[3O5-OS-22a] OS-22

[3O5-OS-22a-02] Implementation of a Composing System with Time-Variable KANSEI Parameter Control Using Large Language Models

〇Ryosei Kawaguchi¹, Haruhiro Katayose¹ (1. Kwansei Gakuin University)

Keywords:Large Language Model, Music Composing, Parameter Control

One application of generative AI is to control generated content through parameters. For content with a temporal structure, such as music and stories, methods have been proposed to specify their progression using time-series semantic parameters. For example, the story generation system TaleBrush enables users to set the main character’s fortune level along the timeline. Similarly, the composing system SOUNDRAW allows users to adjust the intensity of a composition every four measures. However, in existing systems, the controllable parameters are fixed. To address this issue, we propose a music editing system that utilizes a large language model (LLM), allowing users to flexibly set time-series semantic parameters. As a result, we confirmed that users could control the musical progression based on specific attributes such as "robotic feel" or "strength." On the other hand, more ambiguous concepts like "romantic excitement" or "urban feel" were difficult to control as intended by users. This paper discusses the potential of this user interface and the remaining challenges.