Thai has no shortage of text, speech, or documents. It has a shortage of labelled data. We build the supervision ourselves and publish what transfers.
arXiv:2609.03502v118 pages
An 82M model small enough to run on the device, trained entirely on speech a larger model generated, and the benchmark that says what it still gets wrong.
Kunat Pipatanakul, Potsawee Manakul, Warit Sirichotedumrong, Sittipong Sripaisarnmongkol, Pakorn Nathong, Phatrasek Jirabovonvisut
Three of the twelve teacher voices. None was recorded.
arXiv:2609.03595v120 pages
A 0.9B document model taught on 45,723 reconstructed pages, and the controlled experiments that say which part of a synthetic document actually transfers.
Kunat Pipatanakul
The 82M fixed-voice Thai and English student, open weights, for on-device inference.
The 0.9B Thai document model trained without a single real OCR label.
The same weights quantized for llama.cpp.
1,531 sentences, each built around one expression Thai speech models get wrong: a brand, a name, a rare compound, a chat spelling.
210 long sentences for scoring where a Thai voice breaks, which no character-error number measures.
Thai forced alignment scored against exact duration ground truth, for word and token boundaries.
Whether a Thai transcription model uses the vocabulary you give it, and whether it invents words when the list is irrelevant.
Every page Common Crawl read as Thai across 70 monthly crawls, 2019 to 2026: 932 million captures over 6.7 million domains.
The evaluation framework behind the speech report: keyword accuracy, pause placement, speaker similarity, speaking rate.
The document reconstruction pipeline: erase a page's text, fit the label back into the same region, render it in a sampled typeface or real ink.
Reference inference for the speech model.
Reference inference for the document model.
Type Thai, hear the 82M model read it.
Drop a Thai page in, read what the model returns.
This work is a collaboration between Wayu Research and Paxa Labs, self-funded by Wayu Research, with the Typhoon team co-authoring the speech report and reviewing the document one.