Full-Duplex

Full-duplex describes a connection in which both sides can send and receive simultaneously. Its counterpart is called half-duplex: there, only one direction is ever free, while the other must wait. A walkie-talkie is the classic example of half-duplex. You press a button, speak, say “over” and release it — only then can the other person respond. A telephone, by contrast, operates full-duplex: both parties talk, both listen, and if need be, talk over each other. The term originally comes from communications engineering, but today it is used above all for talking AI systems.

Why conversations with AI long felt stilted

The first voice assistants were technically half-duplex. You said a wake word, spoke your sentence, and then there was a pause. The system processed the recording and responded. Meanwhile, the microphone was essentially deaf. Interrupting was not supported.

That’s exactly what makes such conversations feel unnatural. Human conversations largely consist of overlaps. We say “mhm” and “right” while the other person is still talking. We cut in when an answer is heading in the wrong direction. And we can tell from pauses whether someone is finished or just thinking.

A full-duplex system can replicate this. If the AI gives a rambling answer, you simply say “stop, shorter” — and it stops. That sounds like a minor detail, but it dramatically changes how the experience feels. Only this makes a voice assistant feel like a conversation partner rather than a machine responding to voice commands.

Listening while speaking yourself

The biggest technical problem is mundane yet stubborn: the speaker is talking right next to the microphone. So the device mainly hears itself. Without a countermeasure, the AI would mistake its own voice for user input. This is addressed by so-called echo cancellation. The system knows its own output signal and subtracts it back out of the microphone signal. Ideally, only the human voice remains.

The second building block is continuous evaluation of the microphone. The model constantly checks whether speech is currently coming in and whether it is directed at the system. A cough or a conversation in the background should not interrupt the response. A clear “no, wait,” however, should. This decision is made in a fraction of a second.

Modern systems no longer solve this with a chain of speech recognition, text model, and speech output. Instead, a single model processes audio directly and also generates audio directly. This saves time. Instead of one to two seconds of delay, typical response times then come in at a few hundred milliseconds. That roughly matches the pace of a human conversation.

From network technology to customer service

In classic technology, full-duplex has long been standard. A network cable transmits in both directions simultaneously, as does the mobile network during a phone call. Video conferencing software also operates full-duplex, which is why talking over each other works there too — with all the well-known side effects.

In product announcements today, the term mainly comes up in connection with voice models. When a provider advertises a voice mode as “naturally interruptible,” full-duplex is behind it. Automated phone assistants in customer service also rely on it, since on the phone every delay feels especially disruptive.

A common misconception: full-duplex does not automatically mean a system answers intelligently. It only describes the flow of conversation, not the quality of the content. An assistant can be smoothly interruptible and still talk nonsense. Conversely, a slow system can deliver very good answers. The two properties are independent of each other.