Why we let you interrupt us

Listen to two friends talking for a minute. Really listen.

What you'll notice is that they almost never wait for each other to finish a sentence cleanly. One starts answering before the other has stopped asking. The other jumps back in to clarify. There's a "yeah, yeah" muttered halfway through, a "no wait — " interruption, a "go on, go on" overlap. Sentences finish in the middle. Subjects change without anyone saying so. Whole exchanges happen in fragments.

This is what conversation actually looks like. Tidy turn-taking — I speak, then you speak, then I speak — is the exception, not the norm. It's how we talk to strangers we want to impress, or to AIs in 2022.

Jaimy is built so you can talk to it the way you'd talk to a friend. You can interrupt. You can change your mind mid-sentence. You can mutter agreement while the tutor's still going. The fancy word for this is barge-in, and most language apps don't have it.

It seems like a small feature. It's not.

The problem with strict turn-taking

When an app forces you to wait for it to finish before you can respond, three things happen.

First, the conversation feels fake. Even if everything else is perfect — the voice quality, the topic, the difficulty — that one piece of artificiality breaks the spell. You're suddenly aware that you are talking to a machine. Your shoulders tense. Your output gets stiffer.

Second, you stop trying. If you know that you can't jump in until the AI finishes, and the AI has just started explaining something at length, your brain disengages. You're not listening to the explanation as a participant. You're waiting your turn. The thing being explained washes over you.

Third, you lose agency. Speaking is a deeply agency-driven act. You decide what to say, when to say it, when to back off, when to push forward. Strict turn-taking takes most of those decisions away from you. It tells you when to talk and when not to. It puts you in the passenger seat of your own conversation.

These problems compound. A speaker who feels they are not in control of the conversation talks less. A speaker who talks less learns less. The whole flywheel slows down.

The technical reason most apps skip this

Barge-in is hard. To do it well, the system has to:

• Be listening for your voice continuously, even while it is talking;

• Distinguish your voice from its own output coming through your speakers;

• Distinguish a real interruption ("wait, but —") from a non-interruption ("mhm");

• Stop its own speech cleanly, not awkwardly mid-syllable;

• Process what you said and respond, all without a noticeable pause.

Most language apps look at this list and decide it's not worth the engineering. They stick with turn-taking, accept the unnatural feel as a cost, and hope users don't notice.

We think users notice in their bones, even if they couldn't tell you what's wrong. The session feels formal. The conversation feels staged. The whole product feels a generation behind.

So we did the work.

What barge-in unlocks

Once you can interrupt, several other things become possible that weren't before.

Repair becomes natural. When you start a sentence and immediately realise it's going wrong, you can stop and restart, exactly the way you would in real life. Without barge-in, you have to finish the broken sentence and then ask to redo it. The repair feels surgical instead of natural.

Backchannelling is allowed. "Mhm. Mhm. Right. Yeah. Go on." These tiny utterances are how you signal to a speaker that you're tracking and they should keep going. In any natural conversation, you do this without thinking. Jaimy understands these for what they are — not interruptions — and keeps talking.

You can argue with Jaimy. Halfway through a Jaimy sentence you realise you disagree, and you say so. The tutor adapts. This is a small thing in any individual moment and a transformative thing across hundreds of moments. You are not being lectured at. You are talking to someone.

The pace is yours. Some learners want to think. Some learners want to ride the momentum. With barge-in, both work. The fast learner doesn't have to wait through information they already have. The slow learner gets the space they need without an impatient tutor barreling ahead.

A small philosophical point

A lot of voice product design comes from a place of slight condescension toward the user. The system is "in charge" of the interaction; the user is given turns; the system explains things; the user receives them.

We think the inverse is closer to right. The user is in charge of the interaction. The system's job is to be useful within whatever the user is doing — and to disappear, in some sense, into the conversation. The best language tutor you have ever had probably did not feel like they were running the session. They felt like they were available for whatever you needed in the moment.

Barge-in is the simplest, most concrete way we know to express that philosophy in code. You speak whenever you want to. We listen. The session is yours.

It's a small feature. It changes how it feels to use the product. And what we are trying to build, more than anything else, is a way of feeling — comfortable, in control, kindly listened to — that adult learners trying to speak a second language almost never get to feel.