Available at: https://digitalcommons.calpoly.edu/theses/3415
Date of Award
6-2026
Degree Name
MS in Computer Science
Department/Program
Computer Science
College
College of Engineering
Advisor
Maria Pantoja
Advisor Department
Computer Science
Advisor College
College of Engineering
Abstract
The recent explosion in digital music streaming has posed a unique problem. The sheer variety of listening options has made it more difficult than ever to explore the landscape of music. Current streaming platforms like Spotify and Apple Music rely on the technique of collaborative filtering, which creates genre and style bubbles, further hampering musical exploration. This creates a paradoxical situation in which the ability to explore the world of music has never been more possible, yet users consistently report not being able to branch out and discover new genres.
Content based recommendation– or the technique of recommending songs based on “musical similarity”– is an attractive alternative, but up until now it has been trammeled by the lack of a clear notion of similarity. Given that music is enormously complex, and dependent on such vast features as timbre, key signature, tempo, and genre– among many others– musical similarity recommendation research has largely been limited to objective measures of similarity, like cover song identification.
This thesis presents a novel system for musical exploration– Phonic– based on the “harmonic map,” a proprietary musical representation that captures essential characteristics of songs, regardless of key, genre or tempo. Several new, efficient algorithms for analyzing song similarity are presented, ranging from the deep, structural level to the gesture/motif level. We demonstrate the effectiveness of Phonic in bridging genre gaps through case studies of famous examples of musical similarity, including the viral “Four Chords” video by Axis of Awesome, in which over 20 famous songs were demonstrated to be highly structurally similar.
Aside from its value as a discovery and recommendation engine, Phonic allows musicologists to surface patterns and musical heritage across a dataset of nearly 6000 songs spanning from Jazz to metal, a process that was previously laborious and cognitively demanding– as research into the Pandora Genome project showed. Furthermore, the musical interpretability of Phonic makes it a natural fit for deep neural nets, which, while remarkably effective, suffer from complicated embedding structures that do not translate to the musical level.
Lastly, we demonstrate more theoretical results such as the combinatorial explosion of the song space, how musical ideas distribute across genres and time, and the most popular patterns in Western music.
Included in
Databases and Information Systems Commons, Systems Architecture Commons, Theory and Algorithms Commons