Duke Listens!: Visit my main blog at MusicMachinery.com

My notes from the ISMIR 2008 - invited speaker talk Abstract: In my talk, I will describe five properties that help make music sound tonal — or “good,” to most listeners. I will then show that combining these properties is mathematically non-trivial, with the consequence that space of possible tonal musics is severely constrained. This leads me to construct higher-dimensional geometrical representations of musical structure, in which it is clear how the various properties can be combined. Finally, I will show that Western music combines these five properties at two different temporal levels: the immediate level of the chord, and the long-term level of the scale. The resulting music is hierarchically self-similar, exploiting the same basic procedures on two different time scales. In fact, one and the same twisted cubic lattice describes the musical relationships among common chords and scales.

In this talk, Dmitri explained his geometrical model of music structure. He uses the model to explain ow harmony and counterpoint. The gist of the talk can be found in this Science article: The Geometry of Musical Chords. A good talk, interesting subject matter (but I must admit I'm skeptical).

More info: The Geometry of Music.

Day 1 of ISMIR. The first plenary session of ISMIR has 3 really interesting presentations right in a row:

• Music Thumbnailer: Visualizing Musical Pieces in Thumbnail Images Based on Acoustic Features by Kazuyoshi Yoshii and Masataka Goto - Masataka and his team always produce the most innovative and fun interfaces for interacting with music. What a great way to start ISMIR.
• MCIpa: A Music Content Information Player and Annotator for Discovering Music by Geoffroy Peeters, David Fenech and Xavier Rodet
• Development of a Music Organizer for Children by Edmond Zhang and Sally Jo Cunningham

It should be a good day.

Yesterday was the tutorial day. Elias and I presented our tutorial on social tags and MIR. The tutorial was well attended - and we had lots of good questions at the end of it all. Much thanks to Elias for being such a gracious co-speaker. We had a very tight timetable for the talk - and I took longer to present my bits than I was scheduled for - but Elias seamlessly shortened his parts to make up for my overages.

Photo by Anita

The collateral information for the talk is at SocialMusicResearch.org. Feel free to register and contribute to the wiki if want to point the community at some work that you are doing or some interesting aspects around social music and MIR.

I'm taking notes during ISMIR - since some people may be interested, I'm blog them.

ISMIR kicks off with a tutorial on the ChucK language, presented by Ge Wang (Stanford) and Rebecca Fiebrink (Princeton). This is a hands on tutorial based on the miniAudicle.

Ge Ge Wang gave an overview of core language features:

• The chuck operator:
  
• Example:

  Impulse i => dac;
  
  while (true)
  {
        1 => i.next;
        100::ms => now;
  }
  

  This program clicks the speaker every 10th second

Next Ge Wang explains all about time and concurrency. Ge Wang is coding on the fly in response to a question ... he has no fear.

Next Ge Wang coded up some demos - cool sounds. Ge Wang demonstrated the audicle - which has all sorts of interesting visualizations of the active shreds.

It's about time for me to finalize prep for the social tags tutorial - so I shall miss the second half on the tutorial. Too bad, since this has been a great intro to ChucK and I'd like to hear more.

We've all seen those iPod unboxing posts ... Now it's my turn -- because look what just arrived in the mail -- The proceedings for ISMIR 2008 - W00t!

My first view of the volume, still in bubble wrap.

Out of the box, but still wrapped in the protective covering. Never touched my human hands.

And finally, out of the cover. It is so very shiny. Shinier than my iPhone.

Here's the back side. It is definitely thicker than my iPod. Notice the crack in the Liberty Bell. Apparently all Version 1 copies of the proceedings are exhibiting the same susceptibility to cracking.

And finally a peek inside -- nice to see Sun on the sponsorship page.

The proceedings are just beautiful. The ISMIR 2008 organizers have done just an incredible job. I am glad I decided to get the hard copy.

During times like this, I like to listen to ice-creamo - its emo music about ice cream.

Elias and I have been working hard to prepare for our tutorial on Social Tags and Music Information Retrieval that we are presenting in a couple of days at ISMIR. We are presenting a whole lot of material in this three hour tutorial. Perhaps too much for anyone to be able to remember it all. With this in mind we are creating a site called SocialMusicResearch.org that will be the clearing house for all of the material that we present during the tutorial. At this site you'll find our slides, links to code, links to data, our bibliography and other pointers to interesting topics relevant to social tags and music information retrieval.

Instead of just creating a static site, we've set up SocialMusicResearch.org as a wiki - so that other researchers that are interested in social tags and music can contribute and help make this the place to go to learn all about social tags and music information retrieval, and social music in general.

We've just started populating the wiki so now's the time to get involved if you like being in on the ground. If you'd like to start contributing to the wiki, sign up and let me know your user ID (email it to [email protected]) and I'll add it to the set of editors.

