Flamenco Music Information Retrieval: Automatic Content-Based Description of Flamenco Music Collections.
Department of Applied Mathematics II
University of Seville
Supervisor: Prof. José-Miguel Díaz-Báñez
Abstract: Flamenco is a rich performance-oriented art music genre from Southern Spain, which attracts a growing community of aficionados around the globe. The constantly increasing number of digitally available flamenco recordings in music archives, video sharing platforms and online music services calls for the development of genre-specific description and analysis methods, capable of automatically indexing and examining these collections in a content-driven manner.
Music Information Retrieval is a multi-disciplinary research area dedicated to the automatic extraction of musical information from audio recordings and scores. Most existing approaches were however developed in the context of popular or classical music and do often not generalise well to non-Western music traditions, in particular when the underlying music theoretical assumptions do not hold for these genres. The specific characteristics and concepts of a music tradition can furthermore imply new computational challenges, for which no suitable methods exist.
This thesis addresses these current shortcomings of Music Information Retrieval by tackling several computational challenge which arise in the context of flamenco music. To this end, a number of contributions to the field are made in form of novel algorithms, comparative evaluations and data-driven studies, directed at various musical dimensions and encompassing several sub-areas of computer science, computational mathematics, statistics, optimisation and computational musicology. A particularity of flamenco, which immensely shapes the work presented in this thesis, is the absence of written scores. Consequently, computational approaches can solely rely on the direct analysis of raw audio recordings or automatically extracted transcriptions, and this restriction generates set of new computational challenges.
A key aspect of flamenco is the presence of reoccurring melodic templates, which are subject to heavy variation during performance. From a computational perspective, we identify three tasks related to this characteristic – melody classification, melody retrieval and melodic template extraction – which are addressed in this thesis. We furthermore approach the task of detecting repeated sung phrases in an unsupervised manner and explore the use of deep learning methods for image-based singer identification in flamenco videos and structural segmentation of flamenco recordings. Finally, we demonstrate in a data-driven corpus study, how automatic annotations can be mined to discover interesting correlations and gain insights into a largely undocumented genre.
Chapter 2: Melody Classification
Chapter 4: A Geometric Approach to Template Extraction
Chapter 4: A Multiple Sequence Alignment Approach to Template Extraction
Chapter 5: Discovering Melodic Repetition
Chapter 6: Image-Based Singer Identification in Flamenco Videos
Chapter 6: Structural Annotation Using a Multi-Label CNN.