Applied AI Research Portfolio Blog Get In Touch

Introducing OhanASR v0.1

State-of-the-art open speech recognition for Armenian — one model for recorded audio, one for live transcription.

#1CER, open & closed
3.49%Character error rate
16.12%Word error rate
820 hArmenian speech
OhanASR v0.1 — Open Armenian Speech Recognition, non-streaming and streaming

Overview

Today we're releasing OhanASR, a family of open-weight speech recognition models for Armenian. The flagship, OhanASR-v0.1, is the strongest open Armenian ASR model available: on ArmBench-ASR it ranks first by character error rate and second by word error rate across every system we evaluated — open or closed.

The family is named after Dzenov Ohan — Ohan the Loud-Voiced — from the Armenian epic Daredevils of Sassoun, whose shout could be heard across the mountains. It felt right for models whose whole job is listening to Armenian.

Armenian is spoken by millions of people, yet it has long been an afterthought for speech technology: large multilingual systems treat it as one of a hundred languages, and the open options have lagged far behind commercial APIs. OhanASR closes that gap and puts a state-of-the-art model in everyone's hands.

Two models, one family

We start from a 2-billion-parameter multilingual speech model and fine-tune it on 820 hours of Armenian speech spanning formal speech, everyday conversation and literary readings. From there the family splits in two:

  • OhanASR-v0.1 — the accuracy-first model for recorded audio: interviews, lectures, podcasts, archives, call recordings.
  • OhanASR-Streaming-v0.1 — continued training on top of OhanASR-v0.1 so it can emit text incrementally as audio arrives. Built for live captions, voice agents and anything that can't wait for the speaker to finish.

Results

We evaluate on ArmBench-ASR, our open benchmark for Armenian speech recognition. It pools five very different kinds of audio — crowd-sourced read speech (Common Voice 26), FLEURS, recited poetry, film dialogue and broadcast news — so a model can't win by being good at just one register.

Overall

ModelWERCERWER-normCER-normRank (WER / CER)
OhanASR-v0.116.12%3.49%6.96%1.76%2nd / 1st
OhanASR-Streaming-v0.119.07%4.47%8.57%2.46%10th / 9th

Combined ArmBench-ASR scores; lower is better. Ranks are among all open and closed models on the leaderboard. "-norm" scores are computed after text normalization (punctuation, casing and orthographic variants), which better reflects how readable a transcript is.

The streaming model gives up about three points of WER in exchange for real-time output — a modest price, and it still places in the top ten overall, ahead of many systems that only work offline.

By domain

Read speech and news are close to solved: on FLEURS, OhanASR-v0.1 gets fewer than one character in a hundred wrong after normalization. Poetry is harder on raw WER — mostly punctuation and archaic spellings, which normalization largely forgives. Film dialogue, with overlapping voices, music and slang, remains the real frontier.

Figure 1. Raw (un-normalized) WER and CER per ArmBench-ASR domain. Use the WER / CER toggle to switch metrics.
DomainWER-normCER-normWER-LLMCER-LLM
Common Voice 266.28%1.23%4.73%1.04%
FLEURS3.88%0.93%3.16%0.82%
News4.08%1.17%2.95%0.85%
Poems6.26%1.44%5.77%1.33%
Movies20.21%8.98%21.47%9.68%
Combined6.96%1.76%6.04%1.60%

OhanASR-v0.1, normalized and LLM-judged scores per domain. "-LLM" scores use a language model to normalize references and hypotheses before scoring, discounting differences a human reader wouldn't count as errors. Full per-domain numbers for both models are on the model cards.

Hear it

A news-style clip — dense with brand names, numbers and foreign terms, exactly the kind of audio that trips up speech recognizers.

OhanASR-v0.1 transcript 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։

And the streaming model in action — the subtitles below are OhanASR-Streaming-v0.1's live output, produced as the audio plays.

Figure 2. Real-time Armenian transcription with OhanASR-Streaming-v0.1.

Known limitations

This is a v0.1, and we'd rather you hear about the rough edges from us:

  • Formal vs. colloquial Armenian. The models sometimes mix registers — for example writing «ա» where the formal form is «է».
  • Code-switching. Speech that mixes Armenian with English or Russian isn't always handled well; foreign words may come out transliterated into Armenian script.
  • Unclear audio. In noisy or mumbled segments the models may drop words they can't resolve rather than guess.

Get the models

Both models are released under CC BY-NC 4.0 — free for research, education and other non-commercial use. For commercial licensing or a custom deployment, get in touch.

What's next

We're already training the next version, focused on three things: higher accuracy — especially on conversational and film audio — lower streaming latency, and robustness to noise, accents and code-switching. OhanASR is also one half of a voice stack: pair it with OhanTTS for speech generation and Ayb for language understanding.

Try it, break it, and tell us what you find — feedback is how v0.2 gets better.