.6771 | 2.2222 | 40 | 0.0928 | 25.0 |
| 0.0289 | 2.7778 | 50 | 0.0660 | 34.0426 |
| 0.0289 | 3.3333 | 60 | 0.0484 | 32.9787 |
| 0.0289 | 3.8889 | 70 | 0.0241 | 25.5319 |
| 0.0056 | 4.4444 | 80 | 0.0184 | 28.7234 |
| 0.0056 | 5.0 | 90 | 0.0111 | 0.0 |
| 0.0019 | 5.5556 | 100 | 0.0088 | 0.0 |
Intended Uses & Limitations
Intended Uses
- Speech-to-text transcription of Quranic recitation for Surah Fatiha.
- Educational tools to assist in learning and practicing Quranic recitation.
- Research and analysis of Quranic audio transcription methods.
Limitations
- This model is fine-tuned specifically for Surah Fatiha and may not generalize well to other chapters or non-Quranic Arabic audio.
- Variability in audio quality, accents, or recitation styles might affect performance.
- Optimal performance is achieved with high-quality audio inputs.
Training and Evaluation Data
The model was trained on The Truth 2.0 - Surah Fatiha dataset, which comprises high-quality audio recordings of Surah Fatiha and their corresponding transcripts. The dataset was meticulously curated to ensure the accuracy and authenticity of Quranic content.
Training Procedure
Training Hyperparameters
The following hyperparameters were used during training:
- Learning Rate: 1e-05
- Training Batch Size: 16
- Evaluation Batch Size: 8
- Seed: 42
- Optimizer: Adam (betas=(0.9, 0.999), epsilon=1e-08)
- Learning Rate Scheduler: Linear
- Warmup Steps: 10
- Training Steps: 100
- Mixed Precision Training: Native AMP
Framework Versions
- Transformers: 4.41.1
- PyTorch: 2.2.1+cu121
- Datasets: 2.19.1
- Tokenizers: 0.19.1