Post-doctoral position on browsable representations of music content

We are offering a postdoc position on browsable representations of music content for 1 to 2 years starting in Fall 2010. Applications must be made online before March 20, 2010 at
http://www.inria.fr/travailler/mrted/en/postdoc/details.html?nPostingTargetID=8304
Questions about the position or the subject should be addressed by email to the principal investigator.

Principal investigator: Frédéric Bimbot (frederic.bimbot@irisa.fr)
Co-principal investigator: Emmanuel Vincent (emmanuel.vincent@irisa.fr)

Description of the project

As a consequence of massive production and diffusion of digital music, music companies and listeners are expecting increasingly advanced features from online music services and personal music archiving software. However, music pieces currently remain difficult to represent in a way that allows fast visualization and browsing of collections containing several thousand pieces.

Algorithms exist that can extract and quantify certain properties of a music piece via numerical or symbolic features. Nevertheless, the exploitation of these features is limited by the lack of a representation paradigm allowing intuitive browsing of a music piece or a collection of pieces according to a given combination of features and simple visualization of the structure of each piece or the relationships between pieces derived from these features.

The subject of this postdoc consists of proposing and validating a generic approach to the representation of music content allowing joint exploitation of multidimensional feature sets including features of different
- meanings (e.g. rhythm vs. melody)
- types (e.g. numerical vs. symbolic)
- time scales (e.g. beat-level vs. song-level)
- confidence levels (e.g. trusted features vs. missing features).

The outcomes of the postdoc are expected to provide advances in several areas by :
- providing a framework for the implementation of intuitive graphical user interfaces for multi-feature browsing of music pieces or music collections (despite possibly missing features)
- offering tools for the interpretation and diagnosis of novel feature extraction algorithms
- introducing, by means of the proposed representations, additional contextual knowledge exploitable by model-based audio signal processing algorithms either in the training or in the decoding phase.

Practical tasks to be addressed include :
- studying existing approaches to the representation of music content, including [1,2,3,4]
- determining desirable properties of browsable representations and finding suitable representation formalisms and data structures (segments, trees, clusters)
- defining and experimenting appropriate visualization and browsing procedures
- evaluating the relevance of the proposed representations and procedures according to a suitable evaluation protocol

Candidate profile

Prospective candidates should have a background in signal processing or computer science with some additional knowledge about musical audio. Proficient coding in Matlab or C++ is necessary. Candidates must have held a PhD degree for less than a year or be about to obtain one.

References

[1] P. Cano and M. Koppenberger, "The Emergence of Complex Network Patterns in Music Artist Networks", in Proc. 5th Int. Conf. on Music Information Retrieval (ISMIR), 2004.
[2] E. Pampalk and M. Goto, "Music Rainbow : A New User Interface to Discover Artists Using Audiobased Similarity and Web-based Labeling", in Proc. 7th Int. Conf. on Music Information Retrieval (ISMIR), 2006.
[3] K. Jacobson & M. Sandler, "Visualizing structured data about music: The k-pie network layout algorithm and applications to music data", in Proc. 2009 Web Science Conference (WebSci), 2009.
[4] L. Sarmento, F. Gouyon, B. Costa & E. Oliveira, "Visualizing Networks of Music Artists with RAMA", in Proc. 5th Int. Conf. on Web Information Systems and Technologies (WEBIST), 2009.