Glossary

Barge-in

Barge-in is the ability of a voice system to stop its own output as soon as the caller starts to speak.

  • Also: interruption, talk-over handling
On this page
  1. In practice
  2. An example
  3. A common misunderstanding
  4. Read more

Without barge-in, callers have to wait until an announcement has finished. With barge-in they can correct something (“No, Thursday”) or cut a long answer short. Technically, the system keeps listening while it speaks, detects speech on the other side and stops the output. For natural conversations this is a basic requirement, because people briefly interrupt each other all the time.

The difficult part is telling a real interruption from noise: a cough, an “mhm” of agreement or background noise should not cut off the reply. Systems can therefore be tuned in their sensitivity. What happens after the interruption matters too: the system has to know which part of its answer the caller actually heard.

In practice

Test calls quickly show whether barge-in is tuned well. Interrupt the assistant mid-sentence with a correction, say nothing but “yes” or “okay” in between, cough, and call once from a noisy place. It should fall silent at the correction, keep talking after “okay” and not be thrown by the noise. Short answers help as well. The longer an assistant speaks in one go, the more often it gets interrupted.

An example

A patient calls a dental practice to move an appointment. The assistant begins: “I can offer you Tuesday at nine or Wednesday at …” She cuts in: “Tuesday is no good, I work in the mornings.” With barge-in it stops and offers afternoon slots. Without barge-in it finishes its list, does not hear her objection and then asks which of the appointments it named would suit her.

A common misunderstanding

Barge-in tends to be set as sensitive as possible so that the assistant never talks over anyone. The result is the opposite of a good conversation: every cleared throat and every background sound stops the answer, and the assistant keeps starting over. A workable setting is a compromise. There are also sentences it should finish, such as the notice at the start that an AI is speaking.

Read more