I like this concept, but you need some sort of automated filtering. My first track (because hypem liked it) is a cat purring.

Plug but it's a useful one: run the tracks through the SCAnalyzer http://labs.echonest.com/SCAnalyzer/index.html and filter for speechiness and duration. If you get a SO ID, you'll also know the artist via fingerprinting and you can do more filtering at that level.