- It uses Whisper to transcribe and label speakers from your audio with timestamps - Formats the transcription in Markdown for the show notes - Creates chapters with timestamps for each topic using Claude-100k - Generates title ideas as well as tweets with both Claude and GPT-3.5; I've found Claude to be better, but not everyone has access to it.
You can find the code on Github [0]. We've been using this for all of our latest episodes to see a "demo", you can read the latest one with Tri Dao of FlashAttention here [1].
[0] https://github.com/FanaHOVA/podcast-summarizer [1] https://www.latent.space/p/flashattention#details