Snipd uses AI to “unlock knowledge” in podcasts
Podcasting has emerged as a major billion-dollar industry, with advertising revenue in the US alone expected to reach $2 billion this year – a figure set to double by 2024. Against this backdrop, major players in the podcasting space are strengthening their podcasting arrangements, with Spotify recently dishing out around $85 million to two companies that specialize in podcast measurement and analytics, while Acast recently snapped up Podchaser — an “IMDb for podcasts” that gives advertisers deeper data insights — in a $27 million deal.
But as the major platforms lock horns in the pursuit of podcasting riches, smaller players continue to arrive on the scene with their own ideas on how to advance the podcast medium for creators and consumers alike.
One of those is Snipd, a Swiss startup building a podcast app that uses AI to transcribe content and sync with notes apps, automatically generate book-style “chapters” and, starting this week, deliver podcast highlights in a TikTok-style personalized feed.
Beyond search and subscribe
Like other so-called “podcatcher” apps, Snipd works by users searching for and subscribing to podcasts that are of interest to them – this can be anything from true crime to history and sports. But Snipd strives to be much more than just another podcatcher, in terms of how it analyzes the content of episodes to help listeners curate and get to the heart of the details that matter.
For example, Snipd can create “chapters,” which separate each episode into navigable segments with their own title, while it can also generate transcripts of entire shows.
On top of that, users can manually create “clips” while listening to an episode, allowing them to save their favorite moments and add notes to each clip.
With Snipd’s latest launch, which is available on Android and iOS this week, the company is channeling its inner TikTok by presenting users with a highlight reel of sorts, automatically pulling out what they think are the most memorable moments from a variety of podcasts. It then assigns an AI-generated title to each clip, presenting them in a feed that users can navigate by scrolling up and down.
From there, listeners can save each clip to their library, or—if they like what they hear from the short segment presented by Snipd—skip directly into the full podcast episode.
It’s worth noting that with the app’s latest update, users are now prompted to select their favorite topics (eg “history” or “music”), which Snipd uses to generate these highlights. This means the episode feed is not based clean on users’ podcast subscriptions, as it also pulls in content such as Snipd thinking they will be interested in based on their chosen subject, among other “signals.”
“The goal of the algorithm is to present the user with content they are interested in – for this we use different signals,” Snipd co-founder Kevin Smith explained to TechCrunch. “Whether the user subscribes to a particular program is a strong signal, and therefore much of the displayed content comes from a user’s subscriptions. But there are many other important signals, such as what the user has listened to, highlighted and saved, or what is currently trending among other users.”
While this may be interpreted as a positive move by those looking for help finding new and useful podcasts, it may annoy users who only want to watch content that they have specifically subscribed to. But Snipd eventually plans to give listeners more granular controls over what content appears in their highlights feed, including the ability to filter out clips from podcasts they haven’t specifically subscribed to.
It’s also worth noting that Snipd’s new feed focuses on recently released podcast episodes, particularly those released in the previous two weeks – in the future there are plans to adopt a more YouTube-like approach, when it comes to suggesting older content that Snipd believes is relevant and interesting.
Aside from the new TikTok-inspired highlights feed, Snipd users can still access AI-powered highlights for every episode in their main subscription list, regardless of the current episode.
The app automatically generates highlights for the more popular podcasts, using criteria such as how many of its users subscribe to a show. And for new or less popular podcasts, users can manually “request” Snipd to work its magic, giving them highlights, chapters, transcripts, and everything else within about 20 minutes.
AI at work
But what exactly is Snipd looking for when considering what content to present in the “highlights”? How can it know which segments are more valuable than others? According to Smith, it’s all about how users have interacted with episodes historically – it analyzes what type of content sparks the most interest, and then feeds that data back into the AI training engine.
“Our AI learns by analyzing the content of old episodes, comparing which parts of those episodes were highlighted the most by our users and which parts were not,” Smith said. “The most insightful parts of an episode are often highlighted by our users, while less interesting parts are often skipped and not highlighted. Our AI has learned to use the actual content of the conversation to identify these parts and can recommend them in new episodes.” »
Smith added that Snipd mostly creates its AI models in-house, and specifically for the language models, it starts with large pre-trained models similar to GPT-3, which are already capable of understanding a lot about text and language.
