For a decade, voice control in the home meant sending a recording of your kitchen to a large company and hoping. The technical justification was real — the models were too big to run locally.
That constraint has gone. Speech recognition, language understanding and speech synthesis all now run at good quality on hardware that fits in a rack. This changes the calculation in ways worth being specific about.
What changes when it runs locally
It works when the internet does not. A cloud assistant during an outage is an expensive ornament, and outages are exactly when people want to turn lights on.
It is faster. Round-tripping to a data centre adds latency you can feel. Local inference responds quickly enough that the interaction stops feeling like a query and starts feeling like a control.
Nothing leaves the house. Not the audio, not the transcript, not the pattern of when you are home and what you asked for. For clients who declined voice control entirely on privacy grounds — and there are many — this is the change that makes it viable.
No subscription, no sunset. Cloud assistants have features removed, get rebranded, and occasionally get discontinued. A local stack keeps working.
What a local stack looks like
Four pieces, all of which have good open implementations:
- Wake word — small model running constantly on each satellite, cheap enough for a microcontroller
- Speech to text — the heaviest piece, running centrally on the rack machine
- Intent handling — either direct matching for common commands, or a small language model for anything phrased loosely
- Text to speech — synthesis, back to the room the request came from
The pattern that works best in practice is tiered: match the common commands directly, because "kitchen lights off" should never wait on a language model, and fall back to the model only when the direct match fails. Most requests in a house are the same twenty phrases, and they should be instant.
Where the difficulty actually is
Microphones, not models. Recognition accuracy in a real room is dominated by the hardware and the acoustics. A far-field array in a quiet study is close to perfect. A single microphone in a kitchen with an extractor running and music playing is not, and no model fixes that. Budget for decent satellites and place them thoughtfully.
Barge-in. Interrupting the assistant mid-sentence requires echo cancellation good enough to hear you over its own output. This is one of the clearer remaining gaps against commercial products.
Naming things. Voice fails most often because the family calls a room something different from the system. "The snug", "the back room" and "the TV room" may be one place. Get the names from the family during commissioning, not from the floor plan.
Where voice is still the wrong answer
Voice is genuinely good for hands-full moments — cooking, carrying something, in bed — and for anything that would otherwise mean finding an app.
It is genuinely worse than a keypad for everything else. A keypad is instant, silent, unambiguous, works for a guest who does not know the command vocabulary, and does not require anyone to talk to the ceiling in front of visitors. If a room's most common action needs a sentence, that is a missing button, not a voice problem.
We specify voice as an addition to a properly designed set of physical controls. Never as a replacement for them. The houses where voice control is genuinely loved are the ones where nobody has to use it.