An approach to classifying four-part music in multidimensional space - Emerging Trends in Image Processing, Computer Vision, and Pattern Recognition - page 23

StepwiseSopranoMovements The percent of chord changes in which the soprano moves by not more

than a major second.
RootPosition The percent of chords in root position (root in bass).
FirstInversion The percent of chords in first inversion (third in bass).
SecondInversion The percent of chords in second inversion (fifth in bass).
ThirdInversion The percent of chords in third inversion (seventh in bass).
Suspensions The percent of chord changes involving a suspended note that resolves to a chord tone.
SecondaryDominants The percent of chords borrowed from other keys that act as launch pads to
chords that do belong in the key (diatonic chords). FPC handles all “V-of” chords (i.e., V/ii, Viii, V/IV, V/V, V/vi) and
all “V 7 -of” chords except V7/IV. “V 7 -of” chords are simply recorded as “V-of” chords since they perform the same
function.

After all 19 metrics are computed per piece, a TXT file is produced for backup ( Figures 9
and 10 ) .

FIGURE 9 Flow chart for collecting Piece statistics.

FIGURE 10 Sample TXT file for a Bach chorale containing 19 metric values (percentages).

5 Collecting Classifier Statistics — Training Pieces
Only
The previous two steps — Parsing the XML and Collecting Piece Statistics — apply to the load-
ing of both training and test data. Step 3, however, applies to training data only. If the user has
clicked “Load Training XML,” FPC now begins the final step before it is ready to start classi-
fying test pieces: Collecting Classifier Statistics.
In the sections that follow , “classifier” with a lowercase “c” refers to the Piece object's string ield
while “Classifier” with a capital “C” refers to the Classifier object.

5.1 Approach
For each training piece belonging to the same classifier, a Classifier object is created. The mean
and standard deviation are computed for each metric from all the pieces of the classifier and
then stored in the Classifier object. For any piece, metrics outside three standard deviations of
the mean are thrown out, and the means and standard deviations are recalculated. Again, the