M.Sc Thesis

M.Sc StudentHaviv Hakimi Shunit
SubjectDeep Neural Models for Jazz Improvisations
DepartmentDepartment of Computer Science
Supervisor PROF. Ran El-Yaniv
Full Thesis textFull thesis text - English Version


A major bottleneck in the evaluation of music generation is that music appreciation is a highly subjective matter. When considering an average appreciation as an evaluation metric, user studies can be helpful. The challenge of generating personalized content, however, has been examined only rarely in the literature.

 In this work, we address generation of personalized music and propose a novel pipeline for music generation that learns and optimizes user-specific musical taste. We focus on the task of symbol-based, monophonic, harmony-constrained jazz improvisations. Our personalization pipeline begins with BebopNet, a music language model trained on a corpus of jazz improvisations by Bebop giants.

BebopNet is able to generate improvisations based on any given chord progression.

 We then assemble a personalized dataset, labeled by a specific user, and train a user-specific preference metric that reflects this user's unique musical taste. Finally, we employ a personalized variant of beam-search with BebopNet to optimize the generated jazz improvisations for that user.

We present an extensive empirical study in which we apply this pipeline to extract individual models as implicitly defined by several human listeners. Our approach enables an objective examination of subjective personalized models whose performance is quantifiable. The results indicate that it is possible to model and optimize personal jazz preferences and offer a foundation for future research in

personalized generation of art.

 We further extend this generation method and present feature-guided improvisation generation that allows users to define a combination of musical features for BebopNet to optimize. For this purpose, we define numerous measurable musical features inspired by jazz music theory. Additionally, we inspect the possible benefits of using active learning to learn the user-specific preference metric, and the possibility of replacing our optimization step in the personalization pipeline with reinforcement

learning. We discuss our experiments, results, and conclusions.

 We also briefly discuss opportunities, challenges, and questions that arise from our work, including issues related to creativity.