Khmer Pronunciation (G2P)

A small Transformer (7.5M parameters) that converts Khmer words to their NCKL Khmer-script respelling, with - between syllables:

ក្របី → ក្រ-បី

Training data: phonsobon/khmer-G2P. Trained model: phonsobon/khmer-g2p (83.2% word accuracy, 4.2% character error rate on the test split).

Install

pip install khmer-g2p

khmer-g2p ក្របី កម្ពុជា          # ក្របី  ក្រ-បី / កម្ពុជា  កាំ-ពុ-ជា

from khmer_g2p import KhmerG2P
g2p = KhmerG2P()                 # downloads the model from the Hub on first use
g2p("ក្របី")                     # 'ក្រ-បី'
g2p(["សម្ដេច", "កម្ពុជា"])        # list in, list out

Train on Google Colab

Choose Runtime → Change runtime type → GPU, then run the cells in order.

Train locally

pip install -r requirements.txt
python g2p/prepare_data.py               # writes g2p/data/{train,validation,test}.tsv
python g2p/train.py --epochs 100         # best model -> g2p/checkpoints/best.pt

Options: --batch 256, --lr 5e-4, --eval-every 5, --patience 4 (stop after this many validation checks without improvement).

Use the training checkpoint directly

python g2p/predict.py ក្របី កម្ពុជា

from predict import KhmerG2P   # run from inside g2p/
g2p = KhmerG2P()
g2p("ក្របី")                  # 'ក្រ-បី'
g2p(["សម្ដេច", "កម្ពុជា"])     # list in, list out

Release

pip install -e ".[dev]"
python scripts/push_to_hub.py            # g2p/model/best.pt -> safetensors -> phonsobon/khmer-g2p
python -m pytest tests                   # tests the model on the Hub
python scripts/evaluate.py               # word accuracy on the test split
python -m build && twine upload dist/*   # publish to PyPI (bump __version__ first)