|
--- |
|
language: |
|
- ar |
|
license: apache-2.0 |
|
base_model: openai/whisper-small |
|
tags: |
|
- fine-tuned |
|
- Quran |
|
- automatic-speech-recognition |
|
- arabic |
|
- whisper |
|
datasets: |
|
- fawzanaramam/the-truth-1st-chapter |
|
metrics: |
|
- wer |
|
model-index: |
|
- name: Whisper Small Finetuned on Surah Fatiha |
|
results: |
|
- task: |
|
name: Automatic Speech Recognition |
|
type: automatic-speech-recognition |
|
dataset: |
|
name: The Truth 2.0 - Surah Fatiha |
|
type: fawzanaramam/the-truth-1st-chapter |
|
args: 'config: ar, split: train' |
|
metrics: |
|
- name: Word Error Rate (WER) |
|
type: wer |
|
value: 0.0 |
|
--- |
|
|
|
# Whisper Small Finetuned on Surah Fatiha |
|
|
|
This model is a fine-tuned version of [openai/whisper-small](https://huggingface.co/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 |
