Made my very own voice interface where I don't even need to look at a screen. Use a combination of my own bespoke g2p with kokoro, parakeet-ulta for ASR, hooks into claude code where each project gets it's own session. Had to make an Android app so the bluetooth headset multi-function button could trigger my "turn". I hate pre-emptive voice agents and getting the watered down voice agent stuff. So I have an intentional turn concept. I'd rather be patient and wait rather than get the dumb version of all the things I use.

I get full claude code, can use the official app for text interface, get to rant and ramble without a screen, and kokoro delivers a great consistency. I even added some tools for the agent to patch G2P lexicon when I hear something that sounded weird. Doing that was a bit pedantic but that is half the fun. So it is bespoke all the way down. Been loving it.