Offline voice control: no account, no network, no cloud

Some of our lights listen for spoken commands without any connection at all — the recognition runs on a chip a few centimetres from the LED. It is a genuinely different product from asking a smart speaker to turn the light on, and the differences run in both directions.

Two kinds of voice, and the difference is not intelligence

Cloud voice is what most people mean by voice control. You speak to a smart speaker or a phone, the audio goes to a server, the server works out what you meant, and a command comes back down to the light. The vocabulary is open-ended, the light is a device inside somebody else's ecosystem, and the whole chain needs an account and a working internet connection.

Offline voice puts a small speech recognition engine inside the fitting. There is a wake word and a fixed set of commands, both burned in when the product is made. No account, no pairing to an assistant, no audio leaving the room. It is not a cut-down version of the cloud approach — it is a different design with a different set of things it is good at.

What offline is genuinely better at

1. It answers immediately. Recognition happens where the light is, so there is no round trip to a server. The gap between finishing the phrase and the light changing is short enough that it reads as instant, which cloud voice usually does not.
2. It works when the network does not. The router reboots, the broadband drops, the account gets logged out on somebody's phone — none of it reaches the light. A power cut is the one moment you most want a light to respond to a voice, and it is exactly when a cloud assistant is least likely to.
3. Nothing leaves the room. There is no microphone stream going anywhere, because there is nowhere for it to go. For a bedroom, a child's room, or anyone who simply does not want an always-listening cloud device in the house, this is the entire argument and it does not need any other one.

What it gives up

1. The command set is fixed. You cannot invent a phrase. There is a list, the list was decided at manufacture, and using the product means learning it rather than the other way round. Most sets run to a dozen or two commands — on, off, brighter, dimmer, colours, a few scenes.
2. It knows only itself. "Turn everything off" needs something that has a model of what "everything" is. An offline chip in one fitting has no idea the other fittings exist. Whole-house logic is what an assistant or a hub is for, and it is the thing offline voice structurally cannot do.
3. The wake word is not yours to change. It can usually be customised when a product is specified — that is a manufacturing option, and a common one for export models — but not by the person who bought the light.

False triggers matter more than vocabulary

When comparing offline voice modules, the specification people read is the command count and the specification that decides whether they like the product is the false trigger rate. A module that wakes up during the television news is worse to live with than one that understands ten phrases instead of twenty.

Two figures describe this and a serious supplier will have both: the false acceptance rate — how often it responds when nobody addressed it — and the false rejection rate, how often it ignores a command that was properly given. They trade against each other. Tuning a module to be harder to trigger accidentally also makes it harder to trigger deliberately, and the useful question is what those numbers look like with background noise present, because the quiet-room figures are always good.

Where the module sits matters as much as the chip in it. A microphone behind a diffuser hears a muffled room; one facing the ceiling hears the ceiling. On a fitting mounted high in a room with hard surfaces, the reflections arrive late enough to smear the phrase. None of this is visible on a datasheet, which is why a fitting that recognises well on a desk can disappoint once it is installed — and why it is worth testing one at the height it will actually live at.

Languages are models, not settings

Offline recognition is per-language in a way cloud voice is not. Each language is a separate model compiled into the chip, so a product that understands English and Mandarin is carrying two, and adding a third is a manufacturing decision rather than a menu item. This is the constraint behind an oddity buyers notice: the same fitting sold in two markets may have different command sets, because the sets were built separately.

Model size is the other axis. A larger model recognises more robustly across accents and noise and costs more silicon; a smaller one is cheaper and less forgiving. For a product sold into one market with one accent, small is often the right engineering choice. For an export product that will meet accents nobody tested against, it usually is not — and a supplier who offers both sizes and can say which is in the part you are quoting is telling you something useful about how carefully the rest of it was specified.

Choosing between them

1. If the light needs to work during an outage, in a room with no reliable network, or with no cloud device present at all — offline, and the decision is already made.
2. If the value is in logic across several devices, schedules, or integration with other brands — cloud or a hub. Offline voice cannot do it and no amount of vocabulary will change that.
3. If you want the light to respond instantly to three or four everyday phrases and everything else can be slower — both, if the fitting supports it. This is the arrangement most people are happiest with.
4. Whichever you choose, test it in the room, at the mounting height, with the television on. Every voice product on the market performs well in a quiet showroom.

Which of our fittings carry offline voice, and which app pairs with the ones that do, is set out in which app controls your light.