Silence can be thinking or leaving

What it is

The system has just asked how much salt. For a few seconds the kitchen produces no speech. The person may be staring at a recipe, doing the arithmetic, or they may already have walked the plate into the dining room. The microphone records one thing: no voice. Ambiguous silence collapses “still in the dialogue, composing” and “already out of the dialogue” into a single observation. This is not endpointing’s question of whether an utterance finished, and it is not how many seconds a timer should wait. It asks whether the gap still belongs to the current session.

Why it happens

Once a question yields the floor, the system sees energy under a threshold. Two processes generate that gap. One is cognitive occupancy: reading a label, mental arithmetic, searching a contact list — the mouth has not opened, the body still faces the device, the person is still the addressee. The other is departure: another room, a summons at the door, the conversation already closed in the head. Acoustically the paths do not split, because what is missing is speech, not two kinds of noise.

Treat every gap as “still here, just slow,” and the device keeps talking to an empty room. Treat every gap as “gone,” and it cuts off someone still assembling a sentence. The error is not a timer that is slightly fast or slow; it is one label covering two states. Cues other than speech can split them: footsteps still in the room, a face still at the screen, a filled pause already produced. Without those cues, the first gap is undecided, not a verdict.

Studying it

Instrument a room with video. In half the trials, after the question, people keep reading material (a recipe, a list). In the other half they are told to leave. Give annotators only the microphone track and score how often they label the gap “thinking” versus “left” — that is the system’s information ceiling. Take presence coded from video (still in the device’s facing space) as ground truth; report false departure (cutting off someone still thinking) and false presence (talking to an empty room) separately.

Field logs give weak labels: a quick “I’m here” or a complete answer after the timeout looks like thinking; no speech until teardown, with the next wake in another room, looks like leaving. Do not substitute a lab instruction to “think for three seconds” for departure — that only copies cognitive occupancy.

Where it stops holding

Hands occupied, eyes on a manual: a long gap is still presence; the primary task borrowed the channel. A phone face-down during a call can mean the mouth left the mic while the person stayed — presence is not “aimed at the microphone.” In a multi-party room, the addressee’s silence may be someone else taking the floor, not a walk-away. For people with motor or articulatory slowness, inferring departure from a few seconds of no-speech systematically harms them. Devices that see a face have evidence acoustics lack; moving those conclusions onto a speaker-only product overstates separability.

Applying it

• Do not encode the first no-speech event as “user left.” Treat it as undecided: a very short check (“still there?”) or a listening light, not an immediate teardown or an action.
• Use presence sensors when they exist: ongoing noise, occupancy, a face at the screen can raise “left” on its own axis. Without those channels, do not pretend the distinction is available.
• Log “no speech after a question” separately from “no speech after the system finished a statement.” The first is waiting for an answer; the second is a session already loosening.
• In walkthroughs, lay first-person video beside the mic track. If the person is still in a drawer and the system already says it will stop asking, thinking was labeled as leaving.

Related

• Same group: M1.07.2 A re-prompt must change its wording · M1.07.3 After repeated no-response, end rather than keep waiting
• Nearby: M1.11 Silence, Timeouts and Re-prompts · M1.08 Turns and Floor Management · C7.02 Endpoint Detection
• Search terms: ambiguous silence · thinking versus departure · presence after a prompt