After Google Assistant's reliability plummeted on Nest Minis, one user switched to a fully local setup using HomeAssistant and llama.cpp. The core problem was Google's assistant becoming progressively 'dumber' and unreliable for basic tasks, coupled with privacy concerns from always-listening microphones. The solution involved replacing Google Home with HomeAssistant's local voice assistant powered by llama.cpp, running on a UnRaid NAS.

This setup avoids cloud dependencies entirely. Hardware choices included a Beelink MiniPC with USB4, an eGPU enclosure, and various GPUs like the RTX 3090 and RX 7900 XTX, tested for response times. Software used llama.cpp for the LLM, Rhasspy for speech-to-text via OpenVINO, and Kokoro TTS for text-to-speech. The critical factor was the LLM prompt, which the user refined through iterative ChatGPT feedback to achieve reliable tool calling and complex command handling. This journey highlights a viable path for privacy-focused, locally hosted voice control.