naazimsnh02/whisper-large-v3-turbo-ar-quran

Whisper Large V3 Turbo - Quranic Arabic

Fine-tuned Whisper Large V3 Turbo model for high-accuracy Arabic Quranic speech recognition.

Model Description

This model is a fine-tuned version of openai/whisper-large-v3-turbo trained on the tarteel-ai/everyayah dataset.

Performance

Test Set Evaluation

The model was evaluated on the full test split of the EveryAyah dataset, comprising 23,473 samples.

Metric	Value
Word Error Rate (WER)	1.18%
Character Error Rate (CER)	0.34%

Validation Set Evaluation

Performance on the validation split (20,976 samples) during training:

Metric	Value
Word Error Rate (WER)	0.86%
Eval Loss	0.0040

Training Details

Training Data

• Dataset: tarteel-ai/everyayah
• Train Samples: 167,908
• Validation Samples: 20,976 (used during training)
• Test Samples: 23,473 (used for final evaluation)

Training Hyperparameters

LoRA Configuration:

• LoRA Rank (r): 128
• LoRA Alpha: 128
• LoRA Dropout: 0.05
• Target Modules: q_proj, v_proj, k_proj, out_proj, fc1, fc2

Training:

• Max Steps: -1
• Epochs: 2
• Batch Size: 24
• Gradient Accumulation: 2
• Effective Batch Size: 48
• Learning Rate: 5e-05
• LR Scheduler: cosine
• Warmup Steps: 500
• Weight Decay: 0.01
• Max Grad Norm: 1.0
• Optimizer: AdamW Fused

Usage

Quick Start (Recommended)

For audio of any length, including recordings longer than 30 seconds:

from transformers import pipeline
import torch

# Load the model with automatic chunking for long audio
pipe = pipeline(
    "automatic-speech-recognition",
    model="naazimsnh02/whisper-large-v3-turbo-ar-quran",
    device=0,  # Use GPU (or -1 for CPU)
    torch_dtype=torch.float16,
    chunk_length_s=30,      # Process in 30-second chunks
    stride_length_s=5,      # 5-second overlap between chunks
)

# Transcribe audio of any length
result = pipe("long_quran_recitation.wav")
print(result["text"])

For Short Audio (< 30 seconds)

from transformers import pipeline

# Simple usage for short audio
pipe = pipeline(
    "automatic-speech-recognition",
    model="naazimsnh02/whisper-large-v3-turbo-ar-quran",
    device=0
)

result = pipe("short_ayah.wav")
print(result["text"])

Data Processing

• Audio filtered to ≤ 30 seconds
• Text filtered to ≤ 448 tokens (Whisper's limit)
• Audio resampled to 16kHz
• Language: Arabic (Quranic)
• Task: Transcription

Model Architecture

• Base Model: OpenAI Whisper Large V3 Turbo
• Parameters: 809M total, ~2M trainable (LoRA)
• Training Method: LoRA fine-tuning
• Precision: FP16/BF16
• Context Length: 30 seconds per chunk

Limitations

• Optimized for Quranic Arabic: Best performance on formal Quranic recitation
• Not suitable for: Conversational Arabic, dialects, or non-Quranic content
• Audio quality: Best results with clear, high-quality recordings

Training Framework

• Unsloth for 2x faster training
• Modal for cloud GPU infrastructure
• Weights & Biases for experiment tracking

Citation

@misc{whisper-quran-2025,
  author = {Syed Naazim Hussain},
  title = {Whisper Large V3 Turbo - Quranic Arabic},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/naazimsnh02/whisper-large-v3-turbo-ar-quran}}
}