A detailed account of migrating from Google Home/Nest Minis to a fully local Home Assistant voice assistant powered by llama.cpp and local LLMs. Covers hardware selection (eGPU setup with Beelink MiniPC), GPU and model comparisons, STT/TTS stack (Faster Whisper, Piper), LLM prompt engineering to fix unwanted behaviors, custom wake word training with microWakeWord, and automation-based music playback via Music Assistant. Key lessons include using higher-quantization GGUF models from HuggingFace, the critical role of prompt design, and solving edge cases like emoji in TTS output and false activations.
Sort: