I built a self-healing Voice Assistant today that is using the non-AI pipeline.
It’s not self-healing in realtime though, but the goal was to figure out why some commands work and others don’t, then have AI write updated automations for 'em.
So I have an agent scheduled to go through my failed voice chats, figure out why they failed, and help me write or rewrite intent automations that help add guardrails specific to my house, people names, etc to help make my voice control always work the way I want it to.
I’m not writing any of this. I know that something didn’t work, but an agent is going to fix that for me. If it can’t figure it out, it can send me a push notification to my phone through Home Assistant MCP.
I swear some folks already did this like this guy:
The difference with mine is I’m not using AI to process the requests in realtime, that way, I retain my fast responses. Over time, Voice Assistant (through Intent Automations) learns more and more about my household and the ways I ask for things, and suddenly, it works exactly the way I want it without failed messages.
That’s despite the fact that Whisper converts a lot of the things I say incorrectly (because it will always do that). I switched to NeMo Parakeet and still had issues, so it’s not just Whisper.
a single-letter typo is all it too to destroy the message