Model Overview

The Quran Muaalem model is a multi‑level CTC system based on Wav2Vec2BERT. It predicts:

• A phoneme sequence (primary head)
• Sifat attributes (secondary heads)

Where the model lives

• Wrapper + inference: src/quran_muaalem/inference.py
• Architecture: src/quran_muaalem/modeling/modeling_multi_level_ctc.py
• Config: src/quran_muaalem/modeling/configuration_multi_level_ctc.py
• Tokenization: src/quran_muaalem/modeling/multi_level_tokenizer.py

Key runtime settings

• model_name_or_path (default: obadx/muaalem-model-v3_2)
• dtype (default: torch.bfloat16)
• device (CPU or CUDA)

How to interpret outputs

Decoded outputs are assembled into MuaalemOutput objects with phonemes + sifat. See Outputs for schema and examples.

For researchers

When reporting results, include:

• Model version
• Vocab / tokenizer version
• Loss weights per level
• Evaluation metrics (PER, Sifat F1, alignment accuracy)