Akkhara-mini (18M)

Akkhara-mini converts Sing Khmer (Khmer typed with Latin letters, as in chat) into Khmer script with a ranked top-5 list, like a typing assistant (IME). It is small and fast, made for phones and keyboards, and runs fully offline in the Akkhara keyboard.

input	output (#1)
nh jg tv sala rean	ខ្ញុំចង់ទៅសាលារៀន
b pg lvol ot torn tomne te o	បងកំពុងរវល់អត់ទាន់ទំនេរទេអូន
send link mk b	send link មកបង
sok sbay ot bro	សុខសប្បាយអត់ bro

English chat words stay as typed (benchmark, laptop, deadline), with the Khmer loan spelling in the top-5 when one exists (deadline → ដេតឡាញ, internet → អ៊ីនធឺណិត).

Benchmark

Both Akkhara models on the same tests (same beam search, 5 beams): a held-out test split (7,821 sentences), 39 real chat sentences typed and labeled by a native speaker, 86 context tests (the same chat word read from its neighbours, e.g. ng = និង / នឹង / ហ្នឹង) and 7 glued-word tests.

model	test top-1	test top-5	test CER	real chat top-1 / top-5	context #1 / top-5	glued
akkhara-mini (18M)	58.1%	76.6%	4.55%	85% / 92%	80 / 99%	6/7
akkhara-byt5-small (300M)	62.0%	79.7%	3.73%	87% / 95%	80 / 98%	7/7

• Real chat and context words: about as good as the big model (85–87% right at #1, 81 of 86 context words), while being 16× smaller.
• Chat abbreviations: tmr → ស្អែក, sry → សុំទោស, gn → រាត្រីសួស្តី, btw → ម្យ៉ាងទៀត, wyd → ធ្វើអីហ្នឹង.
• Formal news-style text: the bigger INGMUYLEANG/akkhara-byt5-small is a little more accurate.
• CER = share of wrong letters (lower is better). Top-5 = the right sentence is among the 5 suggestions.

Use

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

model_id = "INGMUYLEANG/akkhara-mini"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSeq2SeqLM.from_pretrained(model_id)

inputs = tokenizer(["nh jg tv sala rean"], return_tensors="pt")
out = model.generate(**inputs, num_beams=5, num_return_sequences=5, max_new_tokens=160)
print(tokenizer.batch_decode(out, skip_special_tokens=True))
# ['ខ្ញុំចង់ទៅសាលារៀន', ...]

# fastest (typing): greedy #1 only
out = model.generate(**inputs, num_beams=1, max_new_tokens=160)

For long text, cut it into ~6-word chunks and join the results.

ONNX (phones, web, other languages)

The onnx/ folder holds an ONNX Runtime export: encoder.onnx + decoder.onnx, or the 8-bit encoder.int8.onnx + decoder.int8.onnx (18.5 MB, used in the keyboard), plus tokenizer.json and abbreviations.tsv (46 abbreviations → full Khmer name: the keyboards show mptc as typed at #1 and ក្រសួងប្រៃសណីយ៍និងទូរគមនាគមន៍ at #2; a plain list, not learned by the model). Run the encoder once per sentence, then the decoder once per Khmer letter (greedy or beam search). On a recent Android phone the int8 model gives the #1 suggestion in about 10 ms.

Model details

• Developed by: Ing Muyleang
• Model type: T5-style encoder–decoder, character tokens (6 encoder + 3 decoder layers, d_model 384)
• Parameters: 17.8M · License: Apache-2.0
• Language: Khmer (km); English chat words in mixed input stay English
• Training data: 398k Sing Khmer → Khmer pairs (incl. 134 Cambodian ministry, university and bank names), collected, generated, cleaned and checked by the author with AI tools (agy, Claude); see INGMUYLEANG/akkhara-sing-khmer (access on request).
• Demo: INGMUYLEANG/akkhara-demo
• Keyboard app (APK): latest release

Limits

• Trained mostly on synthetic pairs; the real-chat test set is small (39 sentences).
• Short abbreviations stay ambiguous (ng, yt, dl, bn); the right answer is often in the top-5.
• It knows rude chat words (they occur in real chat).

Contact

Training details (hardware, time, settings) are not published. Questions: contact the author through the Hugging Face profile or GitHub.

Citation

@misc{akkhara2026mini,
  author    = {Ing, Muyleang},
  title     = {Akkhara-mini: a small character-level Sing Khmer to Khmer script transliteration model},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/INGMUYLEANG/akkhara-mini}
}