Keyword: Khmer language, Speech recognition, Inverse text normalization
Khmer Automatic Speech Recognition (ASR) systems often produce raw text that lacks punctuation, spaces, and numbers, making it less readable. We propose a method to improve the readability of these outputs by utilizing XLM-ROBERTa, a multilingual transformer model trained on 100 languages, including Khmer. Our approach enhances an existing punctuation restoration framework by adding a number recognition component. Trained on Khmer news data, the model transforms raw ASR outputs into more polished, reader-friendly text. The code is publicly available for further research and development.
The Khmer language, like Thai, Lao, and Japanese, does not use explicit word boundaries as seen in languages like English or French. Instead, readers rely on contextual cues to recognize word separations. In Khmer, spaces serve a unique purpose. While they do not indicate word boundaries, their placement carries various meanings, depending on context. Spaces are primarily used to mark pauses, define boundaries, and enhance readability. When working with text output from ASR (Automatic Speech Recognition) models, the text is typically in a normalized form. Therefore, to make it suitable for practical use, it's necessary to restore elements like spaces, punctuation, numbers, and symbols.
In this work, we employ word boundary detection using khmercut, a tool designed for segmentation of Khmer text into word units to make it ready to feed into the model.
Our model architecture is based on XLM-ROBERTA, a multilingual transformer model that serves as the encoder. On top of this, we integrate a bidirectional Long Short-Term Memory (BILSTM) network to capture contextual dependencies. Following the BILSTM, we apply a single fully connected linear layer. The final output comprises two independent classification heads.
We extend the work of xashru's punctuation restoration model by introducing an additional classification head for number recognition. This results in a model capable of handling two types of outputs: punctuation (including space, comma, question mark, exclamation mark, and other symbols) and numeric entities.
A post-processing step is employed after classification, where predicted tokens are converted back into their corresponding numbers or symbols, ensuring proper representation of the original text. This enhances the accuracy of both punctuation restoration and numeric entity recognition.
The model was trained on publicly available Khmer news data sourced from a variety of websites, consisting of a total of 1.5 million tokens. This dataset includes diverse linguistic contexts, contributing to the robustness of the model in handling punctuation and number recognition tasks in the Khmer language.
The training process yielded an accuracy of 97.2%, demonstrating the effectiveness of the architecture in accurately predicting both punctuation and numeric entities in Khmer text.
The example format can be found in the ./data folder
ឌុប 0 0
នាង 0 0
ខារ៉ាអូខេ 0 0
ឡាន 0 0
បុក 0 0
ពីក្រោយ 0 0
ស្លាប់ 0 0
ប្រុស 0 0
ស្នេហ៍ 0 0
របួស 0 0
ធ្ងន់ ។ 0
គ្រោះថ្នាក់ 0 0
ចរាចរណ៍ 0 0
មួយ 0 0
បាន 0 0
កើតឡើង 0 0
កាលពី 0 0
វេលា 0 0
ម៉ោង 0 0
ដប់ 0 NUMBER_B
មួយ 0 NUMBER_I
និង 0 0
ម្ភៃ 0 NUMBER_B
នាទី 0 0
យប់ 0 0
ថ្ងៃទី 0 0
ពីរ SPACE NUMBER_B
ឧសភា SPACE 0
ពីរ 0 NUMBER_B
ពាន់ 0 NUMBER_I
ដប់ 0 NUMBER_I
ពីរ SPACE NUMBER_I
@misc{khmertagger2025,
author = {Seanghay Yath},
title = {KhmerTagger: Inverse Text Normalization for Khmer Automatic Speech Recognition},
year = {2025},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/seanghay/khmertagger}},
note = {Open source project for Khmer punctuation restoration and number recognition using XLM-ROBERTa}
}