Basically, they start with the durations between each note, normalize for differences in tempo, and then create a string that for each note indicates whether the duration goes up, down, or stays the same between each note. Then find the edit distance between that string, and the strings in the database for each song.
You might have to read that a dozen or so times to make any sense of it, but it beats reading the paper if you have work to get done.
This simply can't work for anything other than very simple tunes. Rhythm is less than half of the picture in any modern piece of music, and there are a few basic patterns that repeat across hundreds or thousands of songs.
Nice idea. Shame it doesn't work. Something where users could sing or hum the tune could actually work though. A true 'musical' search engine. I had this idea years ago but I think the implementation would be really hard.
I've heard this idea from multiple places, so I expect it's one of those ideas that is easy to think of but difficult to implement. A similar idea was expressed in the plot of The Music Man, where children are taught to play instruments by imagining that they are whistling.
I wouldn't be surprised if the main implementation problem is that many people are naturally off-key when whistling or humming, and those on-key know the notes well enough that it is more efficient to simply enter music via a piano-like keyboard.
Aha, it's the Eliza trick: if you can look smart for the first fifteen seconds, many people will conclude that you're smart.
One wonders what percentage of people pick one of those two songs to try first. (I'm guessing that several other popular choices, like, say, "America, the Beautiful" and "Twinkle, Twinkle, Little Star", are also pretty easy for the program to pick out.)
Saw this years back, tried it again just now to see if they got it to work by now.
It thought the VU's "Sunday Morning" was Guardiani Del Destino's "Rhapsody" (and their "Rock & Roll" was Sonata Arctica's "Weballergy"), the head of Coltrane's "Giant Steps" was Shakira's "Whenever Wherevr", CSN's "Suite: Judy Blue Eyes" was "A New York Fairytale", Jethro Tull's "Aqualung" was "If Ever I Would Leave You", The Clash's "Atom Tan" was Mental as Anything's "Live it Up", and The Police's "De Do Do Do, De Da Da Da" was Toquinho's "Aquarela". FAIL.
Basically, they start with the durations between each note, normalize for differences in tempo, and then create a string that for each note indicates whether the duration goes up, down, or stays the same between each note. Then find the edit distance between that string, and the strings in the database for each song.
You might have to read that a dozen or so times to make any sense of it, but it beats reading the paper if you have work to get done.
If you want to look into it further, http://cgi.sfu.ca/~gpeters/tapper/tapper.cgi .