ReNikud - Audio-Supervised Hebrew Grapheme-to-Phoneme

Convert unvocalized Hebrew text into IPA for TTS, speech technology, and spoken-language research

🌐 Project Page   |   📄 Research Paper

---

Features

• Context-aware Hebrew G2P (text → IPA)
• Letter-level constrained decoding
• Passthrough for non-Hebrew text
• Runs on ONNX Runtime (no PyTorch)
• ~20 MB, real-time inference

Usage

Inference is published as renikud-onnx on PyPI. Install and download the ONNX weights from Hugging Face (they are not bundled with the wheel):

pip install renikud-onnx
wget https://huggingface.co/thewh1teagle/renikud/resolve/main/model.onnx -O model.onnx

from renikud_onnx import G2P

g2p = G2P("model.onnx")
print(g2p.phonemize("שלום עולם"))
# → ʃalˈom ʔolˈam

See renikud-onnx/README.md for the same install / download / usage flow. For Rust inference, see renikud-rs/.

Architecture

See docs/ARCHITECTURE.md for model design and implementation details.

Training

See docs/TRAINING.md for data preparation, training commands, upload/download, ONNX export, benchmark, and hyperparameters.

License

This project is licensed under the Creative Commons Attribution 4.0 International license (CC BY 4.0). See LICENSE.

Citation

@misc{melichov2026renikud,
  title={ReNikud: Audio-Supervised Hebrew Grapheme-to-Phoneme Conversion},
  author={Maxim Melichov and Yakov Kolani and Morris Alper},
  year={2026},
  url={https://arxiv.org/pdf/2606.20179},
}