Bengt
commented
9 years ago I really like the idea of recommending podcasts by topic. For example, based on the topics that are being talked about in my subscribed podcasts, Podfilter could recommend podcasts that I do not subscribe to. Perhaps there are other views about topics I am interested in that I would usually not consider to hear? This could be really helpful for leaving one's own filter bubble.
I actually built a content-based recommender for (german) YouTube videos for my bachelor's thesis: http://azuvideo.appspot.com/ The system only analyses the text of the video's title and description and matches that against the contents of curricula of different jobs that require training. As can easily be seen, this was not very successful: Many recommended videos have little relevance to the chosen topic. So I would advise against this approach from my personal experience. Making something like this work would certainly require a lot of effort and information retrieval voodoo.
However, content-based filtering of YouTube videos relying on automatic speech-to-text conversion, as presented in the TED talk you linked to, should be very easy "really soon now.". With the complete spoken words of a video transcribed to text, one has a much better representation of the content / topic of a video. Maybe this could be exploited by uploading podcasts as blank videos to YouTube and downloading the automatic captions via the API.
Not so crazy like this: http://www.youtube.com/watch?v=kv_uyUTx5Po, but in the same direction, podfilter.de could analyze: topics, speaker, shownotes, links, textes, comments, whatever, to analyze the content of a podcast. And than build a graph based on that information and compare it with other and rank it by degree of similarity.