Vacana-TextNorm-Khmer · Vacana វចនៈ Khmer text normalization

Work in progress: no model weights here yet. This repo will hold the first model when it is trained and tested.

Vacana ("words") will clean Khmer text in both directions: numbers, dates and money into Khmer words (for speech), and speech-recognition output back into readable text with punctuation and digits. It belongs to the same family as Sota (ASR, listens), Chakkhu (OCR, reads), Tossana (captioning, sees), Akkhara (keyboard, types) and VoxPocket (TTS, speaks).

Planned design (≈4M parameters, estimate)

• Words → text: punctuation and digits for Sota transcripts
• Text → words: numbers, dates, money and units read out in Khmer, for VoxPocket
• Starts from the rule-based number speller used to build Sota's data

Status

Step	Result
Rule-based number speller	done (inside the Sota data tools)
First model	soon

License

To be decided with the first release.

Citation

@misc{ing2026vacana,
  author    = {Ing, Muyleang},
  title     = {Vacana: Khmer text normalization},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/muyleanging/Vacana-TextNorm-Khmer}
}