In the Accidental Tech Podcast #683, Marco Arment discussed a year-long project to add transcripts to his Overcast podcast app. This is a beta release, but he anticipates general availability fairly shortly. That chapter of the ATP #683 is long and detailed – over an hour of discussion. To those who would like to read along – or just read without sound – I have included a transcript of that chapter of the episode here.
The TL;DR summary: Marco has implemented server-side code to Apple’s transcription API. On the smallest configuration of M4 Mac Mini, it generates transcripts at about 100x of real-time audio speed. Marco is hosting a full rack of 36 Mac Minis in a Long Island data center. He has another handful of Mac Minis hosted at his two homes and his Long Island restaurant. In total, there are 48 minimum-config Mac Minis crunching transcripts.
How many transcripts can those machines crunch? Plenty. Quoting from the transcript:
And what this has enabled is Overcast is now transcribing every podcast that has more than one listener. Every single one. Well, there are language restrictions, because I’m using Apple’s API or Apple’s models for this. Their models only support English, French, German, Japanese, Italian, Portuguese, and there’s one more — oh no, that’s it, that’s six. Yeah. If I actually look at what Overcast’s top languages are, the one I really want is Dutch — they don’t support Dutch yet — and that is one of Overcast’s top languages in terms of podcasts being produced that people are listening to. But otherwise, like that covers most of what most people are listening to.
Marco grappled with the problem of dynamic advertising: ads that may be inserted into podcast on a per-download basis. With the help of AIs, he has created algorithms to generate a signature of portions of each podcast. He uses these to synchronize the transcripts with the fixed and variable portions of each download. I do not completely understand what he has done. If you’re curious, search on “signature” in the transcript and read the following paragraphs.
How are these transcripts used? The UI will be limited at first, but even this first implementation is highly functional. Again from the transcript:
There is so much that now follows from this, but this has taken me almost a year and I wanted to get this — like I gotta ship something. I gotta get this out there. And even this, like there were a couple of days last week where I broke the build on my own phone. I broke its ability to use transcripts. So for a couple of days, after I had been using them for a while, for a couple of days I couldn’t use them and I hated it. Like I noticed immediately, because what I do is I swipe over the transcript view, like whenever I’m listening to a big show that has stupid DAI ads, like I’ll hear, great, another ad for the Apple card, awesome. Oh, another ad for Pure Leaf iced tea, okay. And I can go over the transcript and I can skim, skim, skim and just tap right after because it’s very obvious where the ad ends. That’s great. Like features like that, you get used to that.
Or even just like, what did they say? If you’re listening to a podcast and a bus drives by and youmiss a couple of words, you don’t have to seek back and repeat the whole 20 or 30 seconds, you can just swipe swipe swipe and see, oh that’s what they said, okay. Like stuff like that is really great for that kind of thing. It’s also really great for like, you know, if there’s a podcast and you know somewhere in here they mentioned something like this, you can quickly skim around and find it. Or if it’s like, oh, I’ve got to listen to this episode before tonight but I don’t have that much time, but there’s information in it that I would like to know generally what they talked about. You can skim the text of what they are talking about faster than you could hear it. So if they’re talking about baseball, you can skip it. You know?
Like there’s all sorts of little benefits like that. You know, much of the benefits of chapters in podcasts that do it, but just applied to all podcasts. It’s a great experience and it is very clear to me having developed this and now using it: it is extremely clear to me that this feature is table stakes, that all podcast apps need transcripts. And once you are used to navigating a podcast with transcripts, using seek forward and seek back buttons feels like you’re a dinosaur. So it is very obvious to me that podcast listening and podcast navigation require this. For any app that wants to be like a serious podcast experience, to be a good listening experience, you need robust transcript features now. And that will only get more so over time as we kind of evolve our UIs and our playback experiences to use them more and to offer more utility that’s powered by transcripts.
Transcript-assisted podcast navigation sounds like the kind of feature that Alex Lindsay will eat up. 
Will normal users adopt this technology, or will it be relegated to high-information users? How much time will it take Apple itself to have its podcast app transcript-aware? Will Overcast – or some other app – become the TiVo of podcasts?
If advertisements can be signature-analyzed and trivially skipped, what is the future of advertising in podcasts?
Marco’s year-long project is totally awesome. With 48 little-but-mighty M4 Mac Mini servers, he’s able to transcribe any public or private podcast that has been published. Apple’s transcription API currently supports six spoken languages. When Apple’s API expands Dutch and other languages, he’ll rapidly include them. Marco’s pioneering of a transcript-aware Overcast app that understands dynamic ad insertion (DAI) is also totally awesome. For podcast enthusiasts, these two things should totally change the experience of listening to podcasts. It’s somewhat unbelievable that one person was able to do all of this work. Transcribing essentially all podcasts and a transcript-aware player is definitely a disruptive technology.