The short answer
Yes. We have been deploying private voice control for clients who specifically did not want an Amazon, Google, or Apple account associated with their home. The technology is mature, the components are off the shelf, and the experience is on par with consumer voice assistants for the things people actually use voice control to do.
The privacy problem with Echo, Nest, and HomePod
The three big platforms have three different privacy postures. None of them are local-only. Amazon Echo runs everything past the wake word through AWS and ties it to your Amazon account. Google Nest does the same with Google. Apple HomePod with Apple Intelligence is much better, with most processing on-device, but it is still tied to an Apple ID and the convenience features that matter most often require the iCloud account.
For a luxury home -- where the people in the house include staff, family members, guests, and contractors -- the question of whose voice goes into whose account graph is non-trivial. The cleanest answer is "nobody’s." That is a local stack.
The Whisper, Voxtral, Parakeet stack
- Whisper Large v3 (OpenAI, open-weight): the gold standard for local speech-to-text. Accurate, multilingual, runs on a Mac Mini or any modern GPU.
- Voxtral (Mistral): faster on CPU, slightly less accurate, good for low-latency wake-and-respond loops.
- Parakeet (NVIDIA): best in noisy environments. We use it in installs where the microphone is in a kitchen with running water and a cooktop.
Picking one is a hardware question, not a feature question. Our default for a Mac Studio install is Whisper. For a Linux box, we usually go Parakeet.
The Home Assistant Assist pipeline
Home Assistant Assist is the orchestrator. The pipeline:
- Wake word (openWakeWord, fully local) listens for "Restrepo" or whatever phrase the client picks.
- Speech-to-text (Whisper) transcribes the request.
- Intent processing (a small Llama or Phi model, local) parses the transcribed text into a structured action.
- Dispatch sends the structured action to the right system: Crestron processor, Lutron hub, Sonos, Frigate, calendar.
- Confirmation (text-to-speech, local) speaks back the result.
The whole loop completes in under a second on a properly specced Mac Mini. No cloud. No account. No advertising.
Integration with Crestron, Lutron, and Sonos
For Crestron Home, we expose Home Assistant as a documented control source so the Crestron processor sees voice commands as documented API calls. For Lutron, Home Assistant talks to the RA3 or HomeWorks hub via the standard local API. For Sonos, the local stack drives Sonos through its native API; the user does not need a Sonos voice subscription.
The client interaction model: "Restrepo, dim the dining room to thirty percent" gets transcribed, parsed into a dim-to-30 action targeted at the dining room load, dispatched through the Crestron processor (or Lutron hub), and confirmed verbally. Total elapsed time well under a second. No part of the audio left the house. No part of the action depended on a remote service.
"Restrepo, dim the dining room to thirty percent" -- the full path
For the technically curious: the wake word fires on local audio. Whisper transcribes the rest. A small Llama 3.2 3B model running on the same box parses "dim the dining room to thirty percent" into a structured intent: action=dim, area=dining room, value=30. The intent dispatcher matches "dining room" to a known room and "dim" to a known action, then calls the Home Assistant service for the dining room dimmer. Home Assistant routes to the Crestron driver, which sends the documented Crestron API call. The light dims. A text-to-speech model (Piper, local) confirms "dimmed dining room to thirty percent." Done.
Cost and hardware reality
The hardware footprint for this stack starts at about $1,500 for a Mac Mini M4 Pro 24 GB and goes up to $5,000 for a Mac Studio if you want headroom for additional AI workloads. The microphones are PoE ceiling mics, typically Shure or Audio-Technica, $200 to $500 per zone. A six-zone deployment in a 6,000-square-foot home runs $4,000 to $8,000 in hardware plus $5,000 to $12,000 in install labor. The result: voice control of every room with no cloud account.
FIELD NOTE
The local voice experience is genuinely competitive with cloud assistants for home control. It is not competitive with cloud assistants for arbitrary general knowledge. We pair the local stack with optional cloud reach-out for clients who want both, and we let them choose which queries cross the line.