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

The latest Sound Opinions podcast has an excellent segment on The Dawn of Metal about the early history of Heavy Metal with references to

• Steppenwolf
• Blue Cheer
• Hawkwind
• Led Zeppelin
• Black Sabbath
• Deep Purple
• Uriah Heep
• Judas Priest
• Motorhead

It is worth a listen - and while you are at it, just subscribe to the podcast - it is a great way to keep up with the latest trends in music.

The ISMIR organizers have just added the late-breaking and demo sessions to the program page.  It looks like there will be lots of interesting ones to see, including a number that seem to be related to social tags.

Social tag related late-breaking/demo submissions

• Music Retrieval Based on Social Tags: A Case Study
• Herd the Music - A Social Music Annotation Game
• MOODY: A Web-Based Music Mood Classification and Recommendation System
• Creating Transparent, Steerable Recommendations

 There are some others that look really interesting too including:

• A Preference Ranking Model Using a Discriminatively Trained Classifier
• Audio-to-Score Alignment In the Audacity Audio Editor
• FM4 Soundpark: Audio-based Music Recommendation in Everyday Use
• Music and Lyrics: Can Lyrics Improve Emotion Estimation for Music?
• A Geographic Interface for Large Music Collections
• MusicBox: Navigating the Space of Your Music
• Realtime Audio Analysis for Humanoid Robotics and Dance
• Content-based Music Information Retrieval System with Automatic DJ Mixing

As you can see, I am a sucker for anything related to recommendation, visualization, discovery and/or playlisting.  I'm really looking forward to the demo session.  

A note to anyone doing a demo - if listening to music is an essential part of your demo, you may want to consider bringing a set of high-quality over-the-ear headphones for demoees to use.  The ISMIR demo spaces are always very loud (imagine several hundred researchers talking, snacking and listening to music simultaneously). Laptop speakers definitely will not be heard in the demo space.

Elias and I are presenting a tutorial on social tags and music information retrieval in a couple of weeks at ISMIR.  We want to include forward pointers to all of the relevant ISMIR work.  We can easily do that for the accepted papers since the proceedings are already available online. However, I suspect that there may be some posters/demos in the late breaking session that will be relevant as well.  Since the list of late breaking posters and demos has not been published it is hard for us to point to the proper ones.  So, if you have a late-breaking poster or demo that you think might be relevant to a tutorial on social tags and MIR and you'd like us to point to it from our tutorial, please send me the info about it  - title, authors and abstract - to [email protected]

Today Amazon has launched the beta of SoundUnwound - a music web site that looks to be the Wikipedia of music data.  SoundUnwound is combining data from IMDB, Amazon and Musicbrainz along with user-contributed data to serve as a one stop site for all the information you could ever want about an artist, album, track, or label.

Already SoundUnwound has lots of interesting data - multiple genre classifications for an artist, detailed profiles, discographies, info on artist members, similar artists, trivia.  They have a very nice artist timeline.  Here's a subset of the Beatles timeline:

Although SoundUnwound is relying on users to contribute content, they don't seem to be ready to trust all of the data to their users.  User edits are reviewed by the SoundUnwound staff before they become official and seen by everyone.

 Unfortunately, unlike Wikipedia or Musicbrainz, the data that SoundUnwound collects is not going to be freely available - there's no creative commons license.  I am skeptical about any wiki-style site that doesn't make the user-contributed data freely available.

Except for the licensing,  I am really excited about SoundUnwound - but I have a few suggestions that could make it really great:

• Build web services around the data so any 3rd party can get this structured data
• Include access to 30 second clips in these web services
• Make it easy to go between ASINs, MusicBrainz IDs, and SoundUnwound IDs
• Add support for social tags 

Read more at the SoundUnwound blog and read about the Musicbrainz integration at the Musicbrainz blog.

