KENZY.Field notes from building a local-first voice assistant.
Kenzy is an open-source, whole-home voice assistant you own and run yourself. Building it keeps turning up questions the open literature doesn't answer — so we measure them, and publish what we find. Every number here is measured on real hardware, and every experiment is written to be re-runnable by anyone on their own journey.
Can local AI answer as fast as the cloud — on hardware you can actually buy?
Can consumer and lower-end professional GPUs run a private voice assistant with real-time responsiveness on par with Alexa, Google Home, and Siri — but with nothing leaving the house? We measured the whole voice turn — hearing, thinking, speaking — across two model architectures (dense and Mixture-of-Experts), three publicly-buyable GPUs, and a full-duplex speech-to-speech model. The answer: yes — and a 3B-active MoE clears the bar with room to spare, at ~260 ms.
Read the study →Who answers? Choosing the right speaker when two devices hear the same wake word.
Put a voice node in every room and, sooner or later, two of them hear you at once — both answer, both spend a full pipeline. This is the reproducible experiment record behind Kenzy's wake-word arbitration: what commodity microphones actually do to your signal, why loudness is a device property as much as a distance one, and the honest limit — automatic gain control flattens distance so thoroughly that past ~3 ft the "nearest device" is decided by aim, not proximity. Runs on two ~$35 speakerphones and a pair of Raspberry Pis.
Read the study →