Have you considered integrating RHVoice in addition to the Piper models? It's nowhere near as exciting (it uses HMMs instead of cool new shiny neural network go

Have you considered integrating RHVoice in addition to the Piper models? It's nowhere near as exciting (it uses HMMs instead of cool new shiny neural network goodness), but the use-cases are sort of similar (it was originally developed for screen-readers, which require extremely low latencies and high responsiveness.) It doesn't support as many languages, but for some of the languages it does support (Polish for example), I find the quality superior to what Piper can offer.