OmniVoice: An Open-Source TTS Model for Voice Cloning and Voice Design
I've been looking for a TTS tool that offers good voice quality, supports voice cloning, and can be used without relying on expensive proprietary services.
That's when I came across OmniVoice, an open-source Text-to-Speech project from the k2-fsa team.
What caught my attention is that OmniVoice supports both Voice Cloning and Voice Design, while supporting more than 600 languages according to the project's documentation.
Voice Cloning
With Voice Cloning, you can provide a short reference audio and generate new speech using the characteristics of that voice.
For example, the sample audio attached to my original post was generated using OmniVoice with the voice of the late Iranian voice actor Behrouz Razavi.
Voice cloning should of course be used responsibly and with respect for people's rights, consent, and applicable laws.
Voice Design
Another interesting feature is Voice Design.
Instead of providing a reference recording, you can describe the characteristics of the voice you want and let the model generate a suitable voice.
For example, you could describe a voice as:
female, low pitch, british accent
and use that description to generate speech with a corresponding voice.
One important caveat is that Voice Design is not equally optimized for every language, so the results can vary depending on the language and voice description.
Open Source and Self-Hosted
One of the main reasons I found OmniVoice interesting is that it is open source and can be run locally rather than requiring a paid cloud API.
The project also provides instructions for running it on Apple Silicon using MPS.
You can find the source code here:
https://github.com/k2-fsa/OmniVoice
My OmniVoice API
To make OmniVoice easier to use in other projects, I also built and deployed an API based on the OmniVoice project.
You can try it here:
https://huggingface.co/spaces/abmohajeri/omnivoice-api
The idea is simple: instead of setting up the model, dependencies, and inference environment yourself, you can interact with OmniVoice through an API.
Performance
OmniVoice also focuses on fast inference. The project reports very low real-time factors under its benchmark conditions.
Of course, actual performance depends on the hardware, inference settings, input length, and other factors, so benchmark numbers shouldn't be directly generalized to every environment.
Final Thoughts
If you're looking for an open-source TTS solution with Voice Cloning and Voice Design, OmniVoice is definitely worth checking out.
I'll also be publishing a more detailed guide covering how I set up OmniVoice, how Voice Cloning works, and how I built the API around it.
Links
OmniVoice API: https://huggingface.co/spaces/abmohajeri/omnivoice-api
