Full Duplex

When two sides exchange data with each other, there are three possibilities. Either only one side may send at any time while the other listens. Or both may send, but taking turns. Or both send simultaneously, without having to wait for each other. The last case is called full duplex. A phone call works this way: you can talk while the other person is talking, and still hear them. A walkie-talkie does not work this way; there you have to press a button and say “over” at the end.

Why voice AI only feels natural because of this

Older voice assistants operated strictly by taking turns. You say something, wait, the device responds, you wait again. As long as the assistant was speaking, its microphone was practically deaf. Experts call this alternating pattern half duplex. It is technically simpler, but feels unnatural because real conversations rarely proceed in such a cleanly separated manner.

In a normal conversation, contributions constantly overlap. You throw in an “mhm”, you interrupt, you correct yourself mid-sentence. A system operating in full duplex keeps listening even while it is speaking itself. If you say “stop, that’s not what I meant”, it cuts off its answer. This exact difference determines whether a voice AI feels like a conversation partner or like a machine.

The difference is also measurable in networks. A full-duplex line can transmit the full amount of data in both directions. With a half-duplex line, both directions share the same capacity. In practice, this often means a doubling of throughput with the same hardware.

Two channels instead of one

Technically, full duplex requires two separate paths for the two directions. A network cable contains separate wire pairs for this purpose: one for sending, one for receiving. In radio connections, two different frequencies are used instead. Some systems also separate the directions in time, in very short segments, so that it appears simultaneous to humans.

With speech, an additional problem arises. The microphone also picks up its own loudspeaker. Without a countermeasure, the system would react to its own voice. This is prevented by so-called echo cancellation: the software knows the output signal and subtracts it back out of the microphone signal. Ideally, only the human voice remains.

With modern speech models, an ongoing evaluation runs in parallel at all times. The system continuously checks whether a sound is a genuine attempt to interrupt. An affirming “yes, exactly” should not cut off the response, whereas a “no, wait” should. This decision has to be made within fractions of a second, otherwise the conversation feels sluggish.

From network cable to customer service

The term has appeared in product descriptions for decades. Network cards and switches advertise full duplex, while walkie-talkies and simple intercoms operate in alternating mode instead. For headsets used in video conferences, too, this capability is a quality feature, because otherwise one’s own audio gets cut off as soon as someone else speaks.

In tech news over recent years, the term has come to stand primarily for voice AI. When a company introduces a “voice mode”, interruptibility is usually the central selling point. Companies deploy such systems in phone hotlines because callers there are especially impatient and rarely wait for an announcement to finish.

A common misconception is that full duplex simply means “fast”. That’s not true. It’s about simultaneity, not speed. A slow connection can be full duplex, while a very fast one may only be half duplex. When evaluating voice assistants, it is therefore worth taking a closer look: does the system really respond in the middle of its own sentence, or does it merely pause very briefly?