What is backchanneling in voice AI?

Definition

Backchanneling is the short feedback a listener gives while another person holds the floor, such as "mm-hm", "right" or "okay". It signals attention without claiming a turn. For voice agents, it is a sound that must not be treated as an interruption.

How backchanneling works

In ordinary conversation, the listener is not silent. They say “uh-huh” or “yeah” while the speaker talks, nod, or make small sounds that mean “go on”. Linguists call these backchannels. They are short, they usually land at natural pauses in the speaker’s sentence, and the speaker keeps talking after them without treating them as an answer.

That is the key property. A backchannel does not take the turn. “Mm-hm” and “wait, that’s the wrong address” are both short interjections that start while the agent is talking. Only the second changes what should happen next.

Two directions matter for a voice agent.

  • Hearing backchannels. The caller says “okay” or “right” while the agent explains something. The agent should keep going. If it treats the sound as an interruption, it stops mid-sentence, waits for more, and the caller has to say “go ahead”.
  • Producing backchannels. The agent says a brief “mm-hm” while the caller reads out a long number or tells a story. This tells the caller the line is live and the agent is following. Without it, a long stretch of caller speech meets silence, and people start asking “are you there?”

Why it matters for voice agents

Handled badly, backchanneling makes barge-in feel broken. A system that stops on any detected speech, including a one-word acknowledgment, makes the conversation stutter. The usual fixes are to ignore very short speech while the agent is talking, to require a few hundred milliseconds of sustained speech before treating it as an interruption, or to check the words recognized and ignore a short list of acknowledgments. Each has a cost. A longer threshold also delays real interruptions, and a word list fails on acknowledgments it has not seen. There is no setting that solves it, so test with recordings of callers saying “okay” at the wrong moment.

Producing backchannels needs more care than hearing them. An “mm-hm” at the wrong point overlaps the caller, and the agent’s own audio can then be picked up by its speech recognizer. It should also stay out of the middle of a digit sequence the caller is reading out. A simpler alternative is a short confirmation at the end of each chunk, such as “got it, and the next part?”.

Backchanneling varies with language and culture, in how often it occurs and which sounds are used. A threshold tuned on one language can misfire on another.

Example

An illustrative call where the agent reads out a policy:

  • Agent: “Your plan renews on the first of the month, and the price…”
  • Caller: “Mm-hm.”
  • Agent: “…stays the same unless you change your coverage.”

The agent treats “mm-hm” as a backchannel and carries on. If the caller had said “hold on”, the agent should stop.

Guides

Get started today

Talk to an expert.

Connect with a member of our team and learn how Cartesia can help you build world-class voice experiences.

Contact Sales

Start building.

Access our models via API and bring a voice agent into production in minutes.

Try Cartesia