Full Local Home Assistant Voice Assistant on RK3576 (NPU-Accelerated STT + LLM + TTS)
Hi everyone,
I’d like to share a fully local Home Assistant voice assistant that I’ve been running on an RK3576 SBC.
The entire voice pipeline runs 100% locally with no cloud services, no API keys, and no internet connection required. All voice processing stays on the device, providing complete privacy while delivering fast response times through RK3576 NPU acceleration.
AI Pipeline
The following models are used:
- Speech-to-Text (STT): Whisper
- Large Language Model (LLM): Qwen2.5-1.5B-Instruct
- Text-to-Speech (TTS): Piper
All three components are accelerated by the RK3576 NPU.
Features
Fully local voice assistant- NPU-accelerated STT, LLM, and TTS
- No internet connection required
- No cloud services or API keys
- Voice data never leaves your device
- Low-latency responses
- Docker Compose deployment
- Home Assistant compatible
Starting the entire stack only requires:
docker compose up -d
Example Voice Commands
- “Turn on the living room lights.”
- “Set the thermostat to 24 degrees.”
- “What’s the weather today?”
- “Lock the front door.”
- “Summarize today’s calendar.”
- “What devices are currently on?”
Why I Built This
I wanted a voice assistant that:
- Works completely offline
- Keeps all voice data private
- Doesn’t rely on cloud AI services
- Continues working even if the internet is unavailable
- Can run efficiently on low-power embedded hardware
The RK3576 provides enough AI performance to run the complete voice pipeline locally while maintaining responsive interactions.
Resources
I’ve written a detailed step-by-step guide covering the hardware setup, Docker Compose deployment, model configuration, and Home Assistant integration:
Hackster: https://www.hackster.io/h1300923175/make-home-assistant-voice-fully-local-with-rk3576-50b4de
The project is open source on GitHub:
I hope this helps anyone interested in building a private, fully local Home Assistant voice assistant on Rockchip hardware. I’d be happy to answer any questions or hear about similar projects running on other embedded AI platforms.