Welcome to the homepage for the ICML 2009 Workshop Sparse Methods for Music Audio. This workshop will be held on June 18 at McGill University as part of the International Conference on Machine Learning (ICML). Registration is required and costs 70CAD (50CAD student) before June 13 and 150CAD (100CAD student) at the site. See the ICML 2009 website for more details.


The workshop consists of a number of short talks separate by coffee breaks. Lunch is from 12:30 to 2:00, allowing additional time for discussion. The workshop will end with a panel discussion.

Morning Session
09:00-09:30Laurent Daudet (Invited talk), Is sparse coding useful for coding?
09:30-10:00Gert Lanckriet, Using Sparse CCA for Vocabulary Selection
10:00-10:30Matthew D. Hoffman, David M. Blei, and Perry R. Cook, Finding Latent Sources in Recorded Music With a Shift-Invariant HDP
10:30-11:00Coffee break
11:00-11:30Mikkel Schmidt (Invited talk), Audio source separation using sparse non-negative matrix factorization techniques.
11:30-12:00Gautham J. Mysore and Paris Smaragdis, Multipitch Estimation Using Sparse Impluse Distributions and Instrument Specific Priors
12:00-12:30Patrick Wolfe (Invited talk), Sparsity, adaptive time-frequency representations, and fast reconstruction
Aftrnoon Session
2:00-2:30Rajat Raina (Invited talk), Shift-invariant sparse coding for music classification
2:30-3:00Martin Rehn, Richard F. Lyon, Samy Bengio, Thomas C. Walters, Gal Chechik, Sound Ranking Using Auditory Sparse-Code Representations
3:00-3:30Matthieu Kowalski, Sparsity and Structure in Audio Signal Through Mixed Norms
3:30-4:00Coffee break
4:00-4:30Ramin Pichevar, Hossein Najaf-Zadeh, and Hassan Lahdili, Biologically-Inspired Sparse Coding of Music
4:30-5:00Youngmoo E. Kim and Travis M. Doll, Employing Sparsity for Joint Sound Source & Acoustic Channel Estimation
5:00-5:30Panel Discussion

Topic and Motivation

The goal of the workshop is to explore state-of-the-art machine learning methods for processing music audio. Relevant tasks include audio classification, music recommendation, polyphonic pitch extraction and measures of music similarity. The workshop is in response to recent developments in the use of sparse coding for audio analysis, but this will be interpreted in the broadest possible light in organizing the workshop.

Sparse coding is currently a hot topic in many areas of signal processing and machine learning, where it can find more interesting or more useful solutions to underconstrained, high-dimensional problems when compared to traditional regularized least-squares approaches. Music audio is a very good candidate for these approaches, since it is, in its raw form, very high dimensional (over 80,000 values in one second of music from a CD), but can usefully be described as the (underconstrained) superposition of a small number of separate signals, such as individual instruments, all subject to a large number of mutual constraints. This description in fact applies at multiple levels, from the raw audio through to compositional structure.

Despite this opportunity, there have been only a few publications on applying these techniques in music. We would like to hold the workshop as a way to focus, develop, and refine the various perspectives and approaches possible, and to raise the profile of these ideas, particularly to others working in music audio, but also to machine learning researchers who may be curious about working with music audio data. We will try to make the workshop accessible to both these communities.

Impact and expected outcomes

We have high hopes for significant and substantial progress, simply by having an event dedicated to this topic, and bringing together researchers with different, but we believe related perspectives. If all goes well, what will emerge is a more concrete picture of what sparse techniques can (and can not) be hoped to bring to the analysis of music audio. We would hope to publicize such an understanding more widely, perhaps through a journal special issue, or tutorials at related conferences in music analysis and signal processing conferences.


Douglas Eck
Associate Professor, University of Montreal Computer Science
LISA Machine Learning Lab / BRAMS International Laboratory for Research in Brain Music and Sound

Dan Ellis
Associate Professor, Electrical Engineering, Columbia University
LabROSA Laboratory for the Recognition and Organization of Speech and Audio

Philippe Hamel
PhD Candidate, University of Montreal Computer Science
LISA Machine Learning Lab / BRAMS International Laboratory for Research in Brain Music and Sound