Turn-Taking Latency

When two people talk to each other, they take turns. One stops speaking, the other begins. In speech research, this handover is called a speaker change, or turn-taking. Turn-taking latency measures how long this transition takes with a computer program. What is meant is the pause between the last word of the human and the first word of the machine. In conversations between humans, this pause averages around 200 milliseconds, or one fifth of a second. Voice assistants used to take noticeably longer, often one to three seconds.

Why a one-second pause feels wrong

Humans react very sensitively to conversational pauses. Even a delay of half a second is interpreted by us as hesitation. We then suspect uncertainty, rejection, or a bad connection. With a machine, the impression arises that it did not understand the question. Many users then repeat themselves or speak louder. Exactly at that moment the response begins, and both end up talking over each other.

That is why latency is a central metric for voice products. An assistant can be excellent in terms of content and still seem unusable if it answers too late. Companies like OpenAI, Google, and Amazon explicitly advertise low response times for new voice models. Values around 300 milliseconds are now considered the target for systems meant to feel like a real conversational partner.

This is especially important in applications where time costs money. An automated phone service at an insurance company handles thousands of calls every day. If every response is one second too slow, that noticeably lengthens each call. And callers abandon calls more often if the system feels sluggish.

What actually happens during the pause

The delay does not arise at a single point. Classic voice assistants work through several steps in sequence. First, the system must recognize that the human has actually finished speaking. Then the recording is converted into text. This text goes to the language model, which formulates a response. Finally, another program converts the response text into audible speech. Each of these steps takes time, and the times add up.

The first step is surprisingly difficult. A system must distinguish whether someone is pausing to think or has really finished. If it waits too long, it seems sluggish. If it reacts too quickly, it interrupts the user mid-sentence. Many systems therefore wait for a fixed 500 to 800 milliseconds of silence. This safety pause is often the single largest contributor to overall latency.

Newer approaches shorten the chain instead of speeding up individual steps. So-called speech-to-speech models process audio directly and output audio directly, without the detour through written text. In addition, output often begins before the response sentence has even finished being planned. So you hear the beginning while the model is still computing the rest. A common misconception is that a faster chip alone would solve the problem. The architecture of the system usually matters more than raw computing power.

From Alexa to video conferencing

Latency is most noticeable with voice assistants on phones or smart speakers. The familiar moment where you briefly wait in silence after asking a question is exactly this metric. Translation features in video conferences also struggle with this, because network delay is added on top of response time. Even good internet connections easily add 50 to 100 milliseconds.

In product announcements and tests, the term usually appears as a number in milliseconds. Pay attention to exactly what was measured. Some manufacturers only state the model’s computation time and leave out recognition, speech output, and network transit. But for the user, only the perceived pause from their own last word to the machine’s first word counts. This total time is almost always noticeably higher than the advertised figure.