Tag: content identification
YouTube's content ID system – copes with volume, but not pitch changes
An enterprising techy in the Computer Science department at Rochester University, NY, has undertaken an investigation to find out exactly how much you have to mangle audio tracks before YouTube’s content identification system can’t recognise them any more. The findings are quite interesting.
For starters, he learnt that despite YouTube still easily recognising when a song’s volume is massively increased or decreased, it can’t cope with any pitch scaling beyond 5% or so. Also, interestingly, it only seems to recognise the first 30 seconds or so of a track.
The full analysis is right here, and there’s some fantastic comments on the white noise test video, here. Also interesting is that despite infringing copyright on the website 35 times and, at once point, 15 times in an hour, the account hasn’t been banned or removed.