| dc.description.abstracteng | Vocal repertoires of nonhuman animals and especially of terrestrial mammals are often
characterized by their relatively small size of innate vocal types which can show considerable
variation in acoustic structure. To understand the proximate and ultimate causes
that shape the structure of acoustic communication systems in animals, an objective
characterization of the vocal repertoire of a given species is critical, as it provides the
foundation for comparative analyses among individuals, populations, and taxa.
The common approach to characterize vocal repertoires is by using unsupervised clustering
algorithms to identify call types and to define a repertoire's size. Progress in the
field has been hampered by a lack of standard in methodology which can lead to an arbitrary
decision about the size of a species' repertoire. To investigate whether this difficulty
is based on the used methodology or whether it is intrinsic to the acoustic structure of
a given repertoire, the major aim of my dissertation was to investigate and advance the
available methods in the field. To do so, I focused on three main aspects of a vocal
repertoire analysis: (1) how is the analysis affected by the input parameters, i.e. the
acoustic features that are used; (2) how can we quantify the acoustic variation within and
between different vocal types; (3) what is the impact of data set composition, i.e. the call
recordings that are being used in the analysis.
In the first part of my thesis, I re-analyzed recordings from wild chacma baboons (Papio
ursinus) to test the impact of the number and type of acoustic features that are included
in the analysis. To do this, I constructed data sets with the same 912 call exemplars
but with a varying number of acoustic features to describe these calls. To this end, I
had three data sets with 9, 38, and 118 acoustic features as well as a data set with 19
factors derived from a principal component analysis. By comparing and validating the
resulting classifications of two clustering algorithms, namely k-means and hierarchical
Ward's clustering, I could show that the data sets with a higher number of acoustic
features lead to better clustering results than data sets with only a few features. I further
showed that factors are not suited to cluster the chacma baboon's calls. None of the
applied clustering algorithms gave strong support to a specific cluster solution. Since there was substantial acoustic variation within and between the different call types, I
applied an approach based on fuzzy logic that we developed to describe the gradation
within vocal repertoires and which provides a quantitative description of the gradation
within the chacma baboon's repertoire.
To investigate the impact of potential evolutionary forces that shape a species' communication
system, comparative studies that quantify the differences in these systems
between different species are necessary. In the second part of my thesis, I strove towards
such a quantitative comparison by systematically comparing the vocal repertoire of the
chacma baboon with the vocal repertoire of the Barbary macaque, Macaca sylvanus. I
quantified the gradation within and between different call types of both species with an extended
version of the fuzzy clustering approach that was used to characterize the chacma
baboon's repertoire in the first part of this thesis. The analysis confirmed the findings of
previous studies by showing that the repertoire of the Barbary macaque exhibits a significant
larger amount of gradation within and between different call types. An important
aspect of this method is that it allows the quantification of gradation irrespective of the
number of call types by circumventing the problem to settle on one cluster solution when
several solutions are largely equivalent.
In the third part of my thesis, I investigated the influence of the data set composition
that is used for the analysis of vocal repertoires. Specifically, I was interested in the
effects of size- and arousal based differences in the recorded animals and their impact
on clustering results. The differences in body size and arousal were simulated with a
software-based model that simulates muscle characteristics of the larynx and vocal tract
anatomy. With this model I created pseudo repertoires of three distinct baboon call
types that varied in subglottic pressure levels (as a proxy of arousal-based differences)
and vocal fold and vocal tract characteristics (size-based differences). The preliminary
results show that whereas differences in subglottic pressure levels had a minor impact on
the characteristics of vocal repertoires and all three call types can be clearly separated
from each other, differences in body size can hamper classification and characterization
of call types.
In conclusion, I investigated several aspects that have to be taken into account when analyzing vocal repertoires. The composition of the data sets as well as the selection of
acoustic features that are used in the analysis can both have a profound effect on the
classification outcome and on cluster determination. To overcome the often arbitrary decision
about a species repertoire size I developed a method that is useful to describe the
gradation within and between different call types over several cluster solutions and therefore
circumvents the problem to settle on one specific solution. In addition, the method
allows a systematic comparison of different species' vocal repertoires, a prerequisite to
investigate potential driving forces in signal evolution. | de |