Nathan Gathright@marcoarment just announced on @atpfm that @overcastfm is shipping podcast transcripts for every public podcast with more than one listener. No subscription required. The story of how he pulled it off is genuinely absurd.
Apple Podcasts launched transcripts ~2 years ago using massive cloud infrastructure. Cloud transcription APIs would cost Overcast thousands of dollars per day. OpenAI's Whisper is too slow to scale. So what does a solo developer do?
Last summer Apple released an on-device speech transcription API in the iOS 26 beta. It's the Siri dictation model, and it runs locally on Apple silicon. On a single M4 Mac Mini, it could transcribe audio at 200x real-time. 200 minutes of audio per 1 minute of wall clock time.
So he started buying Mac Minis. First two, then a third for EU podcasts. Then more during sales at $450-500 a pop. He eventually called a data center on Long Island to ask about racking a few units. They told him the minimum was basically a full cabinet, 48 rack units.
By the time he'd accumulated 18, a full cabinet started to make economic sense. He signed a contract and had to face the challenges that come with the reality that most data centers simply don't want to deal with one guy renting three rack units.
Final count: 48 Mac Minis. All base model M4s, 16GB RAM. Some in the data center, some at Mac Mini Vault, some at a host in the Netherlands, a few still at his house. Total recurring cost for the whole operation: ~$1,000/month. Cloud APIs would have been thousands per day.
At one point the queue wasn't keeping up despite 48 machines churning away. Turned out a bug was preventing completed jobs from being deleted, so they kept getting re-processed. Fixed the bug, queue dropped to zero in hours. He had way more capacity than he needed.
So he expanded scope. Overcast now transcribes private/membership podcasts too. Built deduplication so one transcript serves all subscribers getting the same audio. Also working backward through archives. Podcasts with 100+ listeners are getting full back catalogs done.
The trickiest problem was dynamic ad insertion. Different listeners get different ads, so the audio doesn't match. The solution was audio signatures that fingerprint the non-ad content, then align the server transcript to whatever version your phone downloaded.
Transcripts are stored on Cloudflare R2 in a custom compressed format with word-level timing and confidence scores. The UI only highlights at the paragraph level for now, but the granular data is there for future features like clip sharing and word-by-word highlighting.
What's next: transcript search, AI-generated chapters, topic summaries, LLM post-processing to fix transcription errors, and someday, phones may be fast enough to just do it all on-device during overnight charging. The Swift code already runs on both server and phone.
Marco's take: once you're used to navigating a podcast with transcripts, using seek forward/back buttons feels like being a dinosaur. Transcripts are table stakes now.