|
|
NWO-VIDI
project MUSIVA Modelling musical similarity
over time
through the variation
principle
Principal
investigator: Dr. Anja Volk |
|
|
|
in the media: e-data & research,
March 2011, pdf
News
Project
Description:
Introduction
The aim of this project is to deliver a
cognition-based computational model on music similarity that
grounds in the variation principle employed in
classical, folk and popular music. The project will integrate
knowledge and methods from Music Information Retrieval,
Musicology and Cognitive Science.
The assessment of similarity is fundamental for cognitive
processes. Cognitive Science provides formal models for the
theoretical framework and empirical measurement of similarity,
demonstrating its crucial influence on our conception of the
world. However, no comprehensive theory exists on how
listeners use similarity to predict, categorize or appreciate
music. This is a major problem in the rapidly growing
discipline of Music Information Retrieval (MIR). MIR
researches methods that allow users to retrieve music that is
similar to musical queries representing their needs via the
Internet. Developing content-based search methods for music
faces the challenge of relating features of the music to
listeners’ experience of similarity. Most methods developed in
MIR extract low-level features from the music, severely
limiting the cognitive plausibility of search results and
hence impairing the usefulness of current retrieval systems.
This project will investigate the fundamental principle of
variation in music studied in Musicology and Cognitive Science
as a means to establish similarity. The project will research
computational modelling of music
similarity in the symbolic domain (i.e. using
perception-related notation, such as MIDI or **kern) based on
the variation of structural elements (such as motives, rhythms
and chord sequences). Specifically, it will take into account
the interaction between global and local features of the music
and will address music as unfolding in time.
The project will deliver a model of music similarity that
covers three major styles, namely classical, folk and popular
music. The envisioned model will be based on cognitive and
structural aspects of music, addressing high-level processes
in establishing similarity. Hence, the project will make a
major step towards cognition-based similarity models in music,
which are urgently needed for the design of meaningful music
retrieval systems. The development of a theoretic framework
for similarity in music will contribute to the search for
general principles of similarity across different domains
envisioned in Cognitive Science.
Aims and Objectives
The overall aim of this project is the
development of a computational model of music similarity in
the symbolic domain (i.e. based on perception-related
notation) that grounds in the variation principle employed in
classical, folk and popular music. Specifically, there are
four key objectives:
Objective 1: To deliver a theoretic
framework for similarity in music based on the variation
principle across classical, folk and popular music. Test cases
for different forms of similarity relations will be identified
that serve as a ground truth for the computational model.
Objective 2: To design segmentation methods
that partition music into perceptual units in which variations
of structural elements occur.
Objective 3: To model local and global
features of the music that are relevant for experiencing
similarity based on variation.
Objective 4: To model the integration of
segments into their context by considering the unfolding of
musical structure over time; this approach links local
elements dynamically with their global contexts. A similarity
model that is based on the interaction of local elements and
global contexts will be developed.
Significance
Modelling similarity based on the variation principle is urgently needed to establish a general theory of similarity in music. This is crucial for the successful design of MIR systems and will open up new research lines in music cognition research. Integrating the concept of variation (Musicology), formal models of similarity (Cognitive Science) and computational modelling (Information Retrieval) into a comprehensive approach to music similarity will deliver important aspects for a general theory on similarity across different domains.
· Music Information Retrieval: The largest part of modelling music similarity in Music Information Retrieval is currently in the audio domain. The extracted low-level features allow the comparison of broad categories of music, such as genres, but are less successful for establishing similarity for finer-grained data, such as for similar songs within the same genre. High-level features extracted in the symbolic domain from formats such as MIDI or **kern are closer to the way people perceive music. This project addresses the high-level processing in establishing similarity in music and hence will deliver an essential requirement for building cognitively plausible search algorithms in Music Information Retrieval. The lack of established concepts of similarity in music causes the lack of suitable ground truths on which algorithms can be tested (Downie et al, 2009). For the three selected music styles, test cases based on the variation principle will establish a ground truth for the computational modelling. The similarity model will deliver a theoretic framework that enables systematic strategies within the empirical cycle for quantitative approaches to similarity in MIR.
· Musicology: The musicological discourse on the variation principle is based on small numbers of selected musical examples and has not led to a general concept of variation. The computational modelling allows to formalize the concept and to explicitly test it on a large collection of music.
· Cognitive Science: While research in music cognition has strongly focussed on the experienced listener of Western classical music (Peretz, 2006), research on music similarity contributes an excellent topic on basic music skills. Similarity is used as a default method to reason about a domain, even if we do not have specific knowledge about it (Goldstone & Son, 2005). Thus, understanding music similarity will demonstrate that accessing music is not reserved to the highly trained specialist. Moreover, our model of music similarity will contribute an underrepresented domain to similarity research in Cognitive Science and hence contribute new aspects to the search for general principles of similarity across different domains.
Approach
Musicology provides analytic descriptions of
music closely related to the phenomenon of similarity that are
based on the variation principle. The term “variation
principle” refers to the variation of musical patterns such
that listeners are able to relate different components of a
piece of music, or entire pieces, to each other and hence
experience similarity relations. Although no formalized
concept of the variation principle has been developed, studies
on techniques to vary musical patterns such that they
establish similarity in classical, folk and popular music
strongly suggest that the variation principle is a universal
principle in music. Recent studies in Cognitive Science
show that similarity relations based on the variation of
characteristic elements as described in Musicology are
recognized both by experts and novices. This project will take
into account different ways of interaction between global
and local features of the music as a basis for
modelling the concrete
manifestations of the variation principle studied in
Musicology and Cognitive Science. The data-rich approach of
computational modelling within Music
Information Retrieval allows to corroborate the
generality of the variation principle in establishing
similarity.
Overview over relevant workpackages
(WP):
WP- 1: Domain modelling and test
cases.
Addressing objective 1, this workpackage
will build domain models for the formalization of variation
described for classical, folk and popular music in Musicology.
Relevant musical features for experiencing music similarity
based on variation will be identified. Appropriate test cases
for similarity relations will be created and evaluated to
serve as a ground truth for the computational modelling.
WP-2: Segmentation.
This workpackage aims to segment
a musical piece into perceptual units addressing objective
2. Stable segments will be determined that serve as
local contexts of salient elements, such that the global
structure of a musical piece is described as a patchwork of
these local contexts.
WP-3: Relevant features.
Addressing objective 3, this workpackage
designs local and global features that are relevant for the
variation principle. Developing appropriate features for
aspects such as melody, harmony and rhythm, and applying them
to different segment sizes, determines change and persistence
of features on different levels of the piece with respect to
local and global information.
WP-4: Interaction between local and global components.
This workpackage will integrate
segments into their global context and thus addresses
objective 4. Integrating local segments dynamically into their
global contexts will be realized through the unfolding of
relevant musical features over time.Similarity
measures based on the interaction of the local and global
contexts will be derived. A computational model of music
similarity will be designed that takes these multiple aspects
into account for a hybrid approach to similarity in music.
WP-5: Theoretic model of similarity in music.
The evaluation of the computational modelling
within WP-2 to WP-4 with respect to the domain modelling of similarity in WP-1 will
lead to a theoretic model on similarity in music.
Research Team
The
research
team includes
·
principal
investigator: Dr. Anja
Volk
·
Postdoc:
Bas de Haas
·
PhD
student: Marcelo Rodriguez-Lopez
·
scientific
programmer
(t.b.a.)
Contact for more detailed project description:
Anja Volk
Department of Information and Computing Sciences
Utrecht University
email and further contact
details