Blog · For developers
Full duplex vs turn-based voice: can your caller interrupt?
People talk over each other all the time. What your voice agent does in that moment tells the caller whether it's listening.
By the CosVoice team · · 5 minute read
What's the quickest way to tell them apart?
Interrupt it. Say "sorry, hang on" in the middle of the agent's sentence. Either it stops, or it finishes what it was saying and then gets round to you. That one behavior splits voice systems into two families, full duplex and turn-based, and most of what callers feel about a voice agent follows from which one they've met.
Picture the line reading back a booking. "So that's a table for four on Friday at seven thirty, under Graham, and I've noted the" and the host cuts in: "Sorry, did you say four? We've only got a two at that time." A turn-based voice carries on to the end of its sentence about the window seat. A full duplex voice stops at "sorry" and deals with the actual problem.
Neither is broken. They're two designs. But the host formed a view of your agent in that second.
How does turn-based voice work?
Like a walkie-talkie with the button hidden. One side talks, then the other. While the agent is speaking it isn't really listening. When it finishes, it listens for the caller to speak and then stop, and takes its turn.
There's a lot to like. It's simple to reason about. The audio is clean, because the system never has to separate its own voice from the caller's. Replies can start quickly. And every turn is a complete thought, which makes for tidy transcripts.
What does full duplex change?
Both directions are live at once. The agent listens while it talks. When the caller speaks over it (the trade calls this barge-in), the agent stops, drops whatever it was about to say, and attends to the new words.
That's three hard problems in a coat. The system has to tell the caller's voice from the echo of its own. It has to tell a real interruption from a cough, an "mm-hm" or a dog. And it has to keep track of how much it actually got out before it was cut off, because the record of the conversation should hold what the caller heard, not what the agent had planned to say.
Isn't talking over someone just rude?
Between people, mostly not. We overlap constantly. "Yeah." "Right." "No, the other one." Overlap is how a listener steers a speaker without waiting for the end of the paragraph. It's how you stop someone reading out an address you already have.
On a call with a business it's also how busy people save time. The person at the host stand on a Friday evening isn't going to wait for a well-formed sentence to finish. If your agent can't be steered, the call takes longer and the human does the adapting. They'll manage. They won't enjoy it.
When is turn-based the better choice?
More often than fans of full duplex like to admit.
- Short, factual exchanges. Opening hours, an address, do you take cards. Nobody interrupts a six-word answer.
- Noisy lines. A workshop, a kitchen, a van with the window down. Background noise can set off barge-in and stop a full duplex voice for no reason.
- When clarity comes first. Reading back a reference number or a date, you want the crispest audio you can get.
- When speed of reply comes first. A front desk where every caller asks one thing and wants it answered now.
And when do you need full duplex?
When the conversation is a negotiation, even a small one. Booking a table when the first time is gone. Getting a quote where the contractor asks questions back. Anything with a hold, a transfer, and a person who picks up mid-sentence. Calls like these are made of corrections, and a correction is an interruption.
Longer answers need it too. If the agent is going to speak for more than a sentence or two, the caller needs a way to say "got it" and move on.
How do you test the one you have?
Ring it and be difficult. Say "mm-hm" while it's talking and see whether it stops (it shouldn't). Cough. Read out a phone number with gaps between the groups and see if it jumps in. Correct it halfway through a read-back.
Then read the transcript, and check it shows what was said on the call in the order it was said. If you'd rather listen before you dial, there are full calls on our demo page.
Which does our line use?
Both, and you choose. Our line has 38 voices in two sets. The new voices can be interrupted: talk over one and it stops and listens, the way a person does. They sound more natural, take a little longer to answer, and on some phone lines they're a little less crisp.
The classic voices are the clearest on a phone line and the quickest to answer. But they finish the sentence before they listen, so a caller who cuts in has to wait for a pause.
Switching is one click in Settings and applies to the next call, and an agent can do it with update_settings. So you don't have to get it right first time. Start with whichever suits the calls you expect, read a week of summaries, and change it if callers are fighting the voice. There's a guide in choosing a voice for your phone line, and the speed side of the trade is in latency in voice AI.
Common questions
What does full duplex mean in voice AI?
Both sides of the call are live at once, so the agent listens while it speaks. If the caller talks over it, it stops and responds to what they said.
What is barge-in?
Barge-in is the caller speaking while the agent is still talking. A system that supports it hears the caller, stops its own speech and listens.
Is full duplex always better than turn-based?
No. Turn-based voices can be clearer and quicker to reply, and they suit short factual calls and noisy lines. Full duplex suits conversations with corrections and back-and-forth.
Can I switch between the two on a CosVoice line?
Yes. Pick a new voice for one that can be interrupted, or a classic voice for the clearest and quickest replies. It's one click in Settings and applies to the next call.
Read next: Where the delay comes from · Choosing a voice for your line · AI voice agent vs a phone menu
Hear it for yourself.
Call our own AI agent and ask her anything, or get a number in your area code in about a minute.