Information Technology Reference
In-Depth Information
Perception-Based Segmentation (PB)
Musicological-Oriented Segmentation
(MO)
Melodies can be segmented accordingly to
theories on human perception. Listeners have
the ability to segment the unstructured auditory
stream into smaller units, which may correspond
to melodic phrases, motifs or musical gestures.
Even if listeners may disagree on the exact location
of boundaries between subsequent units, as high-
lighted by the perceptual experiment described
above, it is likely that perceptually-based units are
good descriptors of a document content because
they capture melodic information that appears to
be relevant for users. The ability of segmenting
the auditory stream may vary depending on the
level of musical training of listeners and their
knowledge of rules on music theory. Yet, a number
of strategies can be generalized for all listeners,
in particular the ones related to the detection of
clear changes in the melodic flow such as large
pitch intervals or note durations. This behavior
can be partially explained by the principles of
Gestalt psychology. Computational approaches
have been proposed by music theorists for the
automatic emulation of listener's behavior (Ten-
ney & Polansky, 1980). PB units do not overlap
and are based on information on note pitch and
duration of monophonic melodies.
Another approach to segmentation is based on
knowledge on music theory, in particular for
classical music. According with music theorists,
music is based on the combination of musical
structures (Lerdhal & Jackendoff, 1983; Narmour,
1990), even if its actual notation may lack of clear
representations of such structures. Yet, they can
be inferred by applying a number of rules, and
part of the analysis of compositions consists in
their identification. It is likely that the same ap-
proach can be extended to less structured music,
like popular or ethnic music. It is assumed that
a hierarchical relationship exists among music
structures, from musical phrases at the lower
level to movements at the higher level. MO units
are computed by analyzing the musical score,
applying rules for structure identification and
segmenting the score in units that correspond
to low-level structures. The computation of MO
units should be carried out using the global in-
formation of the score, but it has been proposed
an algorithm that uses only local information and
gave results comparable to more complex ones
(Cambouropoulos, 1997). Structures may overlap
in principle, but the current implementations do
not take into account this possibility.
Figure 5. Graphical representation of different automatic segmentation (from the top: PB, a statistical
approach not tested, and MO)
Search WWH ::
Custom Search