The caller interrupts a booking lookup to change the location, but a late result from the old lookup still gets spoken. Stopping the assistant's audio is not enough when background work belongs to an earlier request.
For engineers testing interruptions, inspect the relationship between conversation state and operation state. The current caller intent should determine whether an old result remains relevant.
What does interruption during an action change?
An interruption can stop speech, revise inputs, cancel a proposed action or add unrelated context. Those meanings require different handling even when they begin with the same audio event.
The application needs to know which version of the request produced each operation. A result can arrive successfully and still be stale for the caller's current instruction.
Build an in-flight interruption matrix
Use fictional appointments and controlled remote effects. The proposed cases below separate speech behaviour from business correctness.
| Interruption moment | Expected evidence |
|---|---|
| Before lookup submission | Only confirmed current inputs are used |
| During read-only lookup | Old result cannot overwrite a correction |
| Before authorised write | Changed request receives appropriate confirmation |
| After write submission | Remote outcome is reconciled before another mutation |
| During success speech | Saved state remains distinguishable from heard wording |
Do not assume an interrupted local task means a remote service cancelled the work. The disconnect and pending-action guide describes the related uncertain-write boundary.
Observe audio and operation events together
Record when caller speech begins, when assistant playback stops, which input version is active and which remote request was submitted. This exposes failures that a transcript alone can miss.
The LiveKit turn-taking reference describes current interruption mechanisms for that framework. Verify the behaviour of your actual pipeline and tool execution model instead of assuming speech interruption controls every backend task.
Test a correction that changes a consequential field, such as location or date. A harmless change in conversational wording may not reveal stale data reaching the booking action.
Review the final state, not only the overlap
Inspect the destination record and the next message the caller hears. The assistant should explain the current outcome accurately, including cases where the old action already completed and a separate authorised correction is needed.
My voice platform contribution supplies relevant system context. Keep these scenarios in the fictional test set with expected mutations and forbidden effects.
Repeat the set after changes to the model, tool runner or interruption controls. A conversation that sounds more responsive can still become less correct if old operation results are allowed back into the current turn.
Does stopping audio cancel the backend action?
Not necessarily. Inspect the operation's submission and remote completion state separately.
Should every interruption discard all collected details?
Preserve valid context and apply the caller's actual correction rather than resetting the whole workflow automatically.
What should we test first?
Interrupt a controlled lookup with a location change and inspect stale-result handling in the voice AI technical review.