Musical style recognition - a quantitative approach
Submitted to CIM04, 27 october 2003
Peter van Kranenburg
Faculty of Arts, Utrecht University, Utrecht, Netherlands
p.vankranenburg@lodebar.nl
Eric Backer
University of Technology, Delft, Netherlands
e.backer@ewi.tudelft.nl
Desired mode of presentation
talk.
Background in music history / music analysis
In music history it is often considered important to have a reliable overview of the
oeuvre of a certain composer. When critical editions are made, this is even a
necessity. But also to understand the importance of a certain composer, we need to
have a good overview over his work. Unfortunately many pieces exist of which the
composer cannot be determined with great certainty. This can be caused by poor
source material, or by sources that contradict each other. Not rarely this leads to
authorship-discussions. Various kinds of evidence are used to defend an attribution.
These can be categorized into external and internal evidence (Love, 2002). Stylistic
evidence is a subcategory of the latter.
Background in computing
In the subdiscipline of machine learning, many algorithms are developed to extract
knowledge from measurements in order to make automatic recognition of classes of
objects possible. For an overview see e.g. Webb 2002.
Aims
The aim of this experiment is to explore the possibilities of machine learning for
composer attribution. Since no such research is done before, it is desirable to get an
indication of the suitability of these methods, before applying them to 'real' problems.
The focus is on low-level characteristics of counterpoint. So, only polyphonic
compositions are taken into account.
Method
A dataset is made with compositions of five well-known composers: J.S. Bach,
Telemann, Handel, Haydn and Mozart. From Mozart and Haydn only stringquartet
movements are incorporated. The other composers are represented with
compositions in various genres. Of each composition 20 features (style markers) are
measured. Most features are low-level counterpoint characteristics, such as the
amount of certain intervals between the voices compared to the total number of
intervals or the amount of parallel thirds. Besides these, the entropy of the various
sonorities, the fraction of dissonant sonorities and the "steadiness" of the rhythm is
computed. Then the pattern recognition algorithms are used to obtain knowledge
from this data about the uniqueness of each style compared to the others. Also some
classifiers are built. The algorithms used are: k-nearest neighbor classifier, k-means
clustering and a decisiontree (C4.5). The fisher-transformation is used to reduce the
dimensionality. In the transformed space, also a nearest neighbor classifier is trained.
Results
The clustering shows that the compositions of the each composer do form a cluster in
the featurespace (typical error between 10% and 20%). Only the separation between
the stringquartets of Mozart and Haydn is more difficult (error of 35%).
The decisiontree is used to learn which features are important for separating the
styles. These features will appear in the top-nodes. We can learn, for example, that
the style of Bach is isolated from the other composers by taking those compositions
with low amount of parallel thirds and a steady rhythm and a high fraction of
dissonant sonorities. About styles of the other composers, similar observations can
be made.
The nearest neighbor classifier performs very well in the transformed feature space,
with an error-rate of 22% for the stringquartets of Haydn and Mozart, and an errorrate between 4% and 10% for the styles of the other composers. After removing
some outliers, even lower error-rates can be obtained.
Conclusions
These results indicate clearly that it is possible to recognize musical style
automatically. So, this kind of research can be a valuable addition to more traditional
methods of musical style analysis. It offers a quantitative evaluation of the styles
rather than the traditional qualitative descriptions.
References
Love, H. (2002). Attributing Authorship: An Introduction, Cambridge: Cambridge
University Press.
Webb, A. (22002). Statistical Pattern Recognition, Chichester: John Wiley & Sons.
Biographies
name
current position
relevant qualification
main field of research
name
current position
research interests
Peter van Kranenburg
Doctoral student musicology, University of Utrecht,
Netherlands.
M.Sc. Electrical Engineering, 2003, Delft University of
Technology, Netherlands.
supervisor: Eric Backer, Title: Musical style analysis using
statistical pattern recognition.
Composer Attribution.
Eric Backer
Professor of Information- and Communication Theory, Delft
University of Technology/University of Twente; visiting
professor University of South Florida.
Pattern Analysis and Machine Intelligence, Data Mining,
Information Theory.
qualifications
book publications
editorial activities
Ph.D. Delft University of Technology, Dean (1989-1998)
Electrical Engineering, Computer Science and Mathematics
Author, Computer Assisted Reasoning in Cluster Analysis
(Wiley, 1995)
Editor in Chief, Pattern Recognition Letters (Elsevier)