Turn detection

Human turn-taking is coordinated by a bundle of cues that overlap in time: syntactic completion, falling intonation, the end of a gesture, gaze shifts, and stereotyped closings such as "so, yeah". Gaps between speakers are typically short enough that the next speaker must have begun planning before the current one finished, which is why simply waiting for silence produces conversations that feel slower than natural ones.

Semantic turn detectors run a small model over the running transcript to estimate whether the caller intends to continue. This catches cases pure acoustics miss: "I need to change my appointment" is a complete sentence but an obviously incomplete turn, because the caller has not yet said what to change it to. It also catches the reverse — a grammatically fragmentary "yeah, no, that one" that is nonetheless a finished turn.

Turn detection interacts with backchannels. Short acknowledgements such as "mm-hm", "right" and "okay" are explicitly not bids for the floor, and an agent that treats them as turns will interrupt a caller who is simply signalling that they are still listening. Filtering backchannels out before turn logic runs is a small change with a large effect on perceived quality.