The End of Annoying AI Interruptions? LiveKit Turn Detector v1 Tested

[00:00:00]Let me show you something that almost every voice agent gets wrong. Watch what happens when I talk to this one and I take a breath in the middle of a thought. I'd like to order a large pizza.

[00:00:15]And some >> What toppings would you like on that large pizza?

[00:00:19]>> And it already cut me off. I wasn't done and that's the fastest way to make an agent feel robotic and frustrate a user.

[00:00:26]And almost all of the voice agents that you use today do this. So let me show you how to fix this live. So here are two agents with the same setup, the same voice, same prompt. The only difference is the turn detection. On the right is LifeKit's new audio-based turn detector V1, which listens to my audio, not just a text transcription. On the left is the simplest baseline, a plain voice activity detection, which uh just waits for silence. I'm going to talk to one at a time and watch the state badge flip from listening to thinking. That's the moment that it decides that I'm done.

[00:01:05]And then I'm going to say the same line to each, uh starting with the VAD agent on the left. I'd like to order a large pizza and >> What toppings would you like on that large pizza?

[00:01:20]>> I'd like to order a large pizza and garlic bread.

[00:01:29]>> Hey there. I've got a large pizza and garlic bread.

[00:01:32]What toppings would you like on the pizza?

[00:01:34]>> So on the left, a plain VAD heard the pause, it flipped straight to thinking and it jumped in. I wasn't done. Our new turn detection V1 heard that my tone was still trailing upwards and so it continued listening and it let me finish because it realized based on audio that I wasn't I wasn't done. And it's not just beating plain VAD. Let me switch the baseline to our older text-based model, the kind of most agents use today. So, this is a text-based turn detector, which on its own is better than bad, but let's see how it does.

[00:02:08]Yeah, I'd like to order a large pizza and some garlic bread.

[00:02:15]>> for calling. What toppings would you like on that large pizza?

[00:02:18]>> Oh, even the text-based model reads the transcript, it sees a complete sentence and it jumps in. The transcript cannot tell these things apart, only the audio can. Now, for the opposite case, a real ending, let's go back to the V1 model and watch how fast it responds when I actually stop.

[00:02:38]Can you add a soda to my order?

[00:02:42]>> Sure. What kind of soda would you like?

[00:02:44]>> So, it went to thinking right away, there was no awkward delay. It wasn't waiting because it knows that I'm done because of how I said it, not just what I said. So, why does the text-based model still fail? Well, it decides that I'm finished by reading the transcript and the transcript throws away how I said the words. I would like to order one large pizza reads exactly the same whether I'm done or I'm about to keep going. Only the audio carries over my inflection. Life Kit Turn Detector V1 fixes this by listening to my audio directly. Under the hood, it runs two branches at once.

[00:03:22]One reads the meaning of what I'm saying and the other reads the music of my voice, the timing, the pitch, the rhythm. It fuses them into a single prediction of whether I'm done or not.

[00:03:35]Now, this is not just a nice demo. We measured it. We evaluated every model under full endpointing policies, which is the real trade-off between responding fast and cutting people off. At a 300 ms budget, V1 has a 9.9% false cutoff rate, where the next best deep ground flux sits at 12.9% and the plain voice activity detection baseline like the one that we used on the left side is over 55%.

[00:04:04]If we give it a little more room at 600 milliseconds, the lead holds.

[00:04:09]And it is the strongest multilingual model overall across 14 languages, not just English. And so we're open-sourcing the entire benchmark suite and the data set so that you can check the numbers yourself. The best part is how little you have to do. On LifeKit Cloud V1 is already the default. So for most agents, there is no setup needed. If you want to set it up explicitly, that's just a few lines of code. But there are two models.

[00:04:35]V1 is the larger, most accurate version free for agents on LifeKit Cloud. V1 Mini is open access and small enough to run fast on a CPU so you can use it locally or self-hosted. Same idea, you pick depending on where you're running.

[00:04:51]If your agent talks over people, this is the fix today. Try LifeKit Turn Detector V1, read the full benchmark breakdown, and tell us how it does on your own conversations. Links are in the description. If this video was helpful, give it a like and subscribe for more voice AI content like this.