CIVIT MASOT, MIGUEL, Drai-Zerbib, Veronique , CUADRADO MÉNDEZ, FRANCISCO JOSÉ, Escalona, Maria J.
No
Int. J. Hum.-Comput. Interact.
Article
Científica
23/03/2025
001451126900001
2-s2.0-105000894862
Artificial Intelligence (AI) music generation is a trending field, and many different generators are currently under development. However, no standardized evaluation method exists that can help researchers evaluate and compare AI-based music tools. To create a meta-methodology for AI music assessment based on user evaluation, that can be both standardized and deployed as a tailored implementation model adapted to the idiosyncrasies of specific generators and their intended applications, thereby helping future researchers draw comparisons between different systems. Two different decision trees/matrices are proposed to help researchers tailor their specific evaluation studies. As evaluation tools, the paper explores Likert and analytical hierarchy process (AHP) based surveys and emotional state estimations using facial action units, self-assessment, and physiological signals. A proof-of-concept study demonstrates the viability of the proposed tools for user-based AI music generation evaluation studies. A preference for audio music over symbolic music generation was observed, and this will require future research. The implementation of the proposed methodology and tools across the field will be helpful when comparing different systems in future research and to save time in the development of user-based studies. User-based evaluation studies are needed to prevent biases from passing into future iterations of AI music generators.
Automatic music generation; assisted music composition; artificial intelligence; user evaluation; evaluation studies