An approach to classifying four-part music in multidimensional space - Emerging Trends in Image Processing, Computer Vision, and Pattern Recognition - page 30

the last classification was rerun to include it. Consequently, results saw a significant 3-4%
point improvement, suggesting this to be another helpful metric worth isolating in the future.

7.2 Additional Metrics
Because the five metrics that were isolated proved so useful, three new metrics were deve-
loped to analyze similar musical features. The metrics are as follows:

OpenPosition The percent of chords considered “open” by music theory standards. In an open chord, at
least one chord tone is skipped over between soprano and alto, and alto and tenor.
ClosedPosition The percent of chords considered “closed” by music theory standards. Generally, the
soprano, alto, and tenor are packed together as tightly as possible in a closed chord.
SameChordNumberDifferentInversion The percent of chord changes in which the chord number re-

mains the same but the inversion changes.

Like in previous classifications, the new metrics were used to classify music by all four com-
posers in groups of two, three, and four composers at a time. The new metrics were tested on
their own, in combination with the five-metric subset, and in combination with the original 19
metrics.
Interestingly, in most cases FPC was less accurate than previously in its classifications. Ac-
curacy actually suffered the most when these three metrics were used in conjunction with the
ive metrics that had previously performed best. Although these new metrics were aimed at
imitating the properties of the ones that had been so successful, they failed to add any value.

8 Conclusions
It has been shown how FPC uses metrics based on chord structure and chord movement as
input for three classification techniques. Furthermore, it has been demonstrated that conduct-
ing multiple randomized trials with test pieces of known classification allows the accuracy of
FPC's guesswork to be easily measured.
Analyzed results from multiple trials indicate FPC is most reliable when only a handful of
the most effective metrics are used. Root position, first inversion, second inversion, third in-
version, and secondary chords metrics have proven, at least here, to be the most important
factors in distinguishing between composers. A logical direction for future work would be to
test FPC's performance classifying four-part music by time period instead of composer.

References
[1] Ebcioglu K. An expert system for chorale harmonization. In: AAAI-86 Proceedings;
1986:784-788. htp://www.aaaipress.org/Papers/AAAI/1986/AAAI86-130.pdf .
[2] Nichols E, Morris D, Basu S. Data-driven exploration of musical chord sequences. In: Pro-
ceedings of the 14th international conference on Intelligent user interfaces (IUI'09).
ACM, New York, NY, USA; 2009:227-236. doi:10.1145/1502650.1502683.
[3] Doerfler G, Beck R. An approach to classifying four-part music. In: Proceedings of the
2013 International Conference on Image Processing, Computer Vision, and, Patern
Recognition (IPCV); 2013:787-794.