Whisper base AR - YA

This model is a fine-tuned version of openai/whisper-base on an Arabic Quran recitation dataset focused on verse-level speech-to-text transcription. The goal was to create a lightweight ASR system that can accurately transcribe Quranic audio into Arabic text, optimized for clear, male recitation audio.

It achieves the following results:

• Validation set:
  • Loss: 0.0023
  • WER (Word Error Rate): 4.05%
  • CER (Character Error Rate): 1.95%
• Test set:
  • WER (Word Error Rate): 8.2%
  • CER (Character Error Rate): 3.27%

Model description

This model builds upon OpenAI's Whisper base architecture and is fine-tuned specifically for Modern Standard Arabic, with a focus on Quranic verses. Audio samples were cleaned, resampled to 16kHz, and aligned with text for training.

The model is trained using CTC loss in a supervised setting, making it suitable for inference in streaming or batch-based ASR systems. Whisper’s multilingual capabilities were leveraged to build a domain-specific Arabic transcription model.

Intended uses & limitations

Intended uses:

• Speech recognition for Arabic Quran recitations
• Educational tools or Quran learning applications
• Mobile-friendly deployment of ASR for religious audio content
• Fine-tuning or distillation for low-resource Arabic ASR projects

Limitations:

• Optimized for clear, male Quran recitation—performance may degrade with female voices or conversational Arabic
• Not designed for dialectal or informal speech
• Background noise or overlapping speakers may reduce accuracy

Training and evaluation data

The dataset consists of verse-level Quran recitations in Arabic. The recordings were primarily from male speakers with clear tajweed (recitation rules), and aligned to their corresponding Arabic text.

Audio files were resampled to 16kHz and normalized for Whisper compatibility.

Evaluation was conducted on both a held-out validation set and a separate test set to assess generalization.

Training procedure

Training hyperparameters

• learning_rate: 0.0001
• train_batch_size: 8
• eval_batch_size: 8
• gradient_accumulation_steps: 2
• total_train_batch_size: 16
• num_train_epochs: 30
• seed: 42
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• optimizer: AdamW (betas=(0.9, 0.999), eps=1e-08)
• mixed_precision_training: Native AMP

Training was conducted using PyTorch with Hugging Face Trainer API. Metrics monitored include WER and CER.

Training results

This is only the results of the last batch not all batches

Framework versions

• Transformers: 4.51.1
• PyTorch: 2.5.1+cu124
• Datasets: 2.20.0
• Tokenizers: 0.21.0