4 Expert Tips for Building an Accurate AI Meeting Notetaker

Hey everyone, it's Mart here from Assembly. Today I'm sharing four tips for building your AI meeting noteaker. So the first piece of advice that I have is to focus on accuracy. With most meeting note takers, the goal isn't just to transcribe the meeting, but also trigger several post call workflows. So accuracy is really crucial. There are several ways that you can improve accuracy. The first one is to select your model. So if your audio is in English, you can use slam 1 and otherwise you can use universal which supports 99 languages and code switching. We even have a speech models parameter which basically lets you specify like which models you want to use for your request and we'll route your request to the right model depending on the features that you select and the language of your audio file. So once you select the right model, you can also improve accuracy by including a key terms prompt. So in this case, I recommend leveraging metadata that you have about your meeting attendees like their names, the companies they work for, their projects and roles and feeding those into our API as key terms so that it can boost the likelihood of these words being transcribed. So my next piece of advice would be to use speaker labels or multi- channelannel audio. So for a meeting notetaker, it's not just important to know what was said, but we also want to know who said what. That's why at assembly we have a great speaker labels model. If your app has metadata on the number of attendees in the meeting, you can feed that information into our model with the speakers expected parameter. Or if you're just not sure, you can even set a range of possible speakers. If your meeting platform has multi- channelannel audio, then using the multi- channelannel parameter is even better because then you get speaker separation by channels, which means it's 100% accurate. We've also newly released speaker identification which lets you specify the names of the people in the transcript or in the meeting and have those names appear in the transcript instead of like generic speaker labels. So the third tip is to use a hybrid of streaming and async for your note taker. So this is becoming a lot more popular because some users like to keep track of their meetings in real time. It enables users to get caught up on meetings if they're late or for example in a scenario in a meeting where someone loses their train of thought. They have a transcript ready to catch them up. We've also released new languages to our streaming model. So streaming is ready for use cases where there is code switching or if there's a need to support multiple languages. The last tip I have is to use an integration to build out your meeting notetaker. So we have tools like Zoom RTMS or recall which help you build out an early version of your meeting transcriber quickly. These tools speed up your development and the best part is they already work with our API and we have plenty of resources on our docs to help you build it out quickly. So that's all I had and if you have any questions at all, don't hesitate to reach out to our team. All the best with your app and I'll see you again. Bye.