Hobby projects · Household system · AI

Voice Homebot

A private household voice assistant with a central Hermes Agent service and a Swift edge application connected through NetBird.

A locally trained wake-word model listens for the phrase, a local recognizer transcribes the request, and an agent service answers it.

Architecture

An M1 Mac mini runs the central Hermes service, dashboard, and authenticated edge gateway.
The Swift PhoenixEdge application handles wake-word detection, speech recognition, and speech output on the edge.
The wake word is a custom phrase. Its detection model was trained for this household from its own recordings.
Detection and transcription stay on the local machines. The room is not transcribed continuously.
The wake-word model files are checked against recorded hashes when the listener image is built.

The policy installer configures MCP servers, approval for memory writes, tool-loop limits, and voice, dashboard, and Telegram access.
OAuth state is seeded once and preserved so refresh-token rotation survives later deployments.
The central container image is pinned by digest.