Image Processing Reference
In-Depth Information
CHAPTER 2
An approach to classifying
four-part music in
multidimensional space
Gregory Doerfler; Robert Beck Department of Computing Sciences, Villanova University, Villanova, PA, USA
Abstract
Four-part classifier (FPC) is a system for classifying four-part music based on the known classifications
of training pieces. Classification is performed using metrics that consider both chord structure and chord
movement and techniques that score the metrics in different ways. While in principle classifiers are free
to be anything of musical interest, this paper focuses on classification by composer.
FPC was trained with music from three composers—J.S. Bach, John Bacchus Dykes, and Henry Thomas
Smart—and then tasked with classifying test pieces writen by the same composers. Using all two-or-
more composer combinations (Bach and Dykes; Bach and Smart; Dykes and Smart; and Bach, Dykes,
and Smart), FPC correctly identified the composer with well above 50% accuracy. In the cases of Bach
and Dykes, and Bach and Smart, training piece data clustered around five metrics—four of them chord
inversion percentages and the other one secondary chord percentages—suggesting these to be the most
decisive metrics. The significance of these five metrics was supported by the substantial improvement in
the Euclidean distance classification when only they were used.
Later, a fourth composer, Lowell Mason, was added and three more metrics with similarities to the five
that performed best were introduced. New classifications involving Mason further supported the signi-
icance of those five metrics, which the additional three metrics were unable to outperform on their own
for improve in conjunction.
Keywords
Classiication
Clustering
Four-part music
Metrics
Search WWH ::




Custom Search