OmniVoice

OmniVoice is a zero-shot text-to-speech (TTS) model that the k2-fsa team publishes on GitHub, and the README says it supports over 600 languages. Its makers say it is built on a diffusion language model-style architecture, runs inference quickly, and can both clone a voice and design one from a description. For cloning, they suggest 3 to 10 seconds of reference audio. The authors note that voice design was trained mainly on Chinese and English, and they are more cautious about other languages. The standout is the sheer number of languages in one model, with cloning and voice design both built in.

Who it's for:speech developers and researchers, and content makers who dub in many languages and can deploy a model themselves.

Before you start: the code is free under the Apache-2.0 licence; running it locally needs Python, the model files and enough compute. We haven't checked the online demo. Any reference audio you provide goes into the inference process. The README's disclaimer says users are strictly prohibited from using the model for unauthorized voice cloning, voice impersonation, fraud, scams or other illegal or unethical activities, and asks users to follow local laws and ethical norms.

Based on the project's public documentation. Not tested by AI Inspiration World.