We tried four ways to stop our AI talking over people, only one actually worked.

If you have used a voice agent you have had this happen. You pause mid-sentence to think, it decides you have finished, and it starts answering. It does not read as slow. It reads as not listening, which is worse.

Here is what we tried, roughly in order.

Shorter silence threshold. Cut the wait before the agent decides your turn is over. Faster, and much worse, because ordinary speech is full of pauses and now it talks over you constantly.

Longer silence threshold. The opposite problem. It stops interrupting and starts feeling dead. You finish a sentence and sit there wondering whether it heard you at all.

Semantic endpointing. Stop counting silence and judge whether the sentence is actually complete. Genuinely better, and it fails in a specific way: people do not speak in complete sentences, and a trailing clause looks finished right up until it is not.

Then we stopped trying to get the prediction right and made being wrong cheap instead. Keep listening while it speaks, and the moment the person starts talking, cut the playback immediately. It still misjudges the end of a turn constantly. It just no longer matters, because the person takes their turn back and nothing is lost.

The lesson, if there is one: we spent a long time trying to predict when someone had finished, when the fix was to stop needing to be right about it.

If you have built anything conversational, where did you land?