Come check it out yourself
Training code: https://github.com/franciscocarloserra/ttslibre
Weights & Training sets: https://huggingface.co/FranciscoCarlosErra/ttslibre-poc-en
It's still glitchy and robotic but I'm mind blown about the PoC
Meet a tiny robot with a big attitude. Hold a button, tell her "build me a tower", and she does: blind, never seen a thing, absolutely sure she just made a masterpiece. Wrong height, wrong place, zero doubt, plenty of sass.
It lives entirely on your own machine: local model, local ears, local voice, no cloud. Come argue with her.
Built this for myself. It listens to whatever's playing on your laptop, shows characters, pinyin and word meanings as people talk, and I can type a question and get an answer about what I just heard. All local, no cloud.
Not out yet. Curious if anyone else would actually use this.
Let's build a tiny local model, at least as performant both in quality + speed as Supertonic and Kokoro. Let's allow people to actually control it and run it locally.
Let's make it small and performant. Let's publish the training code, the training data and the weights. MIT code, CC0 data and weights. No strings.
Looking for people to help design it, scope it, gather the data, contribute compute or sponsor the training runs. Let's make this happen.
https://github.com/franciscocarloserra/ttslibre
Anyone interested?
Since the Supertonic voice builder has been taken down, and it's surprisingly useful for local sub-second tts renders. I've created a local voice lab and downloadable voicepacks. Feel free to try them out.