Working, this is a speech-to-speech voice assistant that lives entirely on one GPU and talks like a person rather than a command matcher. It copes with interruptions mid-sentence and follow-up questions, needs no cloud AI or subscription, and can be used from a phone.
The build rests on the Ming-Audio-tiny model, released just days earlier. It listens continuously for the Hey Ming wake word, ignores chatter while asleep, and hands complex requests to a Hermes agent. Through Home Assistant it works lights, thermostats, locks, vacuum robots and media playback, and it can fire announcements from automations. On the demo rig, a Radeon RX 7900 XTX picked up secondhand for $750 a month earlier, median latency lands near 230 milliseconds from the end of speech to the first audio of the reply.
Details I admire: the mobile client steps through Paused, Listening, then Hearing you, Thinking and finally Speaking, and offers echo cancellation, noise suppression and gain control options. In the demo the assistant switched the bedroom lights off, then on again, answered a spoken weather question for North Carolina out loud, and analyzed raw audio to identify a fart sound. One party mode request went to the Hermes agent and bathed the room in magenta and purple, arriving about 40 seconds later.
The maker is upfront about the trade-off: snappy replies were chased over raw intelligence, and in the maker's own words it is no genius, with the Hermes agent there for heavier thinking. For my money that is the right balance in a spoken interface. Still unresolved is whether it can be built into Home Assistant Voice PE, a question waiting on a GitHub issue.