Turns and interruptions
VAD detects activity. End-of-turn needs evidence about the complete intent.
Links containing ?t= open the video at a specific second.
The ideas to retain
VAD answers “is there speech?”, not “is the thought complete?”
A voice activity detector classifies regions of audio as speech or non-speech. Silero VAD, for example, exposes speech timestamps and is explicitly designed for voice activity…
A timeout is a policy, not semantic proof
The simplest endpointing policy waits for a silence window:
Endpointing: deciding when to commit the turn
The endpoint is the moment the system stops treating user input as an open turn and allows a response to depend on it.
Jump directly to a section
Read the reviewed transcript
This video has no narration. This transcript reproduces its on-screen text; it does not invent a spoken track.
A pause does not finish the thought
“Change the address… for order 4182.” The internal pause lasts 350 milliseconds in this example.
Closing after 250 milliseconds cuts the request before the order is identified.
VAD detects activity. End-of-turn needs evidence about the complete intent.
Waiting has a cost too
A short window can mistake a pause for the end. A long one can create silence.
Increasing the timeout protects this unfinished sentence, but delays an already complete “yes”.
Tune premature endings and waiting together. A lower latency number is not enough.
Fast start. Evidence for the end.
Detecting the user’s speech early lets the system consider yielding the channel.
Closing the user’s turn too early can change the meaning of the request.
Use separate start and end decisions, with a clear authority for each.
Do they want the turn?
“Uh-huh” can signal attention without asking the agent to stop speaking.
“Wait, change the time” can require stopping the response and listening.
Context decides: detected speech is evidence, not a universal cancellation command.
Stopping generation is not enough
R7 already has one played chunk, more queued chunks and more audio being generated.
The interruption stops new output and removes pending audio from the queue.
Keep the confirmed prefix. Do not confuse generated audio with actual playout.
One turn. One decision.
Two detectors can propose closing T12. They must not start two independent responses.
One authority admits the closure once and assigns response R7.
When T13 replaces the turn, late R7 audio must not reclaim the channel.
