Abstract

Denoising diffusion probabilistic models, or diffusion models, have been successfully used for generating images, audio, and music. This work aims to investigate the potential of employing diffusion models to develop a motif composed by human composers to arouse specific emotions. To achieve this, a dataset consisting of melodies and their corresponding emotion label is constructed for training the diffusion model. The model is conditioned on the user-generated motif and a label displaying the desired emotion category, which opens up an opportunity for human composers to collaborate with computer technology in the field of music composition.