Building an award winning AI product in entertainment
This week, the world’s leading AI conference Cog X named Entale as the best AI product in entertainment.
To celebrate we took the opportunity to speak with Entale’s Chief Product Officer Claire Roberts and Data Scientist Olek Główka who have been leading the charge on building Entale’s AI-powered podcast platform, built to transform the listening and discovery experience. They lift the lid on how they design and implement AI technology to enhance the audio experience and the challenges and opportunities that come with it.
How are you using AI to change the podcast experience?
Claire: When we started Entale, we wanted to transform the podcast experience by allowing listeners to see and engage with anything mentioned in a podcast, then find new things to listen to using this context. Imagine you’re listening to a conversation about a recent best-selling book — at Entale we think you should be able to buy the book or save it for later, then find other podcast interviews with the author, or get notified of any more reviews of the book, all from your podcast app.
We started off by building a manual content management system that allowed podcasters to upload their audio, mark off where related content sits and then publish that to the app. The initial feedback was really positive with audiences spending between 30–50% of listening time interacting with this related content, but with over 1 million podcasts available we had to find a way to make this experience scale. To do this we set about building our own AI tech that could recognise the important information in a podcast, link it to the web, and show that to the listener alongside recommendations at the right time.
Can you explain in more detail how you approach the problem?
Olek: My background is in linguistics, speech science and cognitive neuroscience, and a key part of my work involves using natural language processing techniques to design our solution. In practice that means working with podcast metadata and transcriptions to understand the context of a given episode. We create an in-domain data set by tagging podcast descriptions and transcriptions, effectively highlighting key parts of the conversation that listeners might want to find more information about. We train neural networks to understand which entities are interesting within any given podcast episode. We then extract the entities and use various techniques to link those entities to the correct reference. As you can imagine that’s not easy to map, so we mine as much context as possible to hone in on the signal of what entity belongs to what reference in the real world.
Claire: This process generates a lot of powerful data, and we’re now starting to use that to transform podcast discovery. We don’t just have information about the entities themselves but also the relationship those entities have to other things and how they relate to other podcasts. You can start to see how all that sophisticated data starts to create a really powerful discovery experience that’s specific to a user and what they’re interested in. We believe this provides a much richer experience than the current way in which we find new podcasts which is primarily by broad categories such as ‘comedy’ or ‘arts’.
How do you strike the balance between manual human input vs AI?
Claire: From a product perspective it’s always important to be gathering user feedback to make sure we’re building experiences that consumers want to see. More specifically for our AI product we ask our user base to help improve our related content. We have a UGC experience in Beta in the app which allows users to add and correct content, improving our algorithms and contributing to our unique data set. The more we can crowdsource from our users the more we can cater for diverse opinions and experiences.
Olek: From a machine learning perspective we try to automate as much of the process as we can but we have to be able to have some ground truth to prove that our algorithms are working. We need to benchmark our system so that involves an element of human tagging and processing.
What is the most exciting thing you’ve discovered on this journey?
Olek: One of the more exciting things was the recent project we’ve conducted with the Alan Turing Institute, the national institute for data science and AI. We were involved in their data study group which brings together some of the country’s top talent from data science and AI and we worked with them on our unique podcast recommendation pipelines. As a small startup the ability to be able to collaborate was hugely valuable both personally and professionally.
Claire: For me it’s been taking something that nobody had imagined or done before, building it out as a feature, validating it and then pushing it as far as we have done and seeing the opportunities that has opened up. We now have a vast and unique data set and a level of expertise that no other company has with respect to doing this type of work in the audio industry. The RSS feed, which is the open standard through which a podcast is delivered, is more than 20 years old and no one has really reimagined how we create new experiences on top of it, so to be building Entale as a response to that is really exciting.
What are the challenges to building an AI product in this category?
Claire: We hold ourselves to being able to transform any kind of podcast that is published and with over 1 million podcasts now available and new shows being released all the time it’s quite a big challenge to make sure our technology is applicable to them all. On the flip side we also need to be able to present something useful and relevant to all the different types of people who listen so that scale and diverse range of interests and voices can prove challenging.
Olek: We need to make sure we are presenting balanced information and recommendations whilst not showing anything that could be triggering or insensitive to a user. I actually think what we’re doing in the podcast discovery space is a great driver away from echo chambers as we can present a range of recommendations and conversations around the same topic.
How do you make sure what you’re building is presented in an entertaining and seamless way?
Claire: From the very beginning we have always been focused on presenting a more entertaining and engaging podcast experience than anything else on offer. We reflect that in everything we do from the design and aesthetic of the app to our coral colour palette and the vertical feed style player. At times that can be challenging because people have expectations of what a basic podcast app looks like, so to try and reimagine that for them can be complex. Our efforts were recognised when we won a Webby award for Best Visual Design and that’s something we’re really proud of.
What’s coming up next for Entale and what are you most excited about?
Claire: I’m really excited to be using cutting edge AI technologies to build out our discovery solution and see that powering our in-app recommendations as well as new and exciting product offerings. Importantly, we have designed our technology so it can apply to all spoken word audio and prove transformational for other industries and audio platforms.
Olek: I’m looking forward to doing AI that’s more closely related to human behaviour. I love tracking and modelling human behaviour, interactions and interests so to be able to design a discovery experience with that in mind is really exciting; helping people to pursue their guilty pleasures and fall down a rabbit hole of niche discovery.
Try Entale for yourself, download the app on iOS
