Sync API: Sub-300ms Context-Aware Transcription (Full Transcript)

[00:00:00] Speaker 1: With the Universal 3.5 Pro, we're really excited to introduce the new Sync API. This is a hybrid API that kind of meets in the middle between a real-time API and a pre-recorded API. Think push-to-talk, IVR routing, customized voice agents where you want complete control over the orchestration end-to-end. This is a great HTTP API where you can send in audio between 8 milliseconds and 120 seconds and then get back transcription immediately within 300 milliseconds. It supports prompting, language steering, key terms, and conversational context out of the box. In this demo, I'm going to showcase some of the conversation context features. Let's say that we're transcribing a customer service call. So this is a customized voice agent, you're handling the orchestration end-to-end, and you really just want really fast, accurate transcription. The Sync API is perfect for that. With conversation context, we can add in what the agent is asking of the user to steer the human transcription that is coming back. So let's add what is your birth date to the conversation context. Now I'm going to hold to talk and say my birth date and we'll see the formatting that is applied. 3490. So we can see here the model comes back with 3490 perfectly formatted as a date because it has the context, what is your birth date. We can also see here that we returned the transcription or generated the transcription within 360 milliseconds. So it's very fast and snappy. As soon as you let go of the request, we're going to be processing that and returning right back to you. There's no polling, there's no session management. It is a simple REST API to interact with. Let's try another example in conversation context. What is the expiration date of your card? So let's say this is a reservation or a customer service call and you have to provide information on your card. Now people can pronounce this in different ways. They might say December 12th in the year. They might say 12-28 in the year. And the model will have to infer the right formatting. I'm going to do shorthand and we'll see how the model does. 12-28. So we can see that this came back properly formatted as well. 12-28. We can see the inference turnaround time is under 300 milliseconds at 263 milliseconds. So it's really fast in what we generate. Let's try one more example. So this time we're going to give conversation context. What is your address? So when we say our address, the model should properly format it in a single turn. My address is 1-2-3 East Main Street, New York, New York, 1-1-0-2. So this came back saying my address is 1-2-3 East Main Street, New York, New York, 1-1-0-2. So it didn't split up these entities and it formatted it nicely for the agent. This is really a make or break it for an agent experience. Agents would likely get tripped up, run wrong formatting, have to ask the question over and over again, leading to a frustrated user and likely a poor voice agent experience. With the Sync API, again, you get rapid transcription with the most accurate model that we have on the market, allowing you sub 300 milliseconds turnaround time and complete control over the orchestration for your agent or transcription experience. We can't wait to see what you build with it and we're really looking forward to the feedback. Thanks.