AssemblyAI’s Sync API delivers transcripts in one request (Full Transcript)

[00:00:00] Speaker 1: So you might have already seen my tweet. When I joined Assembly AI almost four years ago, one of the first things that kind of tripped me up was polling for your transcript. And I remember thinking like, surely this is way more complicated than it has to be. Transcription should really only just be one request. So it turns out that instinct was right, it was just a little early. Four years ago, the models weren't fast enough to transcribe a file that quickly without giving up accuracy. And our models have come a long way since then. And it's finally here, because we just released our sync API. It's powered by Universal 3.5 Pro. And it lets you send a short audio file and get back an accurate transcription in the same request. And no polling, no WebSocket, no lambda function, no webhook, just one request in and your transcript out. Let me show you why this is such a big deal, because the difference is really day and night. So this is what our async API lifecycle looks like today. And it's the right tool for a lot of jobs. But just look at how many steps this is. First, you post your audio, and then you get back the audio URL, and then you post the transcript. And then you check the status, you see that it's processing, and then you just loop every few seconds until the status is completed, and you retrieve your transcript text. And the subtle thing here is that once it's completed on our end, you don't see that it's complete until you've made your pull request, which happens every few seconds, which adds latency to your request. So there's this dead time baked into this lifecycle, which means that while it's going much faster, it's not as fast as it could be. With the sync API, we've collapsed all of these requests into one single request. The moment the user starts recording the audio, you can fire off a pre-warm request to kind of pre-establish the connection. So the DNS lookup, the TCP handshake, the TLS handshake, all of that happens while the audio is still being captured, completely off the critical path of retrieving your transcript. By the time the recording stops, the connection is already warm and ready, so there's no handshake left to pay for when you make your post request. You just post your audio bytes straight to the slash transcribe endpoint, and you get the text right back in the response. No upload, no queue, no polling loop. And where this really shines is that the middle ground between our other APIs. The async is fire and forget batch processing, whereas streaming is for a live continuous audio stream. Sync is for when you've got a discrete short clip up to 120 seconds, and you need the transcript back immediately. So we're thinking of real-time voice dictation apps, voice agents, live meeting notes, instant search indexing, customer support agents, anywhere where you need the transcript fast and on a short snippet. To show you the latency, I've built a little demo here. It's just a raw HTML file, and you can grab it from the GitHub gist in the description. Let's record something, maybe about 10 seconds long, and you can see the full life cycle. Hey guys, it's Mart here. We just released our sync API, and I'm so excited to give it a try. A few things about me are that I've worked at assembly for four years, and day to day I do a lot of marketing work here at assembly. Right, so that was 15 seconds, and you can see that we get the response back in 1900 milliseconds. Now for full transparency, I'm in Singapore right now, which is basically as far away from our servers as you can get, and I'm hitting our global endpoint, which means this request should route me to the EU region, but it's still well under two seconds end-to-end for the request. And if you look at the breakdown here, the interesting part is what's not there. There's no connection set up on the clock because it warmed while we were still talking. And so all that's left is the upload step and the model inference time. So when running this out of the US or in the EU, that P50 latency drops to around 134 milliseconds. That's sync API in a nutshell. It's that same accuracy that you already trust from Universal 3.5 Pro, but instead of upload, submit, pull, pull, pull, retrieve, it's one request. Post audio bytes and you get back your text. So give it a try, let me know what you think. I'm very excited to see what you build with this. Bye.