Stochastic processes and database-driven musicology
John Burgoyne · eScholarship@McGill (McGill) · 2012
For more than a decade, music information science and musicology have been at what Nicholas Cook has described as a 'moment of opportunity' for collaboration on database-driven musicology. The literature contains relatively few examples of mathematical tools that are suitable for analysing temporally structured data like music, however, and there are surprisingly few large databases of music that contain information at the semantic levels of interest to musicologists. This dissertation compiles a bibliography of the most important concepts from probability and statistics for analysing musical data, reviews how previous researchers have used statistics to study temporal relationships in music, and presents a new corpus of carefully curated chord labels from more than 1000 popular songs from the latter half of the twentieth century, as ranked by Billboard magazine's Hot 100 chart. The corpus is based on a careful sampling methodology that maintained cost efficiency while ensuring that the corpus is well suited to drawing conclusions about how harmonic practises may have evolved over time and to what extent they may have affected songs' popularity. This dissertation also introduces techniques new to the musicological community for analysing databases of this size and scope, most importantly the Dirichlet-multinomial distribution and constraint-based structure learning for causal Bayesian networks. The analysis confirms some common intuitions about harmonic practises in popular music and suggests several intriguing directions for further research.