# What is turn detection in voice AI?

> Turn detection is how a voice agent decides that the caller has finished their turn and it is time to reply. Good turn detection waits through natural pauses without cutting the caller off, and does not leave awkward silences.

Source: https://agents.elevatebox.in/glossary/turn-detection  
Updated: 2026-10-12

## Why it is hard

People pause mid-thought, say 'umm', or trail off. Weak turn detection either interrupts a thinking pause or waits too long and feels slow. Tuning it well is a big part of why some agents feel natural and others feel like a bad phone menu.

## How agents detect the end of a turn

The simplest method waits for a fixed stretch of silence. Better systems also read the words and the tone: a finished question or a falling pitch signals the turn is over, while 'and...' or 'so the price is' signals more is coming. Code-mixed Indian speech adds fillers such as 'matlab' and 'ante', which a well-tuned agent learns to wait through.

## Frequently asked questions

### What is endpointing?

Endpointing is another name for detecting the end of the caller's speech. The term comes from speech recognition, where the system marks the end point of an utterance before it processes it.

### Why does an AI agent interrupt me?

Usually because its turn detection treats a short pause as the end of your turn. Agents tuned for Indian callers allow for thinking pauses and fillers before they reply. See also [barge-in](https://agents.elevatebox.in/glossary/barge-in).