Latest / Big Iron Radio / Homelab Brief - A Local LLM Voice Assistant in Home Assistant
Transcript
- 0:03Welcome to the HomeLab Brief from Big Iron.
- 0:06Today, wiring a local LLM into Home Assistant
- 0:10as a conversation agent the right way.
- 0:13Quick note, this is an AI voice
- 0:15drawn from a Big Iron guide.
- 0:17Links below.
- 0:19The first counterintuitive move.
- 0:22Don't send everything to the LLM.
- 0:24Home Assistant's built-in assist intent matcher
- 0:27already handles most commands faster
- 0:30and more reliably than any model.
- 0:32Route the LLM, only the fuzzy remainder
- 0:35that fails deterministic matching, not every utterance.
- 0:39The model is your fallback, not your front door.
- 0:42On model size, an 8 billion class model
- 0:46with real function calling training
- 0:48is the practical floor for reliable tool calling.
- 0:51Think LLAMA3, 1-8B Instruct or QN2, 5-7B or 14B.
- 0:57Go smaller and the model narrates instead of acting,
- 1:01it'll tell you it turned on the lights
- 1:02without ever calling the tool.
- 1:05Expose entities deliberately scoped
- 1:07to exactly what you'll voice control.
- 1:10Never expose all.
- 1:11A tight entity list shrinks the context, cuts latency,
- 1:15and closes off a hallucination and prompt injection surface
- 1:18all at once.
- 1:19It's the single highest leverage setting.
- 1:22And keep it safe.
- 1:24Leave locks and security-relevant covers
- 1:26off the LLM's exposed entities entirely.
- 1:30A voice model should not be one hallucinated tool call away
- 1:34from unlocking your front door.
- 1:36One workflow tip.
- 1:38Test from the chat sidebar before you ever touch voice.
- 1:42That isolates the LLM's tool calling accuracy
- 1:45from speech-to-text accuracy,
- 1:47so when something misbehaves you know which have to blame.
- 1:50The full setup's in the show notes.
- 1:53That's your brief.