Applied AI Research Portfolio Blog Get In Touch

Open speech and language models for Armenian

State-of-the-art open models that hear, speak and understand Armenian — the building blocks of sovereign Armenian AI.

OhanASR v0.1

OhanASR v0.1

SOTA open Armenian speech-to-text: #1 by character error rate on ArmBench-ASR among all open and closed models, plus a streaming variant for live transcription.

3.49% CERNon-streamingStreaming
Read
OhanTTS v0.1

OhanTTS v0.1

Streaming Armenian text-to-speech with voice cloning from a few seconds of audio — first sound in ~330 ms, eight languages.

StreamingVoice cloning8 languages
Read
Ayb v0.5

Ayb v0.5

SOTA open Armenian LLM that rivals frontier models on Armenian tasks, trained end to end with an Armenian tokenizer and early vision.

+ Vision0.8B4B9B
Read