A data-driven exploration of rhythmic attributes and style in music

Matthew K. Prockup

doi:10.17918/etd-6896

Humans identify with three basic components of music: melody, harmony, and rhythm, in order to describe and differentiate songs. With these simple components, one can recognize higher level concepts such as the style and other expressive elements of a piece of music. In this thesis, I explore rhythmic components and their relationships to each other, to genre, and other geo-cultural factors (i.e., language) through data driven approaches using audio signals. Working in conjunction with Pandora, I employ a corpus of over 1 million expertly-labeled audio examples across many rhythmic styles and genres from their flagship Music Genome Project. Each song is labeled with more than 500 attributes of rhythm, instrumentation, timbre, and genre. In order to model the rhythmically related information from audio signals, I implement a set of novel and compact rhythm-specific acoustic features. They represent beat-level and meter-level information as well as elements of rhythmic variation and pulse stability. First, the acoustic features are used to predict the presence of human-annotated attributes of the meter and rhythmic feel (i.e., swing). Previous work has studied the general recognition of rhythmic styles in music audio signals, but few efforts have focused on the deconstruction and quantification of the foundational components of global rhythmic structures. Second, I focus on rhythm and its relationship to genre. Genre provides one of the most convenient categorizations of music, but it is often regarded as a poorly defined or largely subjective musical construct. I provide evidence that musical genres can to a large extent be objectively modeled via a combination of musical attributes, with rhythm playing a significant role. Finally, through a set of unsupervised machine learning experiments that employ both the human-labeled attributes and acoustic features, a set of low-dimensional, perceptually-motivated rhythm spaces is designed. These spaces provide grounded and visual insight into the relationships between rhythmic attributes and rhythmic styles. Most previous work strives to automatically predict a specific phenomena (i.e., genre) without a contextual understanding of why a label is applied. This work is motivated by largely the same idea, however, I aim to not only predict the phenomena but also understand the components used to construct it. This opens up the door to a more grounded and intuitive understanding of these components and how they interact to create the different styles of music we enjoy.

A data-driven exploration of rhythmic attributes and style in music

Files and links (1)

Abstract

Metrics

Details

A data-driven exploration of rhythmic attributes and style in music

Files and links (1)

Abstract

Metrics

Details

Drexel University Social media