Tulā ទុលា
A Khmer AI benchmark: every model scored on the same held-out test sets, with the same scoring rules.
Speech recognition (ASR)
CER (character error rate, lower is better) with spaces and punctuation removed. Test set: 12.0 h, 6,513 clips, 16 held-out speakers, real speech only.
| Model | Params | Decoding | CER |
|---|---|---|---|
| muyleanging/Sota-ASR-Khmer | ≈26M | beam 5 | 3.40% |
| muyleanging/Sota-ASR-Khmer | ≈26M | greedy | 3.62% |
| seanghay/Qwen3-ASR-0.6B-Khmer | 0.6B | default | 13.20% |
| seanghay/whisper-small-khmer-v2 | ≈244M | default | 18.57% |
Read this fairly: the test speakers come from the same data family Sota was trained on (different speakers, clean read speech), which favours Sota. whisper-small-khmer-v2 was trained on OpenSLR 42, part of this test set. Phone, noisy and natural speech are not scored yet. More models (Whisper, wav2vec2 and Qwen fine-tunes from the Khmer community): soon
Outside test: FLEURS km (655 clips, 2.4 h)
| Model | Decoding | CER |
|---|---|---|
| muyleanging/Sota-ASR-Khmer (ONNX) | greedy | 9.65% |
| Other models | soon |
Coming boards
- OCR (Khmer printed text) soon
- Khmer ↔ English translation soon
- Natural / phone / noisy speech soon
By Ing Muyleang · part of the Indriya ឥន្ទ្រិយ Khmer model family.