AminMadani/persian-byt5-g2p

Persian ByT5 G2P (Grapheme-to-Phoneme)

  

A lightweight, robust Grapheme-to-Phoneme (G2P) model for Persian, fine-tuned on the comprehensive KaamelDict dictionary (~116,600 phonetic entries) using Google's byte-level google/byt5-small architecture.

🚀 Why Byte-Level G2P for Persian?

Traditional word-piece / subword tokenizers struggle with Persian orthography due to rich morphology (prefixes, suffixes, compound words) and informal spacing.

ByT5 operates directly on raw UTF-8 bytes, providing key advantages:

• —Zero Out-of-Vocabulary (OOV): Transcribes completely unseen words, loanwords, proper nouns, and neologisms.
• —Implicit Diacritic & Short Vowel Inference: Accurately discovers unwritten short vowels (harakat like a, e, o), the grammatical connector (ezafe), and gemination (tashdid).
• —Historical Orthography Resolution: Corrects anomalies such as خوش $\to$ /x o S/ (pronounced with o instead of v).
• —Lightweight & Production Ready: Only ~300M parameters, ideal for real-time TTS synthesis pipelines (e.g., Piper TTS, VITS) on CPU and consumer GPUs.

📊 Benchmark & Training Metrics

• —Base Architecture: google/byt5-small (300M parameters)
• —Dataset: MahtaFetrat/KaamelDict (~116k entries, 95% train / 5% test split)
• —Precision: Full Precision (FP32)
• —Batch Size: 16 per device with 2 gradient accumulation steps (effective batch size = 32)
• —Learning Rate: 5e-4 with AdamW
• —Training Epochs: 3 epochs (~10,386 steps)
• —Final Validation Loss: `0.03755`

🎯 Official Sentence-Level Benchmark (SentenceBench - ICASSP 2025)

Evaluated on the official SentenceBench (ICASSP 2025) test suite across 400 complex Persian sentences using the word-level batched inference pipeline:

• —Homograph Word Accuracy (Context-Free Prior): 41.98% (89 / 212)
• —Sample Exact Matches (0.0% PER): «آیا باید حقیقت را به آنها بگوییم؟» $\to$ ?AyAbAyadhaqiqatrAbe?AnhAbeguyim (100% exact match).
• —Complex Multi-Clause Sentences (< 4.0% PER): «که پیش از انقلاب به خوابگاه دختران و زنان نابینا اختصاصیافته بود...» $\to$ 3.8% PER.

Sample Transcriptions

🛠️ Installation

git clone https://github.com/aminmadaniofficial/persian-byt5-g2p.git
cd persian-byt5-g2p
pip install -r requirements.txt

💻 Usage

1. Python API

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
import torch

MODEL_ID = "AminMadani/persian-byt5-g2p"

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForSeq2SeqLM.from_pretrained(MODEL_ID)

device = "cuda" if torch.cuda.is_available() else "cpu"
model = model.to(device)

def text_to_phonemes(text: str) -> str:
    inputs = tokenizer(text, return_tensors="pt").to(device)
    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_length=128,
            num_beams=2,
            early_stopping=True
        )
    return tokenizer.decode(outputs[0], skip_special_tokens=True)

print(text_to_phonemes("دانشگاه"))
# Output: d A n e S g A h

Or using the built-in helper class:

from inference import PersianG2P

g2p = PersianG2P()
print(g2p.predict(["دانشگاه", "کامپیوتر"]))
# Output: ['d A n e S g A h', 'k A m p i y u t e r']

2. Interactive CLI

# Run interactive transcription shell
python inference.py --interactive

# Transcribe a single word
python inference.py --text "دانشگاه"

📓 Reproduce Training

The complete training script with Google Drive export and Hugging Face Hub integration is provided in notebook.ipynb. You can run it directly on Google Colab's free T4 GPU tier.

📖 Citation

If you use this model or the KaamelDict dataset, please cite the original papers:

@inproceedings{qharabagh2025llm,
  title={LLM-Powered Grapheme-to-Phoneme Conversion: Benchmark and Case Study},
  author={Qharabagh, Mahta Fetrat and Dehghanian, Zahra and Rabiee, Hamid R},
  booktitle={ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages={1--5},
  year={2025},
  organization={IEEE}
}

@article{xue2022byt5,
  title={ByT5: Towards a token-free future with pre-trained byte-to-byte models},
  author={Xue, Linting and Barua, Aditya and Constant, Noah and Al-Rfou, Rami and Narang, Sharan and Kale, Mihir and Roberts, Adam and Raffel, Colin},
  journal={Transactions of the Association for Computational Linguistics},
  volume={10},
  pages={291--306},
  year={2022}
}

📜 License

This project is released under the GNU General Public License v3.0 (GPL-3.0), consistent with the KaamelDict source dataset.