Fooling YouTube Audio Fingerprints: Speed Works, Volume Doesn’t, and All That Matters Is the First 30 Seconds

YouTube and the music industry have a complicated relationship. Warner Music Group yanked its artists from the site and user uploads that contain tracks from the company’s albums are automatically muted. But rival Universal Music Group recently announced a planned new joint music video venture with the site called Vevo.

The main sticking point in all these deals is (of course!) money, but a big part of figuring out who that money goes to hinges on YouTube’s content identification program. For music, YouTube uses (at least in part) Audible Magic’s audio fingerprinting services.

How good is this audio fingerprinting tool, and how does it work? That’s a secret kept in a black box. But enterprising YouTube user retnirpregnif donated a bunch of his or her time to the greater good to try to figure out what’s going on. The experimenter uploading 82 modified versions, mostly of the same repetitive song, I Know What Boys Like by The Waitresses (which I think would be on Universal Music Group, though that’s not specified).

Here are some of the most interesting results:

• The account was not taken down, despite 35 of the 82 test videos triggering Content ID notices.
• The best way to keep a track up without altering it too much is to increase or decrease the speed of playback by about 5 percent (a slowed down example that didn’t get filtered is embedded above). Pitch alterations of more than 6 percent will also work.
• The fingerprinter may only look at the first 30 seconds of a song. It did not catch an upload of the song when the first 0:30 was muted.
• The fingerprinter is really good at changes in amplification, even when it caused a lot of distortion.

Further reading: Way back in the day, we tested audio fingerprinting ourselves, but on TV shows instead of music tracks. It didn’t do well, to say the least.