Recognizing Music

Yesterday I was flying back from a conference in Banff and picked up a copy of New Scientist during a protracted lay-over. The issue features several articles on the cognitive characteristics of music. The thing I found most interesting was how the researchers attempted to classify the essential properties of music, so they could isolate individual effects in people, animals, etc.Maybe the cleverest thing is what they labled as irrelevant, ie, absolute pitch. They point out that in all cultures which have names for notes, notes which differ by an integer multiple of frequency (one or more octaves higher or lower) have the same names. Many cultures have different choices for semitone relative differences, though. They then point out that rhesus monkeys can recognize two songs which are only shifted by one or two octaves as the same song, but sequences shifted by any other amount are not recognized. This makes sense, considering that most things which produce what we call single notes produce some amount of harmonics. Identifying notes an octave apart seems almost unavoidable.

They don’t discuss how people identify two pieces of music as being the same, other than to say that if people have a condition called Congenital Amusia, they cannot recognize “familiar” songs without lyrics. If there is static, a different singer, or an arrangement with some differences, most people can tell that two songs are the same. Specifically, I want to know how a computer or a courtroom, can decide how two songs are the same. The reader should recall the VH1 behind the music where Vanilla Ice sang the difference between the bassline in “Under Pressure” by Queen/Bowie, right before singing the hook in “Ice Ice Baby” There is a difference, note for note. You can hear that they are different, but the court decided that they were very, very similar.

So I was avoiding my thesis a few weeks ago by using a MusicBrainz client to label my mp3’s. It works pretty well, but it tried to identify two different takes of a song (I forgot which one, but I think it was Yardbird Suite) as the same. I’m not sure if this is from users mislabeling two distinct songs with the same name, or if they really have similar signatures in the classifier. The MusicBrainz database doesn’t just identify two songs as the same if the files are the same, it computes a “fingerprint”, which is a number that hopefully encodes the “important” parts of the music. There are many fingerprint possibilities, but MusicBrainz currently uses TRM. Looking at their website, they apparently use the MusicBrainz database, which is user generated, as a fee service for royalty enforcement for internet media. Wow. No technical real info on their webpage, sadly. I’ll do some emailing and see what I can dig up. How far are we from software being able to identify the song Yellow, whether it be done by Coldplay or Richard Cheese?