Research
How our models are built and tested, and where they still fall short. Every paper comes with the code, data and weights behind it, so you can check it yourself.
Paper/
Sota
A 204M-parameter speech recogniser for Khmer and English, and a 73M-parameter one, made by structurally pruning Meta's Omnilingual ASR to 80% encoder sparsity. The larger has a character error rate of 1.77% on the DDD Khmer test set, below its 975M-parameter teacher's 2.47%, at a quarter of the compute.
- Parameters
- 73M / 204M
- CER
- 1.77%
Earlier research
2026
Paper
Netra OCR 1
A 19M-parameter text-line recognizer trained only on synthetic images. It has the lowest character error rate of five systems on every test set, including vision-language models about 150 times its size.
Paper
Netra MT 1
A 77M-parameter translation model trained from scratch on 220 million tokens of English–Khmer text. It beats open models up to 40 times its size on English→Khmer overlap metrics and runs 16 times faster than NLLB-600M on a CPU, but still trails them on Khmer→English.
Paper
AkaraAlpha
A 78.7K-parameter attention-based bidirectional GRU that converts Khmer typed in the Latin alphabet back into Khmer script. It has a lower character error rate than a Transformer three times its size, and a dictionary check makes its first suggestion right 74% of the time.