Downloads · 30 days
29
46% of all-time downloads
Darayut/Omnilingual-ASR-Khm
Omnilingual-ASR-Khm is a automatic speech recognition model from Darayut. Use it when you need speech turned into text. The card lists the license as mit.
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.
Downloads · 30 days
29
46% of all-time downloads
All-time downloads
63
Public
Parameters
320M
1.3 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors1.3 GB · 100%
From the Hugging Face model README
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.
The model is produced in two phases:
omniASR_CTC_300M checkpoint) on a
Khmer speech corpus.Both phases' configs and training scripts are in this model's source repository.
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% |
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.
MIT.