Analyzing YouTube's Audio Fingerprinter

Al Benedetto writes "I stumbled across this article which analyzes the YouTube audio content identification system in-depth. Apparently, since YouTube's system has no transparency, the behaviors had to be determined based on dozens of trial-and-error video uploads. The author tries things like speed/pitch adjustment, the addition of background noise, as well as other audio tweaks to determine exactly what you'd need to adjust before the fingerprinter started mis-identifying material. From the article: 'When I muted the beginning of the song up until 0:30 (leaving the rest to play) the fingerprinter missed it. When I kept the beginning up until 0:30 and muted everything from 0:30 to the end, the fingerprinter caught it. That indicates that the content database only knows about something in the first 30 seconds of the song. As long as you cut that part off, you can theoretically use the remainder of the song without being detected. I don't know if all samples in the content database suffer from similar weaknesses, but it's something that merits further research.'"

Comments

pHash

[b1ng0]

This seems like a good time to pump my own open source project: pHash. pHash is a perceptual hashing library that computes hashes for audio, video and image files, with text and PDF hashing coming soon. We use an algorithm similar to YouTube's audio fingerprinting method but we do not only take into account the first 30 seconds. Although, it's impossible to tell from this basic test whether their algorithm truly only looks at the first 30 seconds, or if the algorithm considers them to be different audio files. If the song is only 1 minute in duration, and 30 seconds is blank, is that really the same audio file as the full 1 minute version? At some point the audio files are not really the same anymore, although the perceptual hashes should be somewhat close to each other. Please give pHash a try. We could use some feedback from the OSS community and would appreciate it greatly.