C4.02.4User-initiated engagement and disengagementdesignresearch

Both engagement and disengagement need conditions the user can execute on purpose

Aliases: explicit clutch on and off · listening control · deliberate enter and leave

What it is

The counter to Midas touch is not “the system guesses whether it should listen now.” It is enter and leave that the user can perform on purpose. Enter turns the candidate stream into a command stream. Leave turns it back off, and later motion is life again. Enter without leave—timeout or stillness doing the rest—keeps listening while the person turns to talk. Leave without enter—“it looked like a gesture”—drags everyday motion into engagement. Both edges must be user-initiated or the clutch is incomplete. Which movement is the entry signal, and how long a timeout may last, are later design layers. The structure comes first: neither edge may be inference-only.

Why it happens

A session needs two reversible edges, analogous to key-down and key-up. Mid-air has no physical edge, so repeatable user-executed conditions have to simulate one. If enter fails, the system keeps listening and Midas touch is intact. If leave fails, the person thinks they have finished while the system still treats arm-swing as next-page. Inferred leave (stillness, walking out of frame, voice dropping) cuts people who paused to think or were interrupted. Inferred enter treats life motion as a sign-up. User-executable means people can name what they just did to start and what they did to stop, and can do both again when they fail. A blink or a button as a gaze clutch is the same structure with a different effector.

Studying it

Run four protocols: enter only, leave only, enter plus leave, and inference throughout. Interleave command tasks with talking, turning, and pausing. Count idle false fires, unexpected cuts during the task, and whether people correctly report listening state. Also measure time from intending to stop to the system actually stopping. Wizard-of-Oz can have a human cut the engagement state first, to bound the gain of “both edges on purpose,” then swap in a real recognizer to see the drop. Do not measure on a pure command sequence; nobody there wants to leave.

Where it stops holding

A single hosted demo, where the presenter's voice stands in for enter and leave, makes the structure look optional. A physical button on a wearable already is a user-executed condition; stacking an in-air open/close on top is unnecessary. When a motor impairment blocks one action, change the effector (other hand, voice, foot) rather than falling back to pure inference. Leave that relies only on walking out of range fails in a living room where people do not actually leave. Timeout may exist as a backup leave; it must not be the only leave.

Applying it

  • Specify a pair of open and close actions people can demonstrate and repeat, and show on the interface whether listening is on or off.
  • In acceptance, insert “go talk, then come back and continue.” Check that people can leave on purpose and re-enter on purpose, rather than the system restoring itself.
  • Treat “enter only, leave by guess” as a defect. A backup timeout is allowed; the user must still be able to shut listening by hand or voice.

Related

  • Same group: C4.02.1 Always-on sensing makes every motion candidate input · C4.02.2 False triggering is structural, not an optimizable defect · C4.02.3 Raising thresholds cannot replace explicit engagement
  • Adjacent: C4.15 Conditions for entering engagement · C4.16 Disengagement and timeout
  • Search: engagement clutch · disengagement · user-initiated listening

Cards in the same group

Quick Actions

Share

Share this page

ios_share

https://hci.top/en/handbook/C4.02.4