Understanding Jazz Improvisation Style with Explainable Music Performer Identification Models

Musical performances are complex phenomena. In the jazz genre, for instance, performers manipulate the harmony, melody, and rhythm of an underlying composition through improvisation. Understanding which elements of a jazz performance best reflect the style of a particular performer can be difficult, however, and may act as a barrier to learning and appreciating this music. Here we use machine learning to address this issue: specifically, we construct machine-learning models that learn to identify particular performers from their recordings, and then we interrogate their decision-making processes post hoc. We construct a series of such models, culminating in a multi-input convolutional neural network architecture that achieves state-of-the-art prediction accuracy (91\% success in identifying twenty famous jazz performers) with an interpretable structure that allows its predictions to be explained in terms of four fundamental musical domains (melody, harmony, rhythm, and dynamics). Our analysis of these models highlights the relative importances of each domain in distinguishing jazz pianists and uncovers a variety of melodic and harmonic patterns associated with each performer. We discuss implications for musicology and pedagogy unlocked by this work and by an accompanying web application, which we make accessible at https://huwcheston.github.io/ImprovID-app/index.html.

Recommended citation: Cheston H, Bance R, & Harrison PMC. Understanding Jazz Improvisation Style with Explainable Music Performer Identification Models. arXiv. 2025; arXiv:COMING_SOON [cs.SD]. COMING_SOON