Information Technology Reference
In-Depth Information
an experIment on the
perceptIon of lexIcal unIts
of the styles of 4 well-known composers of tonal
Western music, namely Bach, Mozart, Beethoven
and Chopin. The criterion for the choice was to
have a good sampling of different kinds of melodic
structure, in which they were all present the dif-
ferent cues that listeners may use for segmenting
the melodies.
Each melodic excerpt was transcribed on a
separated sheet, without reporting composer
and composition names. The excerpts had dif-
ferent tempos, different time and key signatures,
which were all maintained in the transcriptions.
Moreover, the transcribed melodies had different
length, ranging from 7 to 26 bars and from 36 to
192 notes. The musicologist indicated the length
of each excerpt, which depended on the melodic
structure and on the length of the main theme.
Finally, in each sheet there were four lines for
comments, whether the subjects would like to
describe the reasons of a particular choice. The
complete list of the music works, from which ex-
cerpts were taken, is reported in Table 2. Another
short melodic excerpt was used as a graphical
example of the segmentation task.
A group of 17 subjects participated in the
experiment. Subjects were asked to perform the
segmentation task directly on the paper where
the music scores were printed. They were pro-
vided with the melodic excerpts plus one page
of instructions on their task. Instructions also
included a short explanation about the motiva-
tion of the research work and its application to
music retrieval. All subjects were professional
or semi-professional musicians. The choice of
including only musicians in the group is due to a
main consideration. A musician is able to relate
himself directly with the written score, without the
need of someone else's performance. This avoids
the possible bias in the recognition of melodic
contours given by an intermediate interpreta-
tion. Moreover, musicians are more familiar with
the concepts of phrases in music. Beside of the
segmentation task, subjects were required to give
some information about their background in music
It is normally assumed that the dimensions that
form the music flow can be divided in their lexical
units by listeners, depending on the characteristics
of the music structure (Lerdhal & Jackendoff,
1983; Narmour, 1990). This means that it is as-
sumed that listeners are able to single out one or
more dimensions of interest and to segment them.
Segmentation can be considered the process by
which listeners recognize boundaries of lexical
units, being able to recognize the presence of
boundaries according to a number of perceptu-
ally and culturally based strategies. Given these
assumptions, it is not clear the degree by which
listeners agree in recognizing the exact positions
of these boundaries. A similar situation applies
to other application domains of media segmenta-
tion, like text, image, and video segmentation. For
instance, experiments on manual text segmenta-
tion showed that subjects might have different
concepts of the meaning of a textual segment,
and thus recognize boundaries at different loca-
tions, or not agree at all about the presence of a
given boundary.
A perceptual study has been carried out on a
number of subjects to verify the degree of consis-
tency of manual melodic segmentation. Melodic
information has been used as the preferred dimen-
sion for the segmentation task, even if it has to be
noted that melody carries information also about
rhythm and harmony that can be inferred at least
by experienced musicians and musicologists. This
choice is motivated by the fact that most of the
approaches to music retrieval are based on the
melodic dimension, with few exceptions such as
exploiting harmonic information (Lavrenko &
Pickens, 2003).
experimental setting
An expert musicologist was asked to highlight 20
melodic excerpts that he considered representative
Search WWH ::
Custom Search