---
license: apache-2.0
datasets:
- tarteel-ai/everyayah
- MohamedRashad/Quran-Recitations
- ahishamm/QURANICWhisperDataset
language:
- ar
metrics:
- wer
base_model:
- openai/whisper-large-v3-turbo
pipeline_tag: automatic-speech-recognition
library_name: transformers
tags:
- Quran
- Tajweed
- Recitation
- Islam
- Arabic
- whisper
- turbo
---

# Whisper Large v3 Turbo Quran (LoRA Fine-Tuned)

This is a specialized **Automatic Speech Recognition (ASR)** model for **Quranic Recitation** with **tashkeel** or **diacritics**. It is a fine-tuned version of [openai/whisper-large-v3-turbo](https://huggingface.co/openai/whisper-large-v3-turbo), optimized to recognize Quranic Arabic (with diacritics or tashkeel) with high accuracy while maintaining exceptional inference speed.

## Model Performance
- **Word Error Rate (WER):** Achieved **12.69%** on the `ahishamm/QURANICWhisperDataset` test set (20% of the dataset).
- **Accuracy:** The model demonstrates high precision in capturing Quranic vocabulary and standard script nuances.

## Architecture & Trade-offs
This model utilizes the **Turbo** architecture, which reduces the decoder depth from 32 layers (in the standard Large v3) to **4 layers**. 

- **Pros:** Extremely fast inference speed (significantly lower latency than Medium or Large v3).
- **Cons:** Due to the reduced decoder depth, the model has less capacity for long-range context retention compared to the full Large model. This makes it slightly more prone to **hallucinations** and **repetition loops** (e.g., repeating a word during silence), especially if not configured correctly during inference.

**Recommendation:** This model is great for fast verse transcription, **Live Transcription**, or applications where low latency is critical. For offline batch processing where speed is not a priority, a full Large-v3 (or even a medium whisper) model will offer higher semantic stability, especially on long verses.

## Training Details
The model was trained using **LoRA (Low-Rank Adaptation)** in a multi-stage curriculum learning process to ensure stability and precision.

### Datasets
The training and evaluation process utilized a comprehensive mix of professional recitations:
1.  **Training & Validation:**
    - [tarteel-ai/everyayah](https://huggingface.co/datasets/tarteel-ai/everyayah)
    - [MohamedRashad/Quran-Recitations](https://huggingface.co/datasets/MohamedRashad/Quran-Recitations)
2.  **Testing:**
    - [ahishamm/QURANICWhisperDataset](https://huggingface.co/datasets/ahishamm/QURANICWhisperDataset) (Used for final WER calculation).

### Methodology
- **Curriculum Learning:** The model was trained gradually across these datasets to refine its understanding of Tajweed and Quranic sentence structures.
- **Data Augmentation:** To ensure the model remains robust against real-world conditions (non-studio microphones, background noise, varying volumes), diverse audio augmentations (gain adjustments, spectral masking and white noise) were applied during the training process.

## Usage

This model is fully compatible with the Hugging Face `transformers` pipeline. 

**CRITICAL NOTE ON INFERENCE:**
Due to the Turbo architecture's reduced decoder depth, you must carefully control the `stride_length_s` parameter. A long stride (e.g., 5s or more) can cause the model to lose context and enter infinite repetition loops. **It is recommended to keep the stride length to around 2 or 3 seconds.**

```python
from transformers import pipeline

# Load the pipeline
pipe = pipeline(
    "automatic-speech-recognition",
    model="MaddoggProduction/whisper-l-v3-turbo-quran-lora-dataset-mix",
    device=0 # for GPU usage, -1 for CPU
)

# Transcribe audio 
result = pipe(
    "path_to_audio.mp3",
    chunk_length_s=30, 
    stride_length_s=2, # 2s or 3s to prevent loops and hallucinations
    batch_size=8,
    return_timestamps=True,
    generate_kwargs={
        "task": "transcribe",
        "language": "arabic",
        "num_beams": 1           # 1 is sufficient, adjust as needed
    }
)

print(result["text"])