Digital music has changed the world.  It has never been easier to feed your music addiction. 

I personally will never understand why people have bought over 3 billion tracks at iTunes when they can get a much better deal using subscription services like Yahoo! Music Unlimited, which cost less per month than the price of a single drink at a club and allow you to download or play on demand over 2 million tunes.  You may occasionally find me at the bar, getting obstreperous about this subject.  I can usually convince two or three people to try it for every one that I actually get into a fight with.  In the end, I think the problem is that most people don’t know that music subscription services exist or how they work.  There’s also the nagging issue that many of these services suffer from some annoying technical glitches, which can often require a bit of work to get around.  This will hopefully go away once all the major labels agree to drop their requirement that subscription services only offer DRM encoded tracks.

With all this music available though, it’s easy to get lost in the music.  You’ll never be able to actually listen to all of those 2 million tunes you have access to.  There simply isn’t enough time in the day.  So how do you choose what to listen to?  Subscription services can typically deliver about 75% of what you already know you like, but I much prefer finding new music I’ve never heard before.  It’s refreshing and makes you feel good, like sipping Erdinger in the sun on Venice Beach or Leffe au Saint-Michel.

So how do you discover new music in a world of too much choice? I used to really like Webjay for this sort of thing, but it’s gone now.  One of the best ways to find that more eclectic 25% is by surfing music blogs.  Sites like Hype Machine or Musiclibre let you explore the mp3 blog world and find new stuff accordingly - they are a great place to start, but there is much opportunity for cool new discovery tools to be built.

To this end, I had the pleasure of working with DSP guru Malcolm Slaney on this problem last year.  We used signal analysis to analyze web media tracks that otherwise had very little metadata information.  Sadly this is the norm for most audio tracks on the web.  When websites like Yahoo! or Google go out and crawl the web looking for audio files, they often return very little relevant info about the content of the audio files themselves.  We analyzed over 25,000 popular mp3 files and found less than 20% had data in their ID3 tags.  To make matters worse, the actual contents of the ID3 tags were often incomplete or incorrect.  This means that searching for web media using traditional metadata techniques, or traditional web search techniques (ie. the status quo) ignores the vast majority of audio files on the web.

We used MARSYAS to help us analyze our corpus and projected the results into a multidimensional audio feature space based upon the acoustic properties of each track.  Our demo page lets you browse this feature space to find music which is acoustically similar.  Using this technique, we can suggest music that would otherwise never be returned from any existing audio search engine - yet sounds similar to music you enjoy and want to listen to.  Use it to select a seed song which you like, and check out the acoustically similar tunes it recommends: 

http://www.musiclibre.org/research/playlist.php

Popularity: 25% [?]