Characterizing Difficulty Levels of Keyboard Music Scores

The proliferation online music scores for various instruments and musical styles can
be very positive for music learners, who would now witness an increase in score availability
for a variety of styles, which could give them autonomy over what styles and
songs to invest in learning. However, with the increase of data comes the question of
accessibility; how can we aid learners, often without the availability of teachers, to
identify good candidates of music scores to learn next given their current skill level?
To solve this problem computationally, there needs to be a method by which difficulty
can be measured computationally, and effectively. In this master thesis, our
goal is to re-visit computational music score difficulty analysis, which as a research
problem has been sidelined for some years. We apply the 2 approaches that have
been used within the research community on datasets made available to us through
the Trinity College London (TCL) examination board. One of these approaches is
based on symbolic feature extraction, and the other is based on probabilistic cost.
We discuss the strengths and weaknesses of each quantitatively and qualitatively.
Moreover, we devote an entire chapter to reviewing through textual content provided
by TCL such as information within their songbooks and syllabuses to serve
as a foundation for new feature suggestions. 20 features are suggested in addition
to the baseline set, and we examine their usefulness by checking if they can characterize
difficulty better than the baseline set alone. Despite the new feature having
some positive impact, there is still great room for improvement. Finally, after comparing
the feature extraction and the probabilistic difficulty approaches empirically,
we conclude there is no definitive answer on which is currently more robust. Each
approach has its strengths and weaknesses, which are discussed thoroughly, and
perhaps the best next step is to combine both approaches.