Computational methods for percussion music analysis: The Afro-Uruguayan Candombe drumming as a case study

Abstract

Most of the research conducted on information technologies applied to music has been largely limited to a few mainstream styles of the so-called ‘Western’ music. The resulting tools often do not generalize properly or cannot be easily extended to other music traditions. So, culture–specific approaches have been recently proposed as a way to build richer and more general computational models for music. This thesis work aims at contributing to the computer–aided study of rhythm, with the focus on percussion music and in the search of appropriate solutions from a culture–specific perspective by considering the Afro-Uruguayan candombe drumming as a case study. This is mainly motivated by its challenging rhythmic characteristics, troublesome for most of the existing analysis methods. In this way, it attempts to push ahead the boundaries of current music technologies. The thesis offers an overview of the historical, social and cultural context in which candombe drumming is embedded, along with a description of the rhythm. One of the specific contributions of the thesis is the creation of annotated datasets of candombe drumming suitable for computational rhythm analysis. Performances were purposely recorded, and received annotations of metrical information, location of onsets, and sections. A dataset of annotated recordings for beat and downbeat tracking was publicly released, and an audio-visual dataset of performances was obtained, which serves both documentary and research purposes. Part of the dissertation focused on the discovery and analysis of rhythmic patterns from audio recordings. A representation in the form of a map of rhythmic patterns based on spectral features was devised. The type of analyses that can be conducted with the proposed methods is illustrated with some experiments. The dissertation also systematically approached (to the best of our knowledge, for the first time) the study and characterization of the micro–rhythmical properties of candombe drumming. The findings suggest that micro–timing is a structural component of the rhythm, producing a sort of characteristic swing. The rest of the dissertation was devoted to the automatic inference and tracking of the metric structure from audio recordings. A supervised Bayesian scheme for rhythmic pattern tracking was proposed, of which a software implementation was publicly released. The results give additional evidence of the generalizability of the Bayesian approach to complex rhythms from different music traditions. Finally, the downbeat detection task was formulated as a data compression problem. This resulted in a novel method that proved to be effective for a large part of the dataset and opens up some interesting threads for future research.