working on Meridian, my newsletter that aggregates insights from crypto  podcast, I leverage LLMs regularly as leverage. Sometimes to eliminate things that would otherwise be a pain in the neck to write.

transcribing podcast in generally a starting point, base transcriptions can identify that several people are talking but not who they are. So enrichment is needed

The context is in the conversation, and is an easy task for a human. For examples sake, we can say that introductions occur within the first few minutes of a podcast and the enrichment can occur

We sort of do this heuristic in our head of eliminating who the speaker isn't each time we listen to a new podcast. Since we can give the model a fair bit of context it should be able to knock it all down

I asked claude 3.7 Sonnet to write me a prompt for gpt-4o for the speaker assignment and now I have something I can pass a transcript to ,it samples, and provides predictions of which speakers are which

with that said, feel free to sign up to https://www.getmeridian.io/

The last issue of the newsletter covered 22 podcasts, that is a pretty heavy time investment for each issue. So I decided to see how far I could get with some of the easily accessible models