Smart turn: knowing when the turn is over

In the back-channeling post we covered the natural acknowledgment sounds the assistant gives while the caller is talking. A natural conversation asks a far more basic question before that one:

Did the caller really finish, or did they only pause for a moment?

The technology that answers this question is called smart turn.

The turn is harder than it looks

People take turns on the phone. When we finish, the other side takes over. It feels so natural that we never notice how hard it is to teach a machine.

Classic voice assistants make the call with a very simple rule:

• VAD (voice activity detection) notices that the sound stopped.
• A silence timer waits a fixed amount of time, for example 600 milliseconds. If no sound arrives in that window, the assistant assumes the caller is done and starts speaking.

That fixed timer is exactly where the problem comes from.

Two faces of the fixed timer

A single silence window cannot keep two different kinds of caller happy at once:

• Too short a timer: It cuts the slow speaker off. Someone reading an ID or a phone number in pieces (“123... 456...”) pauses naturally, the timer fires, and the assistant either interrupts or assumes they finished and heads the wrong way.
• Too long a timer: The fast speaker hits an uncomfortable gap after every sentence. The conversation feels slow and robotic.

There is no single magic silence window that pleases both a hurried 30 year old and a 70 year old reading digits carefully.

The fix: listen to the speech, not to the silence

Instead of a fixed timer, smart turn is a small AI model that listens to how the caller speaks. It reads the tone, the melody and the stress of the voice. Linguistics calls this prosody.

People signal “I am done” with their voice without noticing:

• A falling tone at the end of a sentence: I am done, your turn.
• A tone left hanging, or a pause in the middle: I am not done, I will continue.

So when the caller says “my number is one, two, three...” and pauses with the tone still up, the model says not done yet, wait, even if the audio is completely silent. When the caller says “...eight, nine.” and drops the tone, it says done, answer now.

Why it matters

What the smart turn approach buys in practice:

• It does not cut the caller off. People who read digits, addresses and long phrases in pieces can speak comfortably.
• It does not slow the fast speaker down. If the caller really finished, the assistant steps in without waiting, so no time is wasted.
• It is light and fast. The decision takes milliseconds and adds no meaningful load to latency.
• It is multilingual. Because it works from prosody, it holds up in many languages, Turkish included.
• It is open source and can run on-premise. That is a critical advantage wherever data is not allowed to leave the organisation.

Smart turn and back-channeling together

The two technologies are complementary halves of natural turn taking:

In the ID number scenario the two work together. Smart turn stops the assistant from cutting the caller off mid number. Back-channeling fills the longer pauses with a natural acknowledgment sound. The result is a flow with no 5 second dead wait at any step, one that treats fast and slow speakers alike as people.

Closing

Leaving the turn to a fixed timer bothers half of your callers on the phone. Listening to the tone of the speech gets the same decision right for both kinds of caller. Combine it with back-channeling and no dead silence is left in the flow.