Another easy way to participate in Social Music Research is to just start tagging relevant web pages with the tag "SocialMusicResearch" on Delicious. Elias talks about this on his research blog.

One of the things we've been working on this summer on Project Aura is a web application called The Music Explaura. The Explaura is designed to help people discover new music. One of the key features of the Explaura is that it offers transparent, steerable recommendations - you can ask the recommender why something was recommended, and you can steer the recommender toward music that you'll like. If you like heavy metal music, but you don't like cookie monster vocals you can point the recommender in the right direction. It's fun to play with and I've been finding lots of new music with it.

All the cool bits in the user interface were implemented by our summer Über-intern Francois Maillet. Francois will be demoing the interface during the Late Breaking/Demo session at ISMIR so if you are interested in seeing the Explaura in action be sure to stop by.

Update: Here's a little video showing steerable recommendations in action

Here's my favorite freakomendation from the iTunes Genius music recommender. The 'Genius' suggest that since I like Led Zeppelin's Stairway to heaven that I might like the song Love will keep us together by The Captain and Tennille.

Here's a first look at the new Genius feature of iTunes 8. First of all, there have been a few reports that the Genius feature will be 'Pandora-like.' This is not true - the iTunes recommendations seem to be rather run-of-the-mill collaborative filtering recommendations based upon the wisdom of the crowds - as compared to the core technology behind Pandora which uses a detailed content-analysis of the music (by a trained staff of musicians) to generate recommendations.

So how are the Genius recommendations and playlists? Well, there are some funny bits to start off with. For instance, if try to get recommendations based upon a Beatles song, the 'Genius' can't help you at all.

Instead, it just falls back on what's popular. It doesn't even try to match the Genre (britpop? 60s psychedelia) or artist (John Lennon, Wings?). Pink and Rihanna are right out. This glaring gap is not just confined to the Beatles, but it seems to occur for other artists like Bob Seger, and AC/DC that have shunned the iTunes store .

Similarly, for long tail artists, the Genius didn't have too much to say. For example, the Genius recognized artist DJ Markitos and was ready to sell me another DJ Markitos album, but couldn't offer me any other recommendations:

When I avoid the long tail, or the iTunes-averse artists, I can coax the Genius to give me some real recommendations. The first thing I notice is that the recommendations seem to be artist-based and not album or track-based. For instance if I select as a seed track Emerson Lake and Palmer's 'Food for your Soul' (which is a big band jazz-fusion track), I get recommendations for Yes, Jethro Tull, The Moody Blues, Procol Harem, and Renaissance. When I pick ELPs - Two Part Invention (Bach on a xylophone), I get the same recommendations for tracks by Yes, The Moody Blues, Procol Harem, Renaissace etc. And surprise, surprise, if I pick a progressive rock track by ELP I get recommendations for the same set of artists.

The same is true for the playlists generated by the Genius - the Genius just picks tracks from similar artists regardless of how well the track is representative for the artist. For instance, on the album 'Return to the center of the earth' by Rick Wakeman the tracks alternate between music and narration. If I use one of the narration tracks as a seed, the Genius gives me run of the mill tracks by progressive rock artists. No track-similarity is used to generate the playlists. This is not Pandora-like behavior.

So, the Genius isn't great with the long tail, and can't give you good recommendations based upon a track or an album. So, how does it do with regular old recommendations? Let's look at a few:

Seed song:: Dazed and Confused by Led Zeppelin.
Recommendations:: Spoon, RHCP, Pure Parie League, Chuck Berry, System of a Down, Madonna, Linkin Park (the Chuck Berry song was the Christmas ditty'Run Rudolph Run')
Comments:I think we are seeing some extreme popularity bias here. The artists just don't match.
My Score: D

Seed song:: My Name is Jonas by Weezer
Recommendations:: Audioslave, Death Cab For Cutie, blink-182, Tenacious D, Bright Eyes, Weezer, The Shins
Comments:No clunkers this time, but nothing new - no novelty or serendipity. I could get a much better list from All Music guide
My Score: C

Seed song:: Fanfare for the Common Man by Emerson Lake and Palmer
Recommendations:: Yes, Focus, Genesis, The Moody Blues, Procol Harum
Comments:All recommendations from the progressive rock bands of the 70s. Again, no clunkers, but this list is not going to help people find new music - no Gentle Giant, no Porcupine Tree or Dream Theater. This is classic rock radio
My Score: C

Seed song:: Fidelity by Regina Spektor
Recommendations:: Ingrid Michaelson, Augustana, Lily Allen, Hellgoodbye, Mat Kearney, KT Tunstall, Corinne Baily Rae, Blue October, OK Go, Death Cab for Cutie, Gwen Stefani, Dashboard Confessional, Jason Mraz, John Mayer
Comments:Regina Spektor is a quirky indie singer-songwriter. One of these recommendations seem good to me: Ingrid Michaelson (hadn't heard of her, and she has that indie vibe), but the rest are just bad, (the Gwen Stefani recommendation is downright laughable). One good recommendation in ten - not too good
My Score: D

