Homo-GE2PE: Persian Grapheme-to-Phoneme Conversion with Homograph Disambiguation

Homo-GE2PE is a Persian grapheme-to-phoneme (G2P) model specialized in homograph disambiguation—words with identical spellings but context-dependent pronunciations (e.g., مرد pronounced as mard "man" or mord "died"). Introduced in Fast, Not Fancy: Rethinking G2P with Rich Data and Rule-Based Models, the model extends GE2PE by fine-tuning it on the HomoRich dataset, explicitly designed for such pronunciation challenges.

Repository Structure

model-weights/
│   ├── homo-ge2pe.zip       # Homo-GE2PE model checkpoint
│   └── homo-t5.zip          # Homo-T5 model checkpoint (T5-based G2P model)

training-scripts/
│   ├── finetune-ge2pe.py    # Fine-tuning script for GE2PE
│   └── finetune-t5.py       # Fine-tuning script for T5

testing-scripts/
│   └── test.ipynb           # Benchmarking the models with SentenceBench Persian G2P Benchmark

assets/
│   └── (files required for inference, e.g., Parsivar, GE2PE.py)

Model Performance

Below are the performance metrics for each model variant on the SentenceBench dataset:

Model	PER (%)	Homograph Acc. (%)	Avg. Inf. Time (s)
GE2PE (Base)	4.81	47.17	0.4464
Homo-T5	4.12	76.32	0.4141
Homo-GE2PE	3.98	76.89	0.4473

Usage

For inference, run the provided inference.ipynb notebook either locally or via the Colab link (recommended for easy setup).

Quick Setup

Install dependencies:
```
pip install unidecode
```

Download models:

git clone https://huggingface.co/MahtaFetrat/Homo-GE2PE-Persian/
unzip -q Homo-GE2PE-Persian/assets/Parsivar.zip
unzip -q Homo-GE2PE-Persian/model-weights/homo-ge2pe.zip -d homo-ge2pe
unzip -q Homo-GE2PE-Persian/model-weights/homo-t5.zip -d homo-t5
mv Homo-GE2PE-Persian/assets/GE2PE.py ./

Fix compatibility (if needed):

sed -i 's/from collections import Iterable/from collections.abc import Iterable/g' Parsivar/token_merger.py

Example Usage

from GE2PE import GE2PE

g2p = GE2PE(model_path='/content/homo-ge2pe') # or homo-t5
g2p.generate(['تست مدل تبدیل نویسه به واج', 'این کتابِ علی است'], use_rules=True)

# Output: ['teste model t/bdil nevise be vaj', '@in ketabe @ali @/st']

Dataset: HomoRich G2P Persian

The models in this repository were fine-tuned on HomoRich, the first large-scale public Persian homograph dataset for grapheme-to-phoneme (G2P) tasks, resolving pronunciation/meaning ambiguities in identically spelled words. Introduced in "Fast, Not Fancy: Rethinking G2P with Rich Data and Rule-Based Models", the dataset is available here.

Citation

If you use this project in your work, please cite the corresponding paper:

@misc{qharabagh2025fastfancyrethinkingg2p,
      title={Fast, Not Fancy: Rethinking G2P with Rich Data and Rule-Based Models}, 
      author={Mahta Fetrat Qharabagh and Zahra Dehghanian and Hamid R. Rabiee},
      year={2025},
      eprint={2505.12973},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2505.12973}, 
}

Contributions

Contributions and pull requests are welcome. Please open an issue to discuss the changes you intend to make.

MahtaFetrat
/

Homo-GE2PE-Persian