The short answer
Interruption handling is the layer above barge-in. Barge-in stops the audio; interruption handling decides what the interruption meant and what to say next. Getting it wrong is conspicuous: an agent that restarts a cancelled sentence from the beginning, or ignores a correction and continues its script, feels broken even when the audio path worked perfectly.
In detail
Not all overlapping speech is an interruption. Backchannels such as "mm-hm" and "right" signal attention and should not stop the agent at all. Corrections — "no, the fifteenth" — must both stop playback and revise the state the agent had recorded. Topic changes abandon the current thread entirely. Treating these three identically produces an agent that either stops constantly or never stops.
When an interruption is genuine, the agent’s belief about the conversation must be corrected. It knows what it generated, but the caller only heard what was played before the cut. Resuming from where the caller actually stopped hearing, rather than from where generation stopped, requires tracking playback position — otherwise the agent refers to information the caller never received.
Recovery strategy then depends on how much was delivered. If the interruption arrived early, restating the point concisely is usually right. If most of it was heard, repeating it is patronising and answering the interruption directly is better. The general rule is to acknowledge the interruption and address it rather than treating it as noise to be spoken over.
Related terms
Barge-in
Barge-in is the ability of a caller to interrupt a system’s audio playback by speaking over it.
Turn detection
Turn detection is the decision that the floor has passed from one participant in a conversation to the other.
Voice agent
A voice agent is software that conducts a spoken conversation with a person over a phone line or a browser connection, understanding natural speech and responding in speech.
Endpointing
Endpointing is the decision that a speaker has finished the utterance they were producing.