|Abstract||In this thesis, we address the problem of automatic composer identification using
information from score-like representations. This task calls for computing features
that capture both specificities of the common practice in each composer’s
period plus individual specificities with regard to notes, chords, durations, etc.
Our goal is to find which are the singular characteristics of each composer and
therefore be able to identify them. This information can be useful for musicians as
well, in order to study which are the common characteristics between composers
and their relationship to the historical period. The particularity of this study is
related to the large number of composers used in it (106 composers) and the large
of works used as well (9876 works).
With a view to achieve this, we have designed features based on music theory,
in order to extract music properties from the scores. We have performed classification
experiments using the information from these features using Random Forest
and SVM classifiers. Using the classification results from theses experiments,
we have used feature selection in order to achieve better accuracies combining
information from different music concepts.
Using the best combinations, we have achieved accuracies up to 46,7% (17
composers), 39,0% (47 composers) and 27,6% (106 composers). In addition, we
present confusion matrices of the classifications and scalability graphs in terms
of accuracy of our approach, fixing the number of composers and the number of
works as well. Finally, we analyse the results and we propose further work related
to the extension of this study, changing parameters of the features.