Very cool. By accumulating lots of these tweaks, I feel like you're also going to have an opportunity to backdoor your way into a great text-to-speech API product as well if you have any interest in going that direction. It seems like the main challenge there is ironing out all the edge cases and you've created an excellent feedback loop for accomplishing that.
We have a small map of tweaks, and our users keep feeding us with more. The model performs great on its own most of the time though.