limpid elbow
Hey everyone! I have been a fan of Home Assistant for many years now but never really put a system together until recently. I was given the project of developing a privacy-preserving voice assistant but decided to give it a twist—adding a constraint that it works fully offline.
My dissertation project at University provided me with the best excuse to use the Home Assistant platform and Radxa's fairly new product, the Fogwise Airbox. It contains a TPU capable of running LLMs locally, and in my testing, it runs on around 40 W of power.
The pipeline utilises Whisper STT which runs on the Fogwise TPU, Piper TTS on the CPU, and the Qwen2.5-3B LLM.
I have created a video of my demonstration, so please let me know if you have any questions!
ancient hull