Quran Whisper Transcription

This repository contains a Python script that utilizes the fine-tuned Whisper model to transcribe Quranic recitations from audio files. The Whisper model used in this project has been fine-tuned specifically on Quranic recitation, enabling accurate transcription of Arabic text from audio recordings.

Features

• Quran-Specific Transcription: The model is fine-tuned on Quranic recitations, providing high accuracy in transcribing Arabic verses.
• Audio Input Handling: Accepts standard audio file formats (e.g., MP3) and processes them for transcription.
• TorchScript Conversion: The model is also converted to TorchScript for easier deployment in production environments.

Getting Started

Prerequisites

• Python 3.7+
• torch
• transformers
• librosa

Installation

1. Clone the repository:

   git clone https://github.com/yourusername/quran-whisper-transcription.git

2. Install the required packages:

   pip install torch transformers librosa

Usage

1. Load the model and processor:

   from transformers import WhisperProcessor, WhisperForConditionalGeneration
   import torch
   import librosa
   
   processor = WhisperProcessor.from_pretrained("tarteel-ai/whisper-tiny-ar-quran")
   model = WhisperForConditionalGeneration.from_pretrained("tarteel-ai/whisper-tiny-ar-quran")

2. Transcribe an audio file:

   audio_path = 'path_to_your_audio.mp3'
   audio_data, sampling_rate = librosa.load(audio_path, sr=16000)
   audio = torch.tensor(audio_data)
   
   inputs = processor(audio, return_tensors="pt", sampling_rate=16000)
   generated_ids = model.generate(inputs["input_features"])
   
   transcription = processor.batch_decode(generated_ids, skip_special_tokens=True)
   print("Transcription:", transcription)

3. TorchScript Conversion (Optional):

   scripted_model = torch.jit.trace(model, torch.randn(1, 80, 3000))
   scripted_model.save("whisper_tiny_ar_quran_scripted.pt")

Notes

• Ensure the audio file is correctly preprocessed (resampled to 16 kHz) for accurate transcription.
• The TorchScript model may have limitations due to unsupported operations in the original model.

Acknowledgments

• Tarteel.ai for providing the fine-tuned Whisper model for Quranic recitations.
• Hugging Face for the Transformers library.

---