Now that the proceedings for ISMIR 2008 are online, I am reading papers that are on topics that interest me.  Today I read the paper Five Approaches to Collecting Tags for Music by Douglas Turnbull, Luke Barrington and Gert Lanckriet.  In this paper, Doug and company compare five approaches to collecting tags for music:  surveys, web mining, harvesting social tags, autotags and games.   The paper presents a well-balanced and complete overview of how tags for music can be collected.  They describe some of the issues with each of the methods.  Now I don't always agree with some of their characterizations about the strengths and weaknesses of the various approaches which means we'll be able to have some interesting conversations in Philly in a couple of weeks.   One topic for discussion: I'd like to get a better understanding of the difference between tags collected in a social context (such as those collected at Last.fm) and those collected in a gaming context. 

 All in all, this is a good paper If you are interested in tags and music information retrieval, this paper should be on your short list.

Last year I posted about my skepticism about the "hit predictors" that claim that they can use machine learning algorithms to predict which songs will be hits.  So when the proceedings for ISMIR 2008 went online, the first paper I downloaded was Hit Song Science is Not Yet a Science by Francois Pachet and Pierre Roy.  In this paper, Pachet and Roy set out to validate the hypothesis that the popularity of music titles can be predicted from acoustic or human features.  It is an interesting paper that concludes:

... the popularity of a song cannot be learnt by using state-of-the-art machine learning techniques.  This large-scale evaluation, using the best machine-learning techniques available to our knowledge, contradicts the claims of "Hit Song Science", i.e. that the popularity of a music title can be learned effectively from known features of music titles, either acoustic or human.

The authors do not close the door on 'hit song science' - they limit their conclusions to say that with the current state-of-the art features, we cannot predict hits - but with better features it may be possible.  They conclude:  Hit song science is not yet a science, but a wide open field.

As Elias has pointed out, the ISMIR proceedings are online.   If you don't want to download the nearly 700(!) pages, you can go to the program page on the ISMIR site to select the papers and topics of interest.  It is really great to be able to get the proceedings before the conference - it will make it easier to plan my trips through the poster sessions (and also avoid a repeat of my worst ISMIR moment).

This looks real fun. RjDj is a music application for the iPhone. It uses sensory input to generate and control the music you are listening to. It is not released yet .. but there's a private beta that you can sign up for.

 

Amazon just purchased Shelfari - a social website for book lovers (it's like Last.fm for books). This makes sense since Amazon has had a hard time getting their users to contribute social data beyond book reviews - data like social tags, as well as taste data associated with book lovers. Personally, I think Amazon made the wrong choice - I like LibraryThing much better - (even if its just because of the Shelfari astroturfing incident).

Now imagine if Amazon wanted to improve its AmazonMp3.com music store with gobs of social data about music - clearly the thing to do would be to buy the company that already has all of the data - that would be Last.fm. Unfortunately for Amazon, Last.fm is now owned by CBS and probably won't be selling it soon.

Anthony Liekens has put together a few web tools that do fun things with Last.fm tags. One tool is Compare User Tag Clouds. This tool takes two last.fm users calculates a tag cloud for each individual based upon their listening habits, and then shows the overlap between the two clouds - that represents the overlap in music preferences. This cloud is a good representation of your common music tastes. I especially like the difference cloud the tool shows that highlights the difference in music preference. Here's the difference cloud between my listening (in green) and Steve's listening (in red).

It is easy to see from this cloud the kind of music that Steve likes that I don't. Also, can you guess what country Steve is from?

Some of been proclaiming that 2008 is the year that the Tag Cloud dies - that this Web 2.0 icon has run its course and will disappear from the web. Certainly there are many sites that have added social tags to their site in such a poor manner that they should be eliminated. We've already seen a few sites such as QLoud, de-emphasize or eliminate tags all together. But I don't think this is the end of social tags - it is just the weeding out of all of the bad implementations of tagging. I think this year we will start to see more innovative ways to use tags to help people organize, explore and discover new content.

Btw, Elias (from Last.fm) and I are giving a tutorial on social tags and music - so if you are interested in social tags and music you might want to check it out.

Here's a recommendation from AmazonMp3.com that registers high on the relevance scale but a bit low on the novelty and serendipity scale.

This is not necessarily a bad recommendation from the point of view of a store trying to sell MP3s, but it is just not the best recommendation for someone who is trying to find some new music.

There's a thread over on the Facebook ISMIR group where people are posting the title and abstract of their late breaking demo session. If you are going to be showing a demo at ISMIR next month then I hope you will add your demo to the list.

