Akhara: A Study on a Style-Conditioned Latent Diffusion Model for Khmer Handwriting Synthesis

Khmer, which is a language spoken by more than 16 million speakers, is considered a low-resource language when it comes to Optical Character Recognition (OCR) and Handwritten Text Recognition (HTR) because of the limited availability of annotated handwriting data and the high cost of manually annotating writing systems that consist of consonant stacking, ligatures, and diacritics.Akhara presents the DiffusionPen framework with an approach that utilizes a CANINE-C character-level encoding scheme without tokenization and a writer-specific MobileNetV2-based style extractor to generate labeled synthetic Khmer handwriting from Unicode text. This paper presents results of applying the presented solution to the task of creating synthetic handwriting for the Khmer language. The quality of generated images was estimated based on visual similarity (Fréchet Inception Distance), OCR readability (Tesseract), and human evaluation.The results have indicated that latent diffusion under conditional style is a scalable approach to manual annotation, which can lead to better performance in OCR/HTR systems of Khmer.