Hey Jeff, interesting article. I do not know how youtube specifically does this, but a standard technique for this sort of thing is to treat points in the video/audio sample as points in a several-hundred-dimension vector space. You then need a way to rapidly identify when any sample point is close to any of the millions of data points in that vector space. The search algorithm needs to be optimized for rapidly finding matches without computing the distance metric millions or billions of times, particularly since the query is likely to be "junk", ie, no match. If you or your readers are interested in a gentle introduction to some of the math involved in this kind of search, I did a series of blog entries on it a few years ago. http://blogs.msdn.com/b/ericlippert/archive/tags/high+dimensional+spaces/ Or, see the paper I based this series on for a less gentle introduction: http://research.microsoft.com/en-us/um/people/jplatt/bitVectors.pdf