Omnilingual-ASR-Khm

A CTC speech-recognition model for Khmer, built on top of Meta's Omnilingual ASR 300M encoder with a custom cluster-level vocabulary and a self-conditioned CTC head.

Training

The model is produced in two phases:

1. Pretrain -- the CTC head and encoder are trained from scratch (the encoder initialized from the public omniASR_CTC_300M checkpoint) on a Khmer speech corpus.
2. Fine-tune -- the pretrained checkpoint is further trained on Khmer-English code-switching data.

Both phases' configs and training scripts are in this model's source repository.

Results

CER on held-out evaluation sets (lower is better):

split	CER	WER
val	2.24%	7.62%
test_fleurs	12.03%	25.66%
test_slr42	14.82%	48.62%

Usage

from huggingface_hub import hf_hub_download
from safetensors.torch import load_file

weights = hf_hub_download("Darayut/Omnilingual-ASR-Khm", "model.safetensors")
config = hf_hub_download("Darayut/Omnilingual-ASR-Khm", "config.json")
vocab = hf_hub_download("Darayut/Omnilingual-ASR-Khm", "vocab.json")

See scripts/inference/transcribe.py in the source repository for a complete loading + decoding example.

License

MIT.