Introducing OhanASR: Open Armenian Speech Recognition Models

Community Article
Published November 20, 2026

OhanASR-Banner

We're excited to introduce OhanASR, a family of open-weight Automatic Speech Recognition (ASR) models for Armenian, named after Dzenov Ohan, the legendary character from the Armenian epic Daredevils of Sassoun, known for his powerful voice.

Our first release includes two models: OhanASR-v0.1, optimized for accurate transcription of recorded speech, and OhanASR-Streaming-v0.1, designed for real-time speech recognition.

OhanASR-v0.1 is built on a 2-billion-parameter multilingual model, fine-tuned on 820 hours of Armenian speech collected from various sources, covering formal speech, everyday conversations, and literary readings. OhanASR-Streaming-v0.1 builds upon this model with additional training to adapt it for streaming transcription.

Our goal is to make high-quality Armenian speech recognition more accessible to developers, researchers, and anyone building applications for the Armenian language.

How Well Do They Perform?

We evaluated both models on ArmBench-ASR, a benchmark covering different types of Armenian speech.

OhanASR-v0.1

OhanASR-v0.1 achieves a combined Word Error Rate (WER) of 16.12% and Character Error Rate (CER) of 3.49%, ranking 2nd by WER and 1st by CER among both open and closed models on ArmBench-ASR.

Dataset WER WER-norm CER CER-norm WER-LLM CER-LLM
Combined 16.12% 6.96% 3.49% 1.76% 6.04% 1.60%
Common Voice 26 9.62% 6.28% 1.77% 1.23% 4.73% 1.04%
FLEURS 10.78% 3.88% 2.03% 0.93% 3.16% 0.82%
Poems 24.35% 6.26% 5.29% 1.44% 5.77% 1.33%
Movies 42.61% 20.21% 15.10% 8.98% 21.47% 9.68%
News 15.33% 4.08% 3.07% 1.17% 2.95% 0.85%

OhanASR-Streaming-v0.1

Alongside the standard model, we're releasing OhanASR-Streaming-v0.1, built on OhanASR-v0.1 with additional training to adapt it for streaming speech recognition.

The model is designed to transcribe speech incrementally as audio arrives, making it suitable for live transcription and other real-time applications.

The streaming model achieves a combined WER of 19.07% and CER of 4.47%, ranking 10th by WER and 9th by CER among open and closed models on ArmBench-ASR.

Dataset WER WER-norm CER CER-norm WER-LLM CER-LLM
Combined 19.07% 8.57% 4.47% 2.46% 7.57% 2.23%
Common Voice 26 11.91% 7.36% 2.20% 1.47% 5.62% 1.25%
FLEURS 13.30% 4.83% 2.52% 1.20% 4.21% 1.10%
Poems 30.10% 7.96% 6.90% 2.16% 7.57% 2.05%
Movies 45.00% 24.50% 20.44% 14.51% 26.65% 14.78%
News 18.02% 6.14% 3.51% 1.54% 4.23% 1.16%

Lower scores indicate better transcription accuracy. WER-norm and CER-norm reflect normalized evaluation.

Listen to OhanASR in Action

OhanASR-v0.1

Listen to an Armenian speech sample and see the model's transcription.

Model prediction:

NVIDIA OpenAIA AWS Mosilla Revolute և տասնյակ այլ միջազգային տեխնոլոգիական անուններ մեկ վայրում Երևանում։ Առաջիկայում դիջիթեք 2026-ը Հայաստանի մայրաքաղաքը կվերածի արհեստական բանականության տարածաշրջանային հանդիպման հարթակի՝ միավորելով տեխնոլոգիան, բիզնեսը և ներդրումային նոր հնարավորությունները։ Դիջիթեք 2026-ն այս տարի վերադառնում է արդեն նոր մասշտաբով՝ աշխարհի նոր ԱԲ հանգույցը խորագրով։ Միացյալ Նահանգներից, Եվրոպայից, արաբական ծոցի երկրներից և կենտրոնական Ասիայից Երևան կժամանեն ձեռներեցներ, ներդրողներ, տեխնոլոգիական ոլորտի առաջատար ընկերությունների և պետական կառույցների ներկայացուցիչներ։

OhanASR-Streaming-v0.1

Watch the streaming model transcribe Armenian speech. The video includes subtitles displaying the model's transcription output.

While there's still room for improvement, these results show that both models are competitive options for Armenian speech recognition, offering solutions for both recorded audio and streaming transcription.

Available for Everyone

We're releasing both OhanASR models under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license, making them freely available for research, educational, and other non-commercial use.

Whether you're researching Armenian speech recognition, building experimental transcription tools, or exploring voice-powered applications, we hope OhanASR can serve as a useful starting point.

Known Issues

While both models deliver competitive results, there are still some known limitations:

  • Formal and colloquial Armenian: The models may occasionally confuse formal and colloquial forms, including variations such as «ա» and «է».
  • Code-switching: Recognition of speech mixing Armenian with other languages is not always reliable. The models may sometimes transcribe foreign-language words in Armenian script.
  • Unclear speech: The models may occasionally omit words that are difficult to recognize, particularly in noisy or unclear audio.

What's Next?

This is just the first release. We'll continue improving OhanASR, focusing on better recognition accuracy, lower streaming latency, and more robust transcription across different speech conditions.

We're excited to see what the community builds with OhanASR and welcome feedback.

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 and OhanASR-Streaming-v0.1 was run on LUMI.

Give OhanASR a try, and help us make Armenian speech technology better for everyone!

Community

Sign up or log in to comment