|Abstract||Music has been playing an important role in the society. With the wide spread of digital and
mobile devices, the consumption of music has been increasing significantly. Therefore, the
demand for automatic processing and analyzing music with computational technologies has
also been pushed to a new ground. It has opened up opportunities for a variety of applications
which were not possible before. Structure, being a fundamental entity of music, is being
used in a lot of those applications to help people organize, navigate, and understand music.
It is also of great importance for the academic areas such as Computational Musicology
and Music Information Research (MIR). However, such application and research are mostly
focused on western music, while there is increasing demand for non-western music as well.
Beijing opera, being one of the most famous traditional operas in China, has influence all
over the world. But few efforts have been made in the study of Beijing opera music with a
computational perspective. In this work, we propose computational tools for the structure
analysis of Beijing opera arias.
Literature review of state of the art research on related topics are performed. Methodologies,
data, and challenges of these topics are analyzed. Then we proceed to the specific
tasks chosen for the structure analysis of Beijing opera arias on different levels. The first
task is the segmentation of singing, percussion and instrumental sections. Features have
been chosen to reflect the nature of the sound in these sections. An Support Vector Machine
(SVM) classifier is built using these features. Experiments on a 34 aria dataset have shown
the effectiveness of this method. The influence of features have also been analyzed.
The second task is lyrics-to-audio alignment. For this task, two datasets of a capella
singing have been created, with human annotations on the phoneme level. The study of
the datasets have been performed and have shown interesting characteristics of the lyrics
of Beijing opera singing. An approach based on Duration-aware Hidden Markov Model
(DHMM) has been proposed to address the challenge of long syllable durations in Beijing
opera arias. The evaluation results demonstrate the ability of the proposed method to tackle
the task of lyrics-to-audio alignment in the case of Beijing opera arias.
The thesis is concluded with main conclusions and a summary of the contributions of