i'd think it wouldn't be too hard to use something like opencv to detect frames that don't have any/much change between them and then correlate that with the audio detection to figure out what can be safely culled like in that fashion.