Tronds lyttearkiv — automatic content reports from the podcasts I listen to
June 28, 2026 · 3 min read
I listen to a lot of podcasts — popular science, heavier research, politics, silly comedy shows, audiobooks and much more. And every now and then I hear something that doesn't quite add up. A claim I'm not sure I buy, or an explanation that feels a little too neat.
I used to grab my phone, pause the episode, and google my way to whether it was true. Now Tronds lyttearkiv ("Trond's listening archive") does it for me — automatically, for every episode I run through it.
What it is
The archive produces structured content reports from podcast episodes. For each episode you get:
- A summary of what the episode was about
- Verbatim quotes of what was actually said
- Real sources for the claims — studies, articles, research
- Critical peer reviews that challenge and nuance the content
- Timestamp links that open the episode at the source at exactly that second
It's like having a curious, source-critical listening buddy who takes notes for you — and who actually checks the things that don't add up.
Right now it holds reports from NRK's science programs Abels tårn and Forskningsfronten, but the setup isn't tied to any particular publisher — it works on podcasts in general, and the archive grows as I feed in more episodes.
Why I built it
The spark was exactly that feeling of something being "off". I've always enjoyed digging into the sources when a claim sounds a little too confident, but it's awkward to do mid-listen. The whole point of the archive is that I now get both the criticism and the sources served up automatically. Instead of taking things at face value — or turning detective myself — I can read the report afterwards and see what holds up.
How it's built
The pipeline goes from raw audio file to finished web page in a few steps:
1. Transcription. The episode is run through Faster-Whisper-XXL, which both transcribes and does diarization — "who says what, when". The result is a timestamped transcript of the whole conversation.
2. Analysis with Claude. The transcript is fed to Claude (Anthropic's language model), which reads through everything and builds a structured representation of the episode — a kind of DSL (a small "article language") describing themes, summaries, verbatim quotes, source references and the critical peer reviews. This is where the heavy lifting happens: understanding what was said, finding what's worth highlighting, and assessing it critically.
3. Timestamp links. Because Whisper provides time codes for everything, each quote can be linked to the right second in the audio. Click a quote and you jump straight to that spot in the episode at the source.
4. Generation and publishing. Finally, an HTML page is generated per episode from the DSL. Those static HTML files are what's hosted on Vercel.
A note on quality
The quotes are verbatim — pulled directly from the transcript, not paraphrased. The sources and peer reviews, on the other hand, are generated by a language model, so they should be read as a good starting point for further checking rather than gospel. But that's exactly the point: I want something that points me in the right direction when something feels off, so I don't have to start from scratch.
This is a hobby project, and it isn't affiliated with the podcast publishers.
Have a look and try it yourself: lyttearkiv.itrond.net