118 lines
3.6 KiB
Markdown
118 lines
3.6 KiB
Markdown
---
|
|
base_model:
|
|
- openai/whisper-small
|
|
datasets:
|
|
- FaisaI/tadabur
|
|
language:
|
|
- ar
|
|
license: cc-by-nc-4.0
|
|
metrics:
|
|
- wer
|
|
pipeline_tag: automatic-speech-recognition
|
|
library_name: transformers
|
|
tags:
|
|
- quran
|
|
- asr
|
|
- arabic
|
|
- speech-recognition
|
|
---
|
|
|
|
<div align="center">
|
|
<img src="https://huggingface.co/datasets/FaisaI/tadabur/resolve/main/tadabur_logo.png" width="100"><br><br>
|
|
|
|
<h1>Tadabur-Whisper-Small</h1>
|
|
|
|
A Whisper Small model fine-tuned on [Tadabur](https://huggingface.co/datasets/FaisaI/tadabur) for Qur'anic speech recognition.
|
|
|
|
[](https://huggingface.co/papers/2604.18932)
|
|
[](https://huggingface.co/datasets/FaisaI/tadabur)
|
|
[](https://huggingface.co/openai/whisper-small)
|
|
[](https://creativecommons.org/licenses/by-nc/4.0/)
|
|
[](https://fherran.github.io/tadabur)
|
|
|
|
</div>
|
|
|
|
---
|
|
## Overview
|
|
**Tadabur-Whisper-Small** is a fine-tuned version of [Whisper Small](https://huggingface.co/openai/whisper-small) on the [Tadabur dataset](https://huggingface.co/datasets/FaisaI/tadabur), as presented in the paper [Tadabur: A Large-Scale Quran Audio Dataset](https://huggingface.co/papers/2604.18932).
|
|
|
|
- **GitHub Repository:** [fherran/tadabur](https://github.com/fherran/tadabur)
|
|
- **Project Page:** [fherran.github.io/tadabur](https://fherran.github.io/tadabur)
|
|
|
|
---
|
|
## Training Iteration
|
|
| Step | Epoch | WER ↓ |
|
|
|:---:|:---:|:---:|
|
|
| 2,500 | 0.15 | 13.78% |
|
|
| 5,000 | 0.30 | 11.20% |
|
|
| 7,500 | 0.44 | 11.15% |
|
|
| 25,000 | 1.48 | **7.89%** ⭐ |
|
|
| 32,500 | 1.93 | 14.75% |
|
|
|
|
---
|
|
## Usage
|
|
|
|
```python
|
|
from transformers import pipeline
|
|
|
|
asr = pipeline(
|
|
"automatic-speech-recognition",
|
|
model="FaisaI/tadabur-whisper-small",
|
|
generate_kwargs={"language": "arabic"}
|
|
)
|
|
|
|
result = asr("path/to/audiofile")
|
|
print(result["text"])
|
|
```
|
|
|
|
Or with the full Whisper API:
|
|
|
|
```python
|
|
from transformers import WhisperProcessor, WhisperForConditionalGeneration
|
|
import librosa
|
|
|
|
processor = WhisperProcessor.from_pretrained("FaisaI/tadabur-whisper-small")
|
|
model = WhisperForConditionalGeneration.from_pretrained("FaisaI/tadabur-whisper-small")
|
|
|
|
|
|
|
|
# Audio must be 16kHz mono
|
|
audio_array, sampling_rate = librosa.load("path/to/audiofile", sr=16000,mono=True)
|
|
inputs = processor(audio_array, sampling_rate=16000, return_tensors="pt")
|
|
|
|
predicted_ids = model.generate(**inputs, language="arabic")
|
|
transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)
|
|
print(transcription[0])
|
|
```
|
|
---
|
|
|
|
## Limitations
|
|
|
|
- Not suitable for speaker identification or diarization.
|
|
- May underperform on noisy or low-quality recordings.
|
|
- Not fully generalized — transcription errors are expected.
|
|
|
|
---
|
|
|
|
## Ethical Considerations
|
|
|
|
This model is trained exclusively on Qur'anic recitation data. Users must engage with outputs respectfully and must not use this model for mockery, distortion, or any disrespectful application involving Qur'anic content.
|
|
|
|
**For research and educational use only.**
|
|
|
|
---
|
|
|
|
## Citation
|
|
|
|
```bibtex
|
|
@misc{alherran2026tadabur,
|
|
author = {Alherran, Faisal},
|
|
title = {Tadabur: A Large-Scale Quran Audio Dataset},
|
|
year = {2026},
|
|
eprint = {2604.18932},
|
|
archivePrefix = {arXiv},
|
|
primaryClass = {cs.SD},
|
|
doi = {10.48550/arXiv.2604.18932},
|
|
url = {https://arxiv.org/abs/2604.18932}
|
|
}
|
|
``` |