Metric AI Releases OhanTTS and OhanASR, Armenian Speech Models Under a Non-Commercial License; the Accuracy Figures Are the Developer’s Own

Metric AI published three Armenian speech models on Hugging Face on 20 November 2026, according to its announcement posts: OhanTTS-v0.1 for text-to-speech, and OhanASR-v0.1 and OhanASR-Streaming-v0.1 for speech recognition. Metric AI describes itself on Hugging Face as an AI research lab based in Yerevan, Armenia. Its OhanASR post says the family is named after Dzenov Ohan, a character in the Armenian epic Daredevils of Sassoun known for his powerful voice. All figures below are Metric AI’s own.
OhanTTS-v0.1: text-to-speech
- What it is: an Armenian-focused speech generator that also speaks English, Spanish, Arabic, Russian, Mandarin Chinese, Japanese and French. Metric AI says it builds on a multilingual speech model and was further trained on about 160 hours of reference recordings across the eight languages, 61 hours of them Armenian. Hugging Face lists the model at 0.8 billion parameters.
- Voices: it supports voice cloning from a 3 to 10 second audio sample plus its transcript, and can create a voice from a short text description with no recording. Metric AI asks that reference voices be used only with the speaker’s permission.
- Streaming: the model sends audio as it is generated, although the full text must be submitted first. On a single shared NVIDIA GB10, Metric AI measured 331 milliseconds to first audio with one client (real-time factor 0.178) and 2,169 milliseconds with 16 concurrent clients (0.841). It says results depend on hardware, request length and server load.
- Quality: on a held-out set of Armenian requests with unseen speakers and texts, Metric AI reports a 14.70% word error rate, a 6.15% character error rate and a speaker-similarity score (SIM-o) of 0.693. The error rates compare an automatic transcription of the generated speech with the requested text.
- Serving: it runs through SGLang-Omni with an OpenAI-compatible /v1/audio/speech endpoint, according to the post.
- Known issues listed by Metric AI: each request produces up to about 20 seconds of audio; Armenian punctuation marks are not always reflected in pauses, stress and intonation; rare names and foreign words may be mispronounced; switching languages within one utterance can cause errors; and results depend on a clean reference recording with an exact transcript.
OhanASR: speech recognition
- OhanASR-v0.1 is meant for recorded speech. Metric AI says it is built on a 2-billion-parameter multilingual model fine-tuned on 820 hours of Armenian speech covering formal speech, everyday conversations and literary readings.
- OhanASR-Streaming-v0.1 is built on that model with additional training for transcribing speech incrementally as audio arrives.
- Results: on ArmBench-ASR, OhanASR-v0.1 has a combined word error rate of 16.12% and character error rate of 3.49%, which Metric AI says ranks second by word error rate and first by character error rate among open and closed models. The streaming model has 19.07% and 4.47%, which it says ranks 10th and 9th. By dataset, v0.1 scores a word error rate of 9.62% on Common Voice 26, 10.78% on FLEURS, 24.35% on poems, 42.61% on movies and 15.33% on news; the streaming model scores 11.91%, 13.30%, 30.10%, 45.00% and 18.02%. Lower is better.
- Known issues listed by Metric AI: the models can confuse formal and colloquial Armenian forms, including «ա» and «է»; recognition of speech mixed with other languages is not always reliable and foreign words may be written in Armenian script; and words that are hard to hear may be omitted.
About the benchmark
ArmBench-ASR is also published by Metric AI. Its 20 August 2026 post, written by authors who also appear on the Ohan posts, describes version 1.0 as five Armenian datasets (Common Voice 26, FLEURS, poems, movies and news) totaling 10,113 clips and about 20.7 hours of audio, used to compare more than 30 open and closed models. The post says the poems, movies and news sets are not open-sourced. In that post Gemini 2.5 Pro ranked first by strict combined word error rate at 13.80%. The OhanASR post says OhanASR-v0.1 ranks second by word error rate on the benchmark. Lens did not open the interactive leaderboard and does not know whether it has changed since August or which scoring view the Ohan ranking uses.
License and access
Metric AI released all three models under Creative Commons Attribution-NonCommercial 4.0 (CC BY-NC 4.0), which it describes as free for research, educational and other non-commercial use. Hugging Face lists all three repositories as gated with manual access approval, so the model cards were not available to Lens.
What Metric AI says comes next
For OhanTTS, Metric AI names bistreaming, which would accept text in chunks and begin speaking while the input is still arriving, along with more robust pronunciation across languages and longer generation. For OhanASR it names better accuracy, lower streaming latency and more robust transcription in varied conditions. It says the models were trained on the LUMI supercomputer in Finland through the EuroHPC Joint Undertaking.
Metric AI also published Ayb, a family of Armenian language models (Ayb-0.5-0.8B, 4B and 9B), the same week. Hugging Face lists them as image-and-text-to-text models, not speech models, and Lens did not read that announcement.
Sources and notes
Sources: Metric AI, “Introducing OhanTTS: Armenian Text-to-Speech” and “Introducing OhanASR: Open Armenian Speech Recognition Models” (both 20 November 2026); “ArmBench-ASR: A Benchmark for Armenian ASR” (20 August 2026); the Hugging Face pages for OhanTTS-v0.1, OhanASR-v0.1 and OhanASR-Streaming-v0.1; and the Ayb announcement, of which only the listing was seen. Notes: Lens read the announcement posts as extracted text, could not open the gated model cards, and did not run any of the models or check the benchmark figures independently. The accuracy, latency and ranking figures are Metric AI’s own, measured on a benchmark Metric AI also publishes. Lens did not contact Metric AI.
Image: the statue of David of Sassoun, the hero of the Armenian epic Daredevils of Sassoun, by Yervand Kochar (1959), outside Yerevan railway station, 25 May 2019. Photo: Armenak Margarian (CC BY-SA 4.0), via Wikimedia Commons; cropped. Stand-in image: not a photo of Metric AI or its models, and OhanASR is named after a different character in the epic, Dzenov Ohan.

