What's this about?
Self-Listening feeds a full-duplex voice model the speech audio that was actually played, alongside user speech and model text, improving recovery after interruptions.
What exactly is an “anchor interruption”?
Text generation, speech synthesis and playback happen asynchronously. If a user interrupts, the model may think it said words the speaker never played; the anchor is the last item the user truly heard.
Why is this worth noting?
Natural voice agents need to listen while speaking and handle backchannels. Closing the gap between intended and realized speech prevents confusing repetitions or skipped steps.