What is end-of-turn detection?

Silence is the crude version of this and the one most systems still use. Wait for a pause of a fixed length, then talk.

Humans do something much richer. We read falling pitch, grammar that has completed, and the rhythm of the sentence, which is why we interrupt each other so accurately.

Better systems mix the signals. A pause after a completed question is treated differently from a pause in the middle of a list.

This is the part of a voice assistant people describe as feeling natural or not, and they almost never name it directly.

Say a caller is rescheduling and says: I'd like to come in Tuesday. Then, after a breath: or actually, do you have anything Monday?

Her first sentence is complete. Grammar says she's done, her pitch fell, and she paused. A system that jumps at that point starts listing Tuesday openings while she's saying the word Monday. Now there are two voices, a wrong answer in progress, and a caller who has to start over.

Compare a different pause. Asked for her address, she says my address is, and stops to check the unit number on a letter. Nobody would take that as finished, because the sentence is hanging open. Filler sounds work the same way. Um and so are how speakers hold their place, and a system that talks over an um is ignoring the clearest signal it'll get.

Neither case can be solved with a timer alone. Both pauses might last 800 milliseconds, and one means go ahead while the other means wait.

Too early and too late both cost you, but not equally. When the assistant jumps in early, the caller stops, the assistant stops, both wait, and both start again. A single collision can easily eat five seconds. It also teaches the caller to speak in short, clipped bursts for the rest of the call, which gives you worse information.

When it waits too long, nothing breaks. The call feels slow, and the caller may say hello to check the line is alive. Most people forgive that more readily than being cut off.

So if a system lets you choose, lean toward patience on questions where people think before answering. An address or a description of the problem are the usual ones. Lean quicker on yes or no questions, where a long wait feels odd.

This is a different job from interruption handling, though the two get blurred. End of turn is about when the assistant starts speaking. Interruption handling is about when it stops.

GreetKeeper publishes no timing figure, because we have not benchmarked one honestly and a number here would be easy to make up.

Judge it on a call instead. Two minutes of your own scenario tells you more about the rhythm than any specification would.

Where the assistant jumps in early, the transcript shows the collision and the prompt is usually the thing to change.