The detailed conference program for ISMIR 2008 has been posted. It is pretty interesting to see all of the sessions for the week laid out. If you are speaking you can find out if you lucky enough to be the first speaker of the conference on Monday morning, or the last speaker of the conference on Thursday. From this program we learn that there are two invited speakers this year at ISMIR: Dmitri Tymoczko of Princeton University is speaking about the Geometry of Consonance, and Jeanne Bamberger of MIT/UC Berkeley gives the keynote: Human Music Retrieval: Noting Time. There is also an 'Invited Panel Discussion' on commercial music discovery and recommendation (but no mention as to who will be on the panel).

I couldn't wait until Rebecca's ISMIR tutorial on ChucK - so I downloaded the latest ChucK distribution, fired it up and worked through the reference guide. It's a lot of fun. ChucK is a strongly-timed, concurrent, and on-the-fly audio programming language that presents a time-based, concurrent programming model that's highly precise and expressive. In some ways, ChucK reminds me of CSound - in that you have lots of control over how to make sounds. The more you know about signal processing and filters and the theory of synthesis the more you can do with ChucK. But unlike CSound, ChucK is a modern program language with classes, objects - all the things that make a Java programmer comfortable, whereas CSound is like coding in assembler.

ChucK is pretty powerful too. Here's a bit of ChucK code by Perry Cook that does a reasonable job imitating the THX deep note - compare this 80 lines of code to the 20,000 lines of code required to make the original.

// THX emulator
//   author: Perry R. Cook (Jan 8, 2007)
// modified: Ge Wang (added parameters up top)

// F-1, B1b,  F1,    B2b,   F2,    B3b,   F3,    A5,    F4,   A6
[ 29.0, 87.5, 116.0, 175.0, 233.0, 350.0, 524.0, 880.0, 1048, 1760,
  29.0, 87.5, 116.0, 175.0, 233.0, 350.0, 524.0, 880.0, 1048, 1760,
  29.0, 87.5, 116.0, 175.0, 233.0, 350.0, 524.0, 880.0, 1048, 1760
] @=> float targets[];

// storage
float initials[30];
float deltas[30];

// parameters (play with these to control timing)
10000 => int steady_samps;
20000 => int sweep_steps;
15000 => int hold_steps;
8000 => int decay_steps;

// UGens
SawOsc s[30];
Gain gl[30];
Gain gr[30];
JCRev rl => dac.left;
JCRev rr => dac.right;

// reverb settings
0.025 => rl.mix => rr.mix;

// variables
0 => int i => int j;

// compute stuff
for( 0 => i; i < 30; i++ )
{
    // random freqs
    Std.rand2f( 200.0, 800.0 ) => initials[i] => s[i].freq;
    // 10 sample updates
    ( targets[i] - initials[i] ) / sweep_steps => deltas[i];
    // initial gain
    0.1 => s[i].gain;
    // random
    Std.rand2f( 0.0, 1.0 ) => gl[i].gain;
    // panning
    1.0 - gl[i].gain() => gr[i].gain;
    // hook up
    s[i] => gl[i] => rl;
    // all the oscs
    s[i] => gr[i] => rr;
}

steady_samps :: samp => now;                            // steady cluster

while( j < sweep_steps ) {
    for( 0 => i; i < 30; i++ ) {
        initials[i] + (deltas[i]*j) => s[i].freq;       // sweep freqs.
    }
    j + 1 => j;
    10 :: samp => now;
}

0 => j;
while( j < hold_steps ) {                               // hold chord
    10 :: samp => now;
    j + 1 => j;
}

0 => j;
while( j < decay_steps ) {
    for( 0 => i; i < 30;  i++) {
        0.1 * (decay_steps-j) / decay_steps => s[i].gain;       // decay gains
    }
    10 :: samp => now;
    j + 1 => j;
}

60000 :: samp => now;                                   // reverb tail

I've just created a Facebook group for those who are going to ISMIR this September in Philadelphia. If you are going to ISMIR (or just wish you were), please join the group. Please feel free to invite people to the group too. You can find the group here: ismir2008 facebook group.