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

[3O5-OS-22a-03] Co-generating songs through a naming game using multiple latent diffusion models

Koki Sakurai¹, 〇Haruto Uenoyama², Tadahiro Taniguchi^1,2, Akira Taniguchi¹ (1. Ritsumeikan University, 2. Kyoto University)

Keywords:Music composition, Latent diffusion model, Multi agent

In this study, we aim to generate music with different musical characteristics by jointly generating music by multiple AI agents. Specifically, we proposed the Metropolis-Hastings Music generation Game (MHMG), which integrates a latent diffusion model with the Metropolis-Hastings naming game, a framework that allows knowledge sharing among agents. In the experiment, two latent diffusion models trained on different genres of music (classical and jazz) were used as agents, and it was verified whether a music piece including the features of each genre could be generated. The experimental results showed that MHMG without fine-tuning retained the characteristics of each genre the best and produced high-quality music.