Turn detection has been one of the more embarrassing failure modes in voice agents — cutting users off mid-sentence or leaving dead air. Fusing acoustic signal (pitch, pacing) with semantic understanding instead of waiting on transcript-only heuristics is the interesting part of this release.