The caller interrupts, the assistant stops, and then an old queued sentence plays over the new answer. Playback cancellation and conversation state have become inconsistent.
For engineers handling voice interruptions, test the entire output lifecycle. Stopping the currently audible fragment is only one part of preventing stale speech from returning.
What creates overlapping audio after interruption?
Possible boundaries include queued speech, asynchronous generation, delayed tool results and output already buffered downstream. The application must determine which response belongs to the current conversational state.
Do not assume cancelling one local generation task clears every speech item or every provider buffer. Verify the actual framework and provider behaviour through observable events and caller-side listening.
Build a stale-output test matrix
Use a fictional booking conversation with controlled delays. Record response identifiers and the input version each response represents.
| Interruption state | Expected evidence |
|---|---|
| Current sentence playing | Appropriate speech yields to caller |
| Next sentence queued | Obsolete output does not resume later |
| Tool result arrives late | Result is checked against current request |
| New answer begins | Only relevant output is scheduled |
| Caller changes action | Old completion wording is not replayed |
The important distinction is between output that remains relevant and output that has become stale. A brief caller acknowledgement does not necessarily invalidate the entire response, while a corrected appointment date can invalidate its central claim.
Inspect the framework's interruption semantics
The LiveKit turn-taking guide describes current speech interruption and queue behaviour. Verify the active SDK and pipeline before relying on a particular cancellation path.
Avoid adding a second independent audio player without a clear ownership rule. Two output mechanisms can each behave correctly in isolation while producing overlap when the caller interrupts.
The running-action interruption guide covers stale business results that can schedule another obsolete response after the first audio has already stopped.
Verify what the caller heard
Compare caller-side playback with application events and the final remote record. Conversation history should not imply that the caller heard information that was cut off before delivery.
My voice platform contribution provides project context. A new release needs its own interrupted-output fixtures and an explicit rule for scheduling, cancelling and replacing speech.
Keep the same tests after provider or buffering changes. A previously hidden delay can make an old race visible even when the application code responsible for the queue has not changed.
Does cancelling generation remove all buffered audio?
Not automatically. Inspect the framework, playback queue and downstream delivery behaviour.
Should every interruption discard the whole response?
Decide from the caller's meaning and the relevance of the pending output, with tested framework behaviour.
What should we inspect first?
Trace output identifiers from scheduling to caller playback during the voice AI conversation review.