Overlap does not have one meaning
A caller can speak while an agent is talking for several reasons: to agree, correct a detail, ask a question, or stop the response. Background speech and the agent’s own voice leaking into the microphone can also produce apparent overlap. Treating all detected speech as an identical event creates unnecessary stops.
Start with an application policy. During a general explanation, a quiet acknowledgment may not need a new turn. During confirmation of a delivery address, a brief “no” can be essential. The words, their timing, and the current task all matter.
Separate detection from the decision
Voice activity detection estimates where speech is present. An interruption decision asks whether incoming activity should take control of the conversation. Acoustic cues such as onset, duration, rhythm, and pitch can contribute, while a transcript can add linguistic context.
Voice activity projection research explores predicting future conversational activity from audio. [2] That supports richer timing models, but it does not establish a universal classifier of user intent. A threshold that works for one language or speaking style may need reevaluation elsewhere.
Use reversible behavior during uncertainty
One possible policy briefly reduces or pauses output while gathering enough evidence to distinguish a correction from incidental sound. Another stops immediately for a confidently recognized request to stop. These are design choices whose value must be tested against detection delay and missed interruptions.
Do not resume a canceled answer blindly after classifying a sound as non-interruptive. Check whether the user has started a longer turn and whether the original response is still relevant. A false interruption should not turn into repeated restarts of the same sentence.
Cancel output across the whole path
When an interruption is accepted, stop generating obsolete content, stop synthesis where possible, and clear audio waiting for playback. Measure the remaining sound on the client. A server-side cancellation event can occur before the listener actually hears the agent stop.
Keep a record of the spoken portion of a response. If the user interrupts halfway through instructions, the next turn should not assume the unheard steps were delivered. Pending external actions require a separate policy: stopping speech does not reverse a completed operation.
Test for both kinds of error
Collect examples of genuine interruptions and non-interruptive overlap, including one-word corrections, acknowledgments, coughs, television speech, and speakerphone echo. Use representative recordings only with appropriate handling and permissions. Include different devices, languages, and speaking volumes.
Report missed interruptions, unnecessary stops, and delay until audible output stops. Review task completion too: an agent that never pauses for acknowledgments can still be poor if it ignores corrections. Choose thresholds according to the cost of each error, and provide a manual stop or alternative interaction path when automatic interpretation remains uncertain.
Research and references
- [1] How Much Does Prosody Help Turn-taking? ↗
- [2] Real-time and Continuous Turn-taking Prediction Using Voice Activity Projection ↗
- [3] Multilingual Turn-taking Prediction Using Voice Activity Projection ↗
Technical review: November 9, 2026. Research supports the underlying concepts; implementation recommendations and illustrative scenarios are editorial guidance, not measured product results.