DEV Community

Natsuki
Natsuki

Posted on • Originally published at blog.natsuki-cloud.dev on

Automated Budget Local LLM Home Assistant Voice Assistant (No Cloud)

#homeassistant #selfhosting #voiceassistant #localai

Motivation

If you want Google Assistant voice control in Home Assistant, you usually need a Nabu Casa subscription. But I wanted three things:

  1. Save money — no monthly subscription.
  2. Keep my privacy — no smart-home data leaving my LAN.
  3. Utilize my unused gaming PC — put my Intel Arc A580 GPU to work.

So I built my own fully local voice assistant — offline, GPU-accelerated, and integrated with Home Assistant.

Inspiration

This project was inspired by NetworkChuck’s video. He showed how to run a Pi satellite. I extended the idea into a automated setup powered by Docker Desktop and Task Scheduler.

My Core Idea

I wanted a system that would start automatically, run fast on my Intel Arc GPU, and stay simple. My setup is built on these pillars:

  • Docker Desktop + Docker Compose → manages Whisper (STT) + Piper (TTS). Auto-start ensures they’re always online.
  • Task Scheduler (Windows 11) → launches Ollama Portable Zip at login.
  • IPEX-LLM → gives GPU acceleration on my Intel Arc A580.
  • Qwen 3-4B Instruct → the local LLM, tuned with a system prompt to only return actionable device commands.
  • Home Assistant Voice Assistants → everything wired up under one assistant with Wake Word, STT, TTS, and LLM.

This combination — Docker Desktop + Compose + Task Scheduler + IPEX Ollama + Qwen 3-4B — is what made it work.

The Result

It actually works better than I expected:

  • Say a command → Pi mic (ReSpeaker HAT) picks it up.
  • Whisper transcribes speech.
  • Qwen interprets it into an intent.
  • Piper responds with voice.
  • Home Assistant executes the action.

All fully offline. No cloud, no Nabu Casa, no Google/Amazon.

Summary

  • Docker Desktop auto-start + Compose keeps STT/TTS reliable.
  • Task Scheduler ensures Ollama is always alive.
  • IPEX-LLM makes Intel Arc GPUs practical for real-time AI.
  • For NVIDIA/Apple users, it’s simpler: just use the Ollama Desktop app.

What’s Next

  • GPU-accelerated Whisper and Piper.
  • Larger models with IPEX-LLM.
  • Smarter intent handling.
  • Multilingual support (English + Japanese).

👉 Want the full setup with configs, batch scripts, and step-by-step Home Assistant integration?

Read it here: Full Blog Post

Top comments (0)