Project Materials
This project explores how a listener can steer real-time music generation through repeated preference feedback while the music is still unfolding. Rather than relying on text alone, the system updates generation over time by comparing successive audio segments and using those comparisons to guide the next control decision.
This page collects example materials from the project. Each example consists of a video and two visualizations: a trajectory showing how the control state changes over time, and an estimated utility map showing which regions of the local control space appear more or less preferred based on the accumulated feedback.
In the prototype system, music is generated as a continuous stream. The stream is divided into successive 10-second segments, and after each new segment, the listener compares it with the immediately preceding one and indicates whether it is preferred, less preferred, or roughly similar.
These repeated comparisons are used to update the system’s estimate of local user preference and to choose the next control setting for subsequent generation. The examples below are intended to illustrate this interaction process and the evolving internal state of the system.
This example shows one run of the prototype system. The figures summarize the trajectory and estimated utility inferred from the interaction, and the video shows how feedback, trajectory, and estimated utility evolve over time together with the generated music.
More examples will be added to this page as the project develops. Each example will include a corresponding video and static visualizations to show how interaction, trajectory, and estimated preference structure evolve over time.