Speaker 1: If you're a YouTuber, a podcaster, or if you're a fan of YouTube videos and podcasts, you've got to check this out. We recently used AI to build this demo link in the description that transcribes any YouTube video, summarizes it, detects the major topics discussed and more. Today, we'll discuss how you can use it, how it works, and how you can build something like it. Ready? Let's go. Alright, so first things first, here's the demo. Follow the link in the description and you'll see this page. Here, you'll enter the link to any YouTube video you want, whether it's a podcast, a sketch, an interview, a tutorial, or anything else. Today, I'm going to be using an episode of Zach Galifianakis' series, Between Two Ferns. So once I enter that link, I'm going to be directed to this page. This page asks me what I want the AI to do. So smart format up here tells the AI not only to transcribe the video, but also to make that transcription pretty, you know, by adding proper punctuation, knowing when to use numerals to represent numbers instead of words, and so on and so forth. I'll check that box. And now I'm also going to check the summarization box over here so that the AI not only transcribes our video, but also summarizes it, like a spark notes or a cliff notes synopsis. Up next, I'm going to click on the topic detection and entity detection boxes so that we can get a list of the major subject discussed, as well as a list of the people, places, and data discussed. Utterances and paragraphs is just another additional layer of formatting to make the transcription pretty. And I'll also turn on language detection just to demonstrate it. This will identify the dominant language spoken in the video, which is particularly useful if you're doing research and you stumble across a video in a foreign language that you just can't pinpoint. Our AI models will be able to tell you exactly what language is being spoken. Finally, I'll check the diarization box. By checking this box, I'm going to tell the AI to label every word with its speaker, that way we know who said what, you know, instead of a transcript that looks like this, we'll get a transcript that looks like this, but all right. Enough talking about the features. Let's see them in action. Now that I've checked every box, I just click get results down here and boom. Results. The first thing we'll see down here is a speaker labeled transcript. So we know who said what. Here's speaker zero is Zach Galifianakis and speaker one is David Letterman.

Speaker 2: Okay. I don't want to get sentimental, but I have to tell you back when I was a kid, I used to, I used to stay up really, really late and, um, good. Thanks. And I was just watching, I would just watch the color bars and the national anthem.

Speaker 1: As you can see, the AI transcribed the whole video, but that's not all. My favorite feature here is the summarization. Our new summarization feature is a DSL app, also known as a domain-specific language model. Basically in the same way that doctors choose specializations, AI can specialize as well. And this AI specializes in summarizing. We built, train, and fine tune this model in-house on a dataset that we gathered and labeled ourselves. And the result is an AI that can listen to any audio, or in this case, watch YouTube video, and then write a synopsis. Here for this example, our AI wrote the following summary. And for comparison, here's a summary that I wrote of the video. Note that the AI summary and my own summary have some pretty close similarities, even some word for word phrases in there. Up next, we have topic detection and entity detection. Long story short, the AI tried to the best of its ability to write a bullet point list of the major subjects discussed in the video. Meanwhile, entity detection extracts the people, places, names, and so on that were discussed as well. For example, the name Steve Jobs was mentioned.

Speaker 2: Did you just wake up from a 15-year nap? You look like Steve Jobs now. Okay.

Speaker 1: And so was Santa Claus.

Speaker 2: Hi, welcome to another edition of Between Two Ferns. I'm your host, Zach Alphanakis. My guest today is Santa Claus with an eating disorder. Thank you very much for inviting me. I appreciate it.

Speaker 1: Finally, the language detected was EN, which is computer speak for English. And if you'd like to see the request that was sent to the server that hosts the AI, you can click on this tab. Now, it's important to discuss a couple of limitations of these AI models. Much like how chat GPT hallucinates or how AI image generators have a tough time with fingers, sometimes transcription models mishear words in the same way that humans do. For example, our model had trouble transcribing Zach Alphanakis' name, as we can see here in the transcription and here in the entity detection tab. But I'll give the AI a little bit of leeway here, since even most humans can't transcribe Alphanakis' last name upon their first attempt. Now, if you're curious as to how this demo works, you can click up here to see the code. A good chunk of it is front end work, but the real meat is in the code that calls the AI. You can find that here. Long story short, all of these features, summarization, language detection, entity detection, diarization, topic detection, and so forth are all available in the DeepGrim API out of the box, seriously. If you look in the code, the most complicated part of building this app is creating the front end components and buttons. Meanwhile, all it takes to utilize an AI is a single function call here, that's it. This is the line of code that tells the DeepGrim AI to parse the audio. The only parameters you need are one, the audio you want to transcribe and analyze, and two, a list of the buttons that the user clicked, like whether or not they want the model to do summarization. That's how simple it is to use DeepGrim's AI. All you need to do is sign up to create an API key. And from that point on, it's just one line of code, two parameters, and a partridge in a pear tree. As in, all you need to do is sign up, create an API key, and you'll be able to apply transcription, diarization, summarization, topic detection, and much more to any audio you desire, whether that's your podcast or your videos, or even recorded Zoom meetings from work, old voicemails, and so forth. Whether you're an individual or an enterprise, you wouldn't be the first to use DeepGrim to make the most of your audio data. Our users have built entire websites with their voices, driven cars with their voice, created live, real-life subtitles to wear, and even hooked up our AI to Stable Diffusion to create art, like the art that you're currently seeing on screen. Not to mention, if you want to create live transcriptions in real time, you can do that as well, whether you want to record yourself live, or if you want to tap into a live radio feed like we did here with BBC. But long story short, whether you're a content creator or an avid content consumer, like me, this demo is linked in the description. And if you want to build anything with our specialized audio expert AI models, sign up for DeepGrim and you'll receive up to 45,000 minutes of free transcription. That's $200 worth, or 750 hours, or 31 days straight worth of audio. We'd love to see what demos you come up with. Now, if you liked this video, leave a like down below. If you have any questions, feel free to chat with us in the comments section. And as always, follow DeepGrim for more AI content.