“We then fine-tune these models for our very specific applications,” noted Smith. “Other models we train from scratch. We then use feedback signals from users to improve the models over time.”
Smith said that in the company’s initial findings, users appeared to use highlights to decide which episode to listen to — so they’ll scroll through different clips until they find something that grabs them, then jump into the full episode. The problem is ultimately one of choice overload — similar to how Netflix “suggests” new shows to watch based on subscribers’ viewing habits, presenting a preview of the show on the main menu screen, Snipd tries to help listeners filter through the podcast noise.
“Our users sometimes subscribe to over 100 shows, especially the ones that are very informative, like the ‘Lex Fridman Podcast’ or the ‘Tim Ferriss Show,'” Smith said. “These episodes are up to five hours long. This makes it extremely time-consuming for listeners to discover the parts they are most interested in.”
To unlock knowledge
Some studies suggest that as many as 74% of listeners use podcasts to “learn something new,” compared to 71% who cite “being entertained” as their main motive and 51% who cite relaxation.
And that’s why Snipd’s self-proclaimed mission is to “unlock the knowledge” in podcasts.
“The main problem we’re solving is getting knowledge from podcasts,” Smith explained. “We look at the entire user journey of engaging with knowledge in podcasts and try to improve it. From discovering the best content, consuming it, storing the knowledge the user will remember, to sharing it with friends.”
Before the latest app update, Snipd has mostly been focused on letting users highlight and save specific bits of knowledge they come across so they can refer back to it later. As such, the app is compatible with headphones, so joggers (for example) can triple-click the button on their headphones to create and save a clip with an automatically generated title, summary and transcript. And given how popular podcasts are among drivers, Snipd also recently launched support for Apple’s CarPlay, allowing users to generate podcast highlights while behind the wheel.
Snipd supports the “knowledge unlocking” mission in other ways as well. For example, users can integrate and sync Snipd with read-it-later service Readwise and note-taking app Notion if they want to read segments or transcripts from their podcasts. On top of that, users can manually export Snipd content to Obsidian, Logseq, Bear and Markdown.
Show me the money
Based out of Zurich, Snipd comprises a team of five, including three co-founders and two employees. The first iteration of the app launched last August, and in the intervening months the company raised an “oversubscribed” $700,000 pre-seed round of funding from backers including early-stage Swiss venture capital (VC) firm Wingman Ventures, also as US-based VC Acequia Capital, which has previously invested in billion-dollar companies such as Square, Pinterest and Wish. Smith said Snipd plans to raise a seed round at some point “in the not-too-distant future.”
All of this brings us to a rather important question around economics – how does Snipd make money? The short answer is that Snipd isn’t making money…yet. But in the future, the company plans to adopt a freemium business model à la other similar podcast apps out there, so this could mean a basic free version supported by ads or promoted content, with some of the funky AI-powered smarts pushed behind a paywall.
This also raises questions about how easy it will be to flourish in a market that includes well-established (and well-funded) incumbents such as Apple, Spotify, Acast and Pocket Casts. Snipd’s AI-powered features are certainly neat, but it’s not clear whether Snipd can garner enough of a user base to build a significant business. Additionally, there are already comparable companies out there, such as Moonbeam, a podcast discovery app that combines machine learning and human curation to deliver personalized podcast recommendations. And so are Airr and Fathom.fm, which are equal in helping listeners get more out of their podcasts, either by aiding in discovery or letting them extract the parts they find most interesting.
In truth, Snipd may be an acquisition or acquisition on the way. Spotify, for example, already offers transcriptions for its own original podcasts, and is no stranger to doling out millions of dollars to podcast-focused startups. Amazon has also recently launched podcast transcriptions.
In a busy space, it’s clear that the big podcast players will continue to seek new ways to add value and differentiate themselves from the competition, and helping their subscribers “unlock knowledge” could be another way to do just that on.
“We see podcasts as one of the largest knowledge bases in the world, and are therefore focused on the knowledge-seeking community,” said Smith. “While our competitors treat podcasts as music you listen to from start to finish, we see them as a series of knowledgeable moments.”