A Dynamic Representation Solution for Machine Learning-Aided Performance Technology

This paper illuminates some root causes of confusion about dynamic representation in music technology and introduces a system that addresses this problem to provide context-dependent dynamics for machine learning-aided performance. While terms used for dynamic representations like forte and mezzo-forte have been extant for centuries, the canon gives us no straight answer on how these terms must be applied to literal decibel ranges. The common conception that dynamic terms should be understood as context-dependent is ubiquitous and reasonably simple for most human musicians to grasp. This logic breaks down when applied to digital music technologies. At a fundamental level, these technologies define all musical parameters using discrete numbers, rather than with continuous data, making it impossible for these technologies to make context-dependent decisions. The authors give examples in which this lack of contextual inputs in music technology often leads musicians, composers, and producers to ignore dynamics altogether as a concern in their given practice. The authors then present a system that uses an adaptive process to maximize its ability to hear relevant audio events, and which establishes its own definition for context-dependent dynamics for situations involving music technologies. The authors also describe a generative program that uses these context-dependent dynamic systems in conjunction with a Markov model culled from a living performer-composer as a choice engine for new music improvisations.