Creative Demands and Practical Limitations of Controllable Music Generation from the Human Perception Perspective

Abstract

Controllable music generation aims to allow creators to guide AI in music creation through intuitive means, but the existing systems are still far from achieving this goal. From the perspective of human perception, this study combines computer music education and practical experience to explore the actual demands of creators for a generation system in terms of timbre control, rhythm arrangement, emotional expression, and editability of results. The study found that current controllable music generation has three deficiencies: First, there is a semantic mismatch between the system's parameter settings and the auditory perception of creators, and the adjustable dimensions are difficult to align with intuitive judgments in music creation; Second, the response of the generation model to control instructions lacks stability and precision, and there are frequent deviations between the creative concept and the final output; Third, the existing tools do not support an iterative workflow of repeated listening and local modification in music production, making it difficult to further edit and refine the generated results. The above problems show that, in order to serve the creative needs of people in their work, controllable music generation should be based on their perception habits and working methods rather than technical performance optimisation.