The question people ask about smart speakers is whether they are always listening. The technically accurate answer is yes, and the useful answer is more complicated, because listening and recording are not the same thing.
How wake word detection works
The microphones are always on. They have to be, or the device could not hear you say the wake word.
What happens to that audio is the important part. It goes into a small buffer of a few seconds, held in memory on the device itself, and is continuously overwritten. A dedicated low-power chip compares the sound against the acoustic pattern of the wake word. Nothing leaves the device.
When the pattern matches, the device wakes, lights up, and begins streaming audio to the company’s servers, where the actual speech recognition happens. That streamed audio is what gets stored against your account.
So the honest summary: constantly listening locally, recording and transmitting only after a wake word.
False wakes are the real issue
Wake word detection is probabilistic, and it is tuned to err towards waking. A device that misses your command is more annoying than one that occasionally wakes by mistake, so manufacturers accept false positives.
Research has repeatedly found smart speakers activating on ordinary television dialogue and conversation, sometimes many times a day. Words that rhyme with or partly match the wake word are enough.
When that happens, several seconds of whatever you were saying is uploaded and stored. Not because anyone targeted you, but because the pattern matcher was wrong. This, rather than deliberate surveillance, is how private conversation actually ends up on a company’s servers.
Humans have listened to recordings
Every major voice assistant company was found, around 2019, to be employing contractors to transcribe samples of user recordings in order to improve accuracy. Several were reported to include accidental activations capturing private conversations.
The practice was not hidden so much as buried in terms nobody reads. After the reporting, the companies made human review opt-in or clearly disclosed, and added controls.
Check whether you are opted in. It is usually a setting named something like “help improve” in the assistant’s privacy menu, and it is worth switching off unless you have a reason to leave it on.
What to actually change
- Turn off human review. The single highest-value setting.
- Set automatic deletion. All the major assistants can delete recordings after three or eighteen months. Choose the shortest option.
- Delete the existing history once, since the retention setting is not retroactive on every platform.
- Use the mute button. On reputable devices this is a hardware disconnect for the microphones, not a software flag, and the light stays on to indicate it.
- Turn off voice purchasing if you have children, or anyone who finds it funny.
- Review linked accounts. Calendars, email and messaging connected to a speaker are readable aloud by anyone in the room.
Listen to what it has
Every platform lets you play back your own recordings, and doing so once is more persuasive than any article.
You will find the commands you expected, and you will usually find a handful of clips you did not intend at all: fragments of conversation, the television, someone in another room. That is the false-wake rate made audible.
Placement is a privacy setting
A speaker in a kitchen or living room hears timers and music requests. A speaker in a bedroom or a home office hears things you would not choose to upload.
Smart displays add a camera to this, and deserve a physical shutter and more thought about which room they go in.
A proportionate view
Smart speakers are not secretly transcribing your life and selling it. The systems work broadly as described, and the companies have real reasons to keep it that way.
What is true is that an imperfect trigger, running constantly in your home, will occasionally capture things you did not mean to say, and those fragments are stored under your name until you tell it otherwise. Ten minutes in the settings reduces that to close to nothing, and the device works exactly as well afterwards.
Join the discussion