k2-fsa releases OmniVoice, a 0.6B zero-shot TTS covering 600+ languages
OmniVoice is a 0.6B voice-cloning and voice-design TTS tagged for 646 languages; code Apache 2.0, weights CC-BY-NC.
k2-fsa released OmniVoice, a 0.6B-parameter text-to-speech model with Qwen3-0.6B-Base as its base. The card says it covers over 600 languages and carries a 646-language tag. It offers zero-shot voice cloning from short reference audio, voice design from attributes such as gender, age, pitch, accent and whisper, non-verbal symbols such as laughter tags, and pronunciation correction through pinyin or phonemes. Output is 24 kHz audio.
The card claims a real-time factor as low as 0.025, which it describes as about 40 times faster than real time. It gives no quality benchmarks and refers to the paper for technical details. Cloning needs a transcript of the reference audio.
Licensing is split. The code is Apache 2.0; the card says the pre-trained model is "licensed under the CC-BY-NC" because of constraints from training data such as Emilia, so the weights are non-commercial. The disclaimer prohibits unauthorised voice cloning and impersonation.
The card gives no release date. The date used is the Hugging Face repo creation date, 30 March 2026; the linked paper is listed as published 1 April.
Six hundred languages is a headline. The card offers no per-language quality figures behind it.
Sources
Editor's first pass (AI-assisted)
- Real 4: Weights are public and an API example is on the card; non-commercial weights limit use.
- Useful 3.5: Zero-shot cloning plus voice design is useful, but quality per language is not documented on the card.
- Notable 4: Over 1,500 likes on Hugging Face and an unusually wide language count.
- Novel 3.5: Language breadth at 0.6B is the unusual part.
Comments