Quranic Wav2Vec2 Phonetic ASR

Model Description

Quranic Wav2Vec2 Phonetic ASR is a fine-tuned wav2vec2 model designed specifically for phonetic transcription of Quranic recitation.

Unlike standard Arabic ASR systems that output orthographic Arabic text, this model outputs a phonetic (sound-level) representation, making it suitable for:

• Tajweed research and education
• Word-level pronunciation analysis
• Forced alignment of Quranic recitation
• Linguistic studies of Classical Arabic phonetics

To the best of our knowledge, this is the first publicly released wav2vec2 model trained explicitly for Quranic phonetic transcription.

---

Key Features

• 🎙️ Outputs phonetic strings, not Arabic text
• 📖 Trained on Quranic recitation, not conversational Arabic
• 🔤 Custom phonetic vocabulary + tokenizer
• 🧠 Compatible with CTC forced alignment
• 🧩 Optimized for word-level Tajweed analysis

---

Intended Use

This model is intended for:

• Quranic pronunciation analysis
• Tajweed educational tools
• Phoneme-level alignment
• Academic research on Quranic recitation

⚠️ Not intended for:

• Modern Arabic ASR
• End-to-end Tajweed grading of full ayat
• Automatic religious judgments
• Replacement of qualified Quran teachers

---

Model Architecture

• Base model: facebook/wav2vec2-base
• Training objective: CTC (Connectionist Temporal Classification)
• Sampling rate: 16 kHz
• Input: Mono audio waveform
• Output: Phonetic transcription string
• Tokenizer: Custom character-level phonetic tokenizer

---

Training Dataset

📦 Dataset Used

Buraaq/quran-md-words

This dataset contains word-level Quranic recitation audio with aligned metadata.

Each sample includes:

• Audio of a single Quranic word
• Phonetic transcription (word_tr)
• Arabic word (word_ar)
• Surah and ayah identifiers
• Word index within the ayah

This dataset was chosen intentionally to:

• Preserve clear phonetic boundaries
• Avoid coarticulation noise from full-ayah audio
• Enable high-accuracy phonetic learning

🔗 Dataset: https://huggingface.co/datasets/Buraaq/quran-md-words

---

Phonetic Vocabulary & Tokenizer

The model uses a custom phonetic vocabulary built directly from the dataset:

• Vocabulary constructed from all unique characters in word_tr
• Character-level CTC tokenizer
• Includes:
  • | as word delimiter
  • [PAD] for CTC blank
  • [UNK] for unknown symbols

This design allows:

• Robust forced alignment
• Fine-grained phonetic decoding
• Word boundary detection via delimiter token

---

Training Procedure

Training Setup

• Framework: Hugging Face Transformers + Datasets
• Platform: Windows-safe (no multiprocessing crashes)
• Audio handling: Hugging Face Audio feature
• Batch size: 16
• Epochs: 20
• Learning rate: 2e-5
• Optimizer: AdamW (default Trainer)
• Precision: FP16 when CUDA available
• Train / Eval split: 90% / 10%

The wav2vec2 feature extractor was not frozen, allowing adaptation to Quranic recitation acoustics.

---

Training Results

On the held-out evaluation split:

• Training accuracy: 99.7%
• Evaluation accuracy: 99.8%
• Test split size: 10% of the dataset

⚠️ Note:
These results reflect word-level phonetic transcription accuracy on a dataset with consistent recitation style.
Performance may degrade on:

• Fast recitation
• Strong coarticulation
• Unseen riwayat styles

---

Example Usage

Load Model and Processor

from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
import torch
import soundfile as sf
import librosa
import numpy as np

processor = Wav2Vec2Processor.from_pretrained(
    "USERNAME/quranic-wav2vec2-phonetic"
)
model = Wav2Vec2ForCTC.from_pretrained(
    "USERNAME/quranic-wav2vec2-phonetic"
)

model.eval()

Phonetic Transcription Example

audio, sr = sf.read("recitation.wav")
# convert to mono
if audio.ndim > 1:
    audio = audio.mean(axis=1)

# resample to 16kHz
if sr != 16000:
    audio = librosa.resample(audio, orig_sr=sr, target_sr=16000)

inputs = processor(
    audio,
    sampling_rate=16000,
    return_tensors="pt",
    padding=True,
)

with torch.inference_mode():
    logits = model(inputs.input_values).logits
    predicted_ids = torch.argmax(logits, dim=-1)

phonetics = processor.batch_decode(
    predicted_ids,
    skip_special_tokens=True
)[0]

print(phonetics)

Example output:

tālik