Blog · 10 October 2026 · 2 min read

How the engine decides it is its turn to speak

End-of-turn timers, interruptions and echo, explained with the trade-offs each setting makes.

People judge a voice agent in the first two seconds, and mostly on one thing: does it talk at the right moments? An agent that answers too early cuts you off. One that answers too late feels broken. Here is how the voice engine makes that call, and what you can tune.

Knowing the caller has finished

The engine has no way to know that a sentence is complete. It knows that the caller has stopped making speech sounds, and for how long. So the end of a turn is a silence timer, and the only question is how long to wait.

One fixed value is always wrong for somebody. The engine uses three:

A short reply such as "umm" is often the start of something longer, so the engine waits longer after it. If the caller does resume before the agent has become audible, the two parts are merged into one turn and answered together.

This wait is the largest single part of the response time. Lowering it makes the agent quicker and makes it interrupt people who pause to think. There is no setting that avoids the trade.

Being interrupted

When the caller speaks while the agent is talking, the engine does not immediately give up the turn. It pauses playback at once, which is what the caller notices, and then waits for the transcript of what was said.

• If it was a real interruption, the reply is cancelled and the caller has the turn. The transcript keeps only the part of the reply the caller had started to hear.
• If it was a short acknowledgement ("ok", "haan"), playback resumes where it paused.

The threshold is a number of words, set per agent. One word makes the agent easy to interrupt; a higher number makes it steadier on lines with background voices.

Not interrupting itself

On many phone lines the agent's own voice comes back, quieter and delayed. Without care, the agent hears itself and stops talking. The engine has an echo guard for this: it compares incoming audio with what it has just played, and compares the transcript of the "interruption" with the sentence it was saying. A close match is treated as echo and ignored.

What it does not do

• It does not cancel echo. It recognises it. That was tuned on simulated echo.
• It cannot tell the caller from other voices. A television in the room is speech.
• Speech-to-text runs after the caller stops, not while they speak. An interruption is confirmed only when its transcript arrives; until then playback stays paused.

Tuning

Every value above is a setting of the engine, and each turn's timing breakdown is logged: how long the end-of-turn wait was, how long transcription took, when the model's first sentence arrived, when the first audio was ready. If the agent feels slow, that record tells you which stage to look at before you touch a timer.

The settings are listed in Settings, and the pipeline in Voice engine.

← All posts