Tha (ថា)

Khmer text normalization: turn written text into spoken words for TTS, and spoken words back into written text for ASR.

Paper: arXiv:2609.30984

Install

pip install tha

Tha needs pynini. On macOS, install it from conda-forge first:

conda install -c conda-forge pynini

Written to spoken

import tha

tha.normalize_text("ទិញនៅ 02/01/2024 តម្លៃ $1.05")
# ទិញនៅ ថ្ងៃទី▁ពីរ▁ខែ▁មករា▁ឆ្នាំ▁ពីរពាន់▁ម្ភៃបួន តម្លៃ មួយ▁ដុល្លារ▁ប្រាំ▁សេន

Input	Output
1,234	មួយពាន់▁ពីររយ▁សាមសិបបួន
12.5%	ដប់ពីរ▁ចុច▁ប្រាំ▁ភាគរយ
5km	ប្រាំ▁គីឡូម៉ែត្រ
10:23	ម៉ោង▁ដប់▁ម្ភៃបី▁នាទី
012 345 678	សូន្យ▁ដប់ពីរ▁សាមសិបបួន▁ហាសិបប្រាំមួយ▁ចិតសិបប្រាំបី
-5	ដក▁ប្រាំ

It also handles ordinals, fractions, phone numbers, e-mails, license plates, ranges, scores and more. Words of one number are joined with ▁, which you can change:

normalizer = tha.Normalizer(separator=" ")  # reuse it, loading takes ~2s
normalizer.normalize("$1.05")  # មួយ ដុល្លារ ប្រាំ សេន

Spoken to written

import tha

tha.inverse_normalize_text("ទិញនៅថ្ងៃទីពីរខែមករាឆ្នាំពីរពាន់ម្ភៃបួនតម្លៃមួយដុល្លារប្រាំសេន")
# ទិញនៅ02/01/2024តម្លៃ$1.05

Input	Output
ដប់ពាន់	10,000
ម៉ោងប្រាំបីកន្លះ	8:30
ដកដប់អង្សាសេ	-10°C
ម្ភៃដកដប់	20 - 10
មួយចំនួន	មួយចំនួន
ពីរោះ	ពីរោះ

A single number word on its own stays a word, and number words inside other words are left alone. For Khmer digits, use tha.InverseNormalizer(khmer_digits=True).

Text cleanup

import tha.normalize
import tha.hashtags
import tha.parenthesis

tha.normalize.processor("មិន​ឲ្យ")                  # មិនឱ្យ
tha.hashtags.processor("Hello world #លុប hello")           # Hello world  hello
tha.parenthesis.processor("Hello (ignored) world")        # Hello world

Development

uv sync
make test

On macOS, pynini is built from source, so install OpenFst first:

brew install openfst
CPPFLAGS="-I$(brew --prefix)/include" LDFLAGS="-L$(brew --prefix)/lib" uv sync

Citation

@misc{yath2026tha,
  author        = {Yath, Seanghay},
  title         = {{THA}: Weighted Finite-State Text Normalization and Inverse Text Normalization for Khmer},
  year          = {2026},
  eprint        = {2609.30984},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  doi           = {10.48550/arXiv.2609.30984},
  url           = {https://arxiv.org/abs/2609.30984}
}