For talks where we don't have detailed "speaking notes" already in the deck itself, we're piping talk videos through whisper.ai and manually popping 'slide break' markers into its timestamped output; that gets rolled into the same deck.json file we build from Keynote.