Seed song:: Money by Pink Floyd
Recommendations:: Eric Clapton, Jimi Hendrix, The Doors, Cream, ZZ Top, Pink Floyd, Deep Purple, Alice Cooper, Blue Oyster Cult, The Rolling Stones, Steppenwolf
Comments:Another set of artists from the greatest hits of the 60s and 70s album. Yawn
My Score: D

I like to rate recommendations on 4 factors:

• Relevance - how well do the recommended items match the taste - the genius does OK here, with a few glaring errors like the Gwen Stefani recommendation for Regina Spektor.
• Novelty - how well does the recommender find items that I haven't heard of - the Genus doesn't fair well at all - when it did recommend items that I hadn't heard of, they were almost always irrelevant.
• Reach - how far does the recommender get into the long tail. The Genius couldn't give recommendations for the long tail items that I tried, nor did it give recommendations that were far into the tail (and I couldn't steer it to hipster items).
• Transparency - how well does the recommender explain the recommendations. A recommender like Pandora does a great job explaining why an item was recommended, however, the Genius doesn't even try to explain its recommendations.

The iTunes Genius is just a run-of-the-mill collaborative filtering recommender - the recommendations are nothing special - there's no advanced content analysis like you get from MusicIP. There's no deep analysis of content and context like The Echo Nest. There's no social community or tags like at Last.fm. There's no transparency in the recommendations like at Pandora. - the Genius just gives rather pedestrian recommendations and playlists.

Apple and iTunes are at a choke point in the music ecosystem - Apple has an opportunity to help people explore for and discovery new music. It is good to see Apple recognize this and at least start to provide recommendation tools. I hope they will start to look at all of the interesting work being done by researchers in the Recommenders and Music Information Retrieval community and try to make their Genius a little smarter.

I've just upgraded to iTunes 8.0. I am eager to try out the new Genius feature (perhaps you didn't notice Steve Jobs mentioning it at the "Let's rock" announcement today). When you enable the 'Genius' feature, iTunes asks you to agree to let iTunes send data about your iTunes library and your play history to the iTunes cloud.

Once you agree, the Genius goes through a 3 step process

I'm a bit puzzled about what the Genius is doing during this first step. I have about 10K songs in my iTunes library. The Genius took about a half hour to 'gather information' about my library. That's about 200 ms per track ... not enough time to do any kind of analysis (not even enough time to decode an MP3, let alone do any significant processing). iTunes already has a database containing all the info it needs. Perhaps it is calculating an item-item similarity matrix for my artists, albums and tracks before it sends the data off too the cloud. This would save Apple a lot of CPU cycles.

Once iTunes has finished gathering information about my library it sends it all to Apple and a few minutes later, Apple sends back my Genius results. Which means that I can now get recommendations for recommendations. More about the recommendations in the next post.

SoundUnwound (the crowd-sourced music metadatabase by Amazon) has an UnSpun poll asking users to vote on which new features.  I've added my two cents voting for (1) Sample audio for artists (2) Web services to let my application get the SoundUnwound data and (3) Use a creative commons license for user-contributed data.    

Band Metrics  has just launched a private beta of its semantic web application that collects, analyzes and displays dynamic popularity and trends about musicians and bands. BandMetrics is a Georgia based, 5 person startup.

The EchoNest has just launched its next set of developer-oriented web services.   These web services include everything you need to roll your own next-generation music website, including services to get reviews, audio, video, news and blog posts about any artist.  Neat stuff!

BBC Radio labs has just launched Radio Pop - social listening for the BBC.  With Radio Pop, your listening habits are scrobbled to the BBC - which will then show you all sorts of charts, graphs, and eight-by-ten colour glossy photographs with circles and arrows about your listening habits (notice I used the British spelling of 'color').

Radio Pop is making the data that they collect available through an API so researchers and developers can play with the data and build new apps and websites around the data.

Unfortunately, I haven't been able to get the Radio Pop Player to work (the player is the thing that scrobbles your listening, and it doesn't seem to be mac-friendly), so I haven't been able to see how well it works or to play with  the data yet.  Still, I think this is a really interesting idea and I am eager to try it out when I can figure out how to scrobble my BBC listening.  I've been really impressed about  the folks at the BBC labs.  They  are thinking hard about how to bridge the divide between traditional radio and the social web.   I think there are lots of opportunities here - especially when you think of the social possibilities when millions of people are all listening to the same show.  Read more at the BBC Radio Labs blog.

Update:  It is working now ... I just had to install realplayer.  Works great.