OhanASR-v0.1
OhanASR-v0.1 is an Armenian automatic speech recognition model. It transcribes Armenian speech into punctuated, cased Armenian text.
Scores
Evaluated on 10,113 clips across five Armenian speech datasets. Scores are percentages; lower is better. Both columns apply number normalization; the second also applies text normalization.
| Dataset | WER with number normalization (%) | WER with text and number normalization (%) |
|---|---|---|
| Combined | 16.1160 | 6.9648 |
| Common Voice 26 | 9.6194 | 6.2831 |
| FLEURS | 10.7776 | 3.8787 |
| Poems | 24.3490 | 6.2608 |
| Movies | 42.6076 | 20.2110 |
| Infocom May 26 | 15.3296 | 4.0780 |
Installation
Use Linux with Python 3.10 or 3.11 (tested with 3.11; omnilingual-asr 0.2.0 does not install on
Python 3.12.1 or newer). Install the audio library required by fairseq2 and the pinned dependencies from this
repository's requirements.txt:
sudo apt install libsndfile1 # Debian/Ubuntu; with conda: conda install -c conda-forge libsndfile==1.0.31
pip install -r https://huggingface.co/Metric-AI/OhanASR-v0.1/resolve/main/requirements.txt
The checkpoint uses the fairseq2 format rather than the Transformers AutoModel format, and is loaded
with the omnilingual-asr library installed above.
Usage
Download the model snapshot and add its folder to Python’s import path to use the bundled helper:
import sys
from huggingface_hub import snapshot_download
folder = snapshot_download("Metric-AI/OhanASR-v0.1")
sys.path.insert(0, folder)
from ohanasr import OfflineTranscriber, load_model
transcriber = OfflineTranscriber(load_model(folder))
print(transcriber.transcribe("audio.wav"))
Reuse the transcriber for multiple files. The helper reads formats supported by SoundFile, converts audio to mono at 16 kHz, and uses bfloat16 on CUDA with float32 rotary frequency tables. On CPU it uses float32.
Long audio
Audio up to 35 seconds is decoded in one piece. For longer recordings, the helper starts with roughly equal segments and looks for a quiet 0.1-second frame within 3 seconds of each proposed cut. It constrains the cut positions so every segment fits the model’s 40-second limit. The helper decodes segments in batches of up to four and joins their transcripts in order. A word cut at a segment boundary may be transcribed incorrectly.
Files
model.pt: OhanASR-v0.1 model weights (float32).ohanasr_tokenizer.model: matching tokenizer (10,291 tokens).model.json: model configuration, file hashes, and detailed scores.ohanasr.py: model loading, audio preparation, and offline transcription.requirements.txt: pinned Python dependencies.LICENSE: CC BY-NC 4.0 license text.
License
This model is released under the Creative Commons Attribution-NonCommercial 4.0 International
license (CC BY-NC 4.0); see LICENSE. You may use, share and adapt it for non-commercial purposes with
attribution.
Citation
If you use OhanASR-v0.1 in your research, please cite it as follows:
@misc{ohan-asr-v0.1,
title={OhanASR-v0.1: Armenian Speech-to-Text Model},
author={Metric-AI-Lab},
year={2026},
howpublished={\url{https://huggingface.co/Metric-AI/OhanASR-v0.1}},
note={Speech-to-Text Model for Armenian}
}
Acknowledgements
We acknowledge the EuroHPC Joint Undertaking for awarding this project access to the EuroHPC supercomputer LUMI, hosted by CSC (Finland) and the LUMI consortium. The training of OhanASR-v0.1 was run on LUMI.
