Changelog

A caller who speaks over the agent now stops it within a few hundred milliseconds instead of about a second. The agent pauses on the caller’s speech onset, then decides from what it hears next: a real interjection commits as a turn, a short acknowledgement (“mhm”, “yeah”, “got it”) mid-sentence lets the agent carry on from where it paused, and a noise that never transcribes releases the pause within about a second.

The interruption_sensitivity levels keep their meaning and move with the change: medium (and null) pauses after roughly 0.2 s of caller speech, high after 0.1 s, low after 0.5 s and still needs three transcribed words. Nothing you send changes.

Two behaviours you may notice on existing agents: the greeting on a phone call can now be interrupted from its first word, and a paused reply no longer resumes over a caller who has already started their next sentence.

Acknowledgement suppression is English-only for now; agents in other languages treat every interjection as a turn, as before.