The code is surprisingly short and compact. It looks as if if the classification could be adapted to extract "highlight" snippets from podcasts as well, which would be a usecase I would be interested in.