Self-Listening for Full-Duplex Speech Model Interruption Recovery
September 3, 2026
Self-Listening addresses anchor interruption in full-duplex models by interleaving user speech, model text, and realized audio playback. By feeding the actual played speech back into the model's input stream, the approach ensures the model's internal state matches its actual vocalized output during conversation.
HOW THIS AFFECTS YOU
●
builderYou can implement more natural conversational agents that recover gracefully from user interruptions.
●
researcherThis provides a method for grounding speech models in their own realized output rather than just text intent.