Always-listening is not always-recording

What it is

A kitchen speaker in a shared apartment keeps a small light on all night. A roommate asks whether it is recording everything. Packaging that says “always on” welds two different states together. Always-listening is on-device keyword spotting over a short buffer that is overwritten unless a wake fires. Always-recording writes audio as a retrievable object, to disk or to the cloud. A dialogue system has to be able to say the two states apart. Mixing them lets “the mic stays up so it can wake” be heard as “every sentence in this room is filed.”

Why it happens

Far-field wake depends on a detector that keeps running. The usual engineering is a ring buffer: a few hundred milliseconds to a couple of seconds cycling on-chip; only a wake sends the following stretch to recognition. That “always hearing” is not “always keeping” — samples that miss the keyword are supposed to be dropped, with no filename and no session id. Always-recording is the opposite: a write, a retention, a chance of being pulled into a training queue. If users, support scripts, and privacy copy share the verb “listening,” the two data flows cannot be interrogated. The cut is object lifetime: listening-state audio has no identity; recording-state audio has one, which is why inspect-and-delete even applies. A lamp that stays on says an acoustic path exists. It does not say which lifetime that path is using.

Studying it

A privacy nutrition-label comprehension test: after a piece of product copy or a label, ask whether a sentence spoken without the wake word becomes a playable record. Score whether listen and record were separated, not whether people “feel safe.” Add a card sort: “ring buffer overwritten,” “cloud recognition after wake,” “opt-in improvement upload” as three piles — see whether they are still glued together after the copy. Walk the device: does the unwoken period produce timestamped audio objects? If yes, it is recording; if not, it is listening. If the box says “we don’t record” and the activity history has clips, the right answers in the quiz are the device’s behavior.

Where it stops holding

Some products upload seconds around the wake “for improvement,” and then the product itself has torn the listen/record border; “it is only keyword spotting” is no longer a defense. Strictly on-chip spotting that never writes a file can have no recording state, and still has to name the listening state. Conference mics and evidence recorders are always-recording by job description; the distinction then becomes whether anyone is treating them as an assistant. Legal “processing” is wider than listen/record; a label that only uses those two verbs may still miss a one-shot inference. The distinction sorts data-flow kinds. It does not settle whether someone in the room feels at ease.

Applying it

• On the settings home, two lines: “while waiting for a wake word, audio is overwritten in a local short buffer and does not create history items,” and “after a wake, whether this request is kept, and where.” One “always on” toggle must not stand for both states.
• Activity history should not contain playable clips from unwoken periods. If it does, label them as recording, not listening.
• Support and the unboxing card need a contrast sentence that answers “does it record all the time,” naming buffer overwrite rather than “we care about privacy.”
• How to check: people who have not read internal docs finish the label and answer three yes/no items (does unwoken speech create a file, is a wake saved by default, is the improvement program a separate consent). If copy leads any item wrong, rewrite. Then list files / capture traffic for an unwoken hour and confirm no new audio objects.

Related

• Same group: M4.02.2 A capture indicator must be visible and unspoofable · M4.02.3 Physical disconnect is the mute people can believe
• Nearby: M4.05 Retention of voice data · M4.07 Felt privacy of always-on microphones · C7.07 Privacy visibility of voice input
• Search terms: always-listening vs always-recording · wake-word buffer · retention