If you run STT and TTS locally, you have to live with some imperfections. E.g. my model very very frequently transcribes my partners name incorrectly, as it is not an English name. But it is fairly consistent.
And on the output side, my LLM gives me often the em-dash in responses, even though I prohibit this in my system prompt - which is super annoying as my TTS piper voice skips over this as if it were no space at all.
I would like to be able to process the text from STT before sending it to LLM and same way Id like to process repsonse from LLM before sending it to TTS. I could then perform simple regexp fixes or even complex custom logic, whatever is necessary.