Speech output must support interruption and replay
Aliases: speech interruption · replay · barge-in
What it is
Interruptibility and replay mean a user can cut off speech mid-utterance and can ask for already-heard content again. Because speech, unlike text, cannot be scrolled back, these two operations are the basic means of regaining control, not bonus features.
Why it happens
Speech vanishes once spoken; a missed fragment cannot be scrolled back, only remembered or regenerated. Listeners drift, get interrupted by their environment, or already have what they needed halfway through, so continued playback wastes attention and may mask what follows. Interruption also allocates conversational roles: output that cannot be interrupted leaves the user waiting and turns a two-way exchange into one-way broadcast.
Studying it
Two measures serve well: interruption latency, the gap between the stop command and actual silence, where longer gaps read as an unresponsive system; and replay cost, the steps and time required to recover a specific passage. Naturally occurring barge-in during conversation can set expectations for acceptable timing. Distinguish recognition latency from synthesis-stop latency, which users routinely conflate.
Where it stops holding
Not all speech should be freely interruptible. In safety announcements, legal notices, or flows requiring full acknowledgement, early termination can drop required information; the right move there is pause-and-record rather than abort. Short alert tones cannot be interrupted and are exempt from replay, since they carry state rather than content.
Applying it
- Make interruption available at any moment and stop audio immediately, without a lingering tail.
- After an interruption, offer explicit continuations: resume, replay this passage, restart, or skip.
- Keep a text record of critical speech so interrupted content can be reviewed rather than recalled.
- Verification: issue a stop command mid-utterance and measure actual stop latency, then ask users to restate the interrupted passage without external notes to confirm the replay path is reachable.
Related
- Within the group: D2.06.1 Speech conveys meaning well but structure poorly · D2.06.2 Speech output occupies the auditory channel and cannot be browsed in parallel
- Adjacent: D2.07 The linear, unscannable nature of hearing · M1.07 Barge-in and turn management in dialogue
- Search terms:
barge-in·speech interruption·replay