PodFeed White Paper
PodFeed is a dynamic recommendation feed, like TikTok or Insta Reels, showcasing engaging podcast snippets. Powered by advanced LLMs, these snippets are indexed and can be enjoyed independently of the full podcast, offering a fresh, convenient way to discover and consume podcast content.
The podcast market is currently series-driven, requiring listeners to sift through episodes they might not find interesting. Most podcast apps, like Apple Podcasts and Spotify, primarily offer subscription options, leaving users to manually filter relevant episodes based on descriptions or topics. This becomes particularly challenging for podcasts that cover diverse subjects across different episodes or even within a single episode. While some genres, like True Crime, focus on specific topics and are best consumed in their entirety, the overall experience often lacks efficient content discovery.
The podcast market is heavily personality-driven, with hosts like Ezra Klein and Joe Rogan inviting guests to discuss a wide range of topics for anywhere from 30 minutes to 4 hours. Given the limited time with their guests, hosts often cover numerous current topics, creating a broad but dense content experience. While this approach is efficient for hosts, it places the burden on listeners to curate the content themselves.
For example, when Joe Rogan hosts Elon Musk, they might discuss Tesla, SpaceX, COVID vaccines, Twitter, Neuralink, and political views on Biden within a 60-minute podcast. While this creates engaging content for the host, listeners may only be interested in specific segments, such as Tesla and SpaceX, but not Musk's views on Twitter or Biden. This makes the content discovery process cumbersome for the user.
Currently, users cannot easily separate podcast segments of interest from those they find irrelevant. While fast-forwarding to the desired content is an option, it is cumbersome and relies on the host providing a detailed index of topics discussed. This approach is often impractical, especially for listeners on the go, such as in a car. As a result, users are forced to listen to entire sections they are not interested in, decreasing overall engagement. Consequently, many listeners only subscribe to podcasts where the "hit-rate" of interesting content is higher, limiting their exposure to a broader range of topics and shows.
LLMs to the rescue! Today's advanced language models can "index" a podcast by analyzing the audio and creating distinctly different snippets. These snippets can be consumed independently of the original podcast, making content discovery and consumption more efficient and user-friendly.
There is a compelling case for creating and distributing podcast snippets independently in the form of a recommendation feed. While technically feasible, this approach would likely violate copyright laws, as podcast owners control the distribution of their content. Although obtaining rights from each owner is possible, it presents an impractically high barrier to entry.
Another challenge in creating a recommendation feed of podcast snippets is the sheer volume of podcasts that need to be indexed and analyzed. Starting with a top-100 or top-1000 list would still miss many niche podcasts that users listen to daily. Additionally, many users subscribe to paid podcasts, such as Stratechery or Substack creators, which cannot be indexed as they are behind paywalls. This limitation further restricts the comprehensiveness and appeal of the feed.
Users are already subscribed to a variety of podcasts, including paid ones. Podcast players typically subscribe to RSS feeds and download relevant episodes for users to listen to, either locally or streamed from the server. PodFeed would follow a similar approach with a few key differences:
This approach is similar to how users receive content from those they "follow" on Facebook or Instagram. By focusing on podcasts users are already subscribed to, PodFeed significantly reduces the initial corpus that needs to be analyzed. Additionally, PodFeed can enrich the recommendation feed with snippets from podcasts the user is not subscribed to, facilitating the discovery of new content. This method also addresses copyright concerns, as PodFeed acts as the user's agent, downloading full podcasts and creating snippets from them.