OmniVoice is a massively multilingual, zero-shot text-to-speech model from the k2-fsa team that can clone voices and speak more than 600 languages. It has more than 15,300 stars on GitHub and was actively updated in 2026.
Built on a novel diffusion language model-style architecture, OmniVoice generates high-quality speech at fast inference speeds without language-specific training. Give it a short reference recording and it clones that voice in a new language; or use voice design to create a voice from a description instead. The model is published on Hugging Face with an interactive demo space, and the repository includes a Python API, command-line tools and the full training and evaluation code for teams that want to fine-tune it.
Beyond raw coverage, OmniVoice offers unusually fine control over how speech sounds. Voice design lets you steer a generated voice with attributes such as gender, age, pitch and dialect or accent โ including effects like whispering โ while non-verbal symbols such as [laughter] can be embedded directly in the text, and pronunciation can be corrected with pinyin or phoneme hints. It is also fast: the project reports a real-time factor as low as 0.025, roughly forty times faster than real time, which makes it practical for interactive applications rather than just offline rendering. The model comes from the k2-fsa community, known for the Next-gen Kaldi speech toolkit.
OmniVoice is free and open source under the Apache-2.0 license, including the model weights on Hugging Face. You only need your own GPU or compute to run it.
Voice application developers, accessibility tool builders, multilingual content creators and speech researchers who need one TTS model for hundreds of languages.