Today’s tech-driven audience is in full control of what content they want to consume when they want to consume it, and how. Things got even more interesting during and after the pandemic when it became absolutely critical to provide customers with a supreme personalized experience that would perfectly cater to their unique needs. And next-generation technologies like AI, Machine Learning, Natural Language Processing, and Computational creativity have turned traditional media and entertainment providers on their heads.
Here’s the fact: Media and Entertainment companies can no longer create one-size-fits-all solutions. Providing customers with a highly personalized experience has become imperative. By using technologies like Machine Learning, companies can identify content topics from live streams such as podcasts in a significantly shorter time, with more precision and bigger efficiency. With so much content being produced daily, it is very tiring for people to spend hours searching for content that would match their needs. Luckily, Natural Language Processing can filter unstructured and structured data and provide users with just the right segments instantly.
This enables users to search for those important micro-moments that add value to their own personalized viewing experience. This also opens a world of opportunities for advertising companies to provide their audience with the right content and drive deeper engagement with them by matching the right ad with their demographic.
There is a huge gap in the podcasting space that prevents growth and monetization of the medium — topics people want are deeply buried inside long listening experiences. One of our clients recognized that gap and saw it as an opportunity to build a solution that would perfectly cater to users’ needs by providing them with just the right audio content they are truly interested to hear.
The client approached us with an idea to create an AI-powered solution that would allow users to find relevant topics in podcasts much more easily without having to spend too much time doing it manually and listening to the entire podcast.
We helped them build an algorithm, a Neural Network for segmentation, that used machine learning to analyze podcast content, identify topics discussed in the show, and surface those topical hooks as search results. At its core, the solution cuts podcasts into smaller audio segments and enables users to easily find only the topics they want to listen to. Now, the user does not have to search through the entire episode manually, because the algorithm makes the entire process more efficient and provides users with just what they need in seconds. For instance, a traveling podcast covering the topic of green energy may include a talk about different solutions that would make the world we live in more sustainable and environmentally friendly. If a listener wants to hear more about the digital plants of the future, this AI-powered algorithm can help them find that particular (otherwise hidden) topic in the podcast right away. These micro-segments empower listeners to focus on their areas of interest while in parallel providing a platform for advertisers to position their content closer to these interest areas.
On top of this, we created a web platform where users can participate in building a world of audio segments using the algorithm. This allowed them to foster a social discovery environment and encourage viral listening.
labeledThe core of the project was scientific research rather than a straightforward implementation, as we explored different architectural solutions for Neural Networks to find out the best ones to help us build our algorithm.
When building this kind of algorithm, it is critical to have a good database and a good label to train and test the algorithm. The higher the quality of data, the better the algorithm. To achieve that, our team spent almost a year exploring different podcasts the client selected which provided us with enough information to start building a tool for labeling data. We created the tool and built a huge database for algorithm training fueled by natural language processing technology. We ended up with a core algorithm that beat most of the competitors on the market. The so-called “Schreder” was almost 80 % successful in cutting and selecting segments based on the database that we created.
The next step was to create a more artistic image representation of those audio segments (the topics people want to listen to with all the keywords). We decided to create an algorithm pipeline with a full episode or a series of many episodes on the input and audio signal with an image representation of the topic, the speakers in the podcast, keywords, and the summary of the topic discussed on the output. We selected a few algorithms and implemented the pipeline. We finally built a fully automated audio segment with a visual representation that listeners could listen to with a click of a button. From a UX perspective, each audio segment has a visual representation (collage and cinematography) supported by our NLP and AI logic.
While listeners/viewers are enabled to go faster through multiple podcast episodes, consuming curated content based on their preference, the content creators are empowered by reaching more relevant audiences with the right material. The synergy created via this new approach signals a new direction that can disrupt the existing advertising model across the global media industry.
Many thanks to Miljan Milanovic for sharing these helpful insights. Stay tuned for more interesting stories! in the meantime, visit the industries page and explore how we can help you build next-generation solutions that deliver data-powered experiences to customers worldwide.