Acoustic turn-taking and endpointing vs. silence VAD
Voice activity detection (VAD) estimates when speech is present. Endpointing decides when an application can act on a finished turn. A pause can mark the end of a thought, a breath, or a moment when the speaker is still choosing a word.
Updated
Listen to the recording
Listen around “I mean…” and “It really has.” Decide when you would start replying. Compare that decision with when you first hear a gap.
“I’m sorry. The price of food has gone up so much. I mean… It really has. What what happened? I don’t know. Like, honestly, I don’t know what happened.”
Audio from Open Yap 1K by The Agentic Data Company, licensed under CC BY 4.0. Oruk excerpted and downmixed this recording. The transcript is a saved model output. Recording credit and changes.
What a silence timer can tell you
A silence-based endpoint waits a set interval after the last speech frame. A longer interval gives hesitant speakers more time, but also makes completed requests feel slower. A short interval can cut off “I mean…” while the listener still expects more.
Turn-taking can also use intonation, lexical completion and which participant holds the floor. A quick “yeah” from a second speaker might acknowledge the first without asking to take over. A single channel can contain both voices, so detecting speech alone does not identify a change of speaker.
Try the comparison
Mark where listeners judge it acceptable to respond. For each candidate endpoint, count interruptions before that mark and measure delay after it. Report both on short commands, hesitations and overlapping turns. The grocery excerpt illustrates the annotation task and has no measured endpoint result.
Configure turn detection and interruptions in LiveKit
Try it on your own audio.
The demo needs no account. The trial needs a card.