The model, with ~31M parameters (ONNX), is tuned for latency and local inference, and comes already exported. I was trying to push the limits of what I could do with small, fast models. Runs 5.6x realtime on a server CPU
It supports voice cloning, voice blending (mix two or more speakers to make a new voice), the license is Apache 2.0 and it uses DeepPhonemizer (MIT) for the phonemization, so no license issues.
The repo contains the checkpoint, how to run it, and links to Colab and HuggingFace demos.
Now, because it's tiny, audio quality isn't the best, and as it was trained on LibriTTS-R + VCTK (both fully open datasets), speaker similarity isn't as good.
Regardless, I hope it's useful.