YM2413-MDB

Abstract

Game music takes a large part in shaping the overall atmosphere of a game. However, as the number of game maps and situations increases and characters become more diverse, there is a demand for creating game music more easily and diversely. In addition, game music is suitable for conditional music generation research because the music had clear composition intentions for specific game scenes. However, despite its significance in terms of industry and research, there is a lack of research data to generate game music. To resolve this problem, we introduce YM2413-MDB, a newly collected 1980s video game music score dataset. The dataset includes about 600 pieces of music from Sega and MSX PC games in the 80s using the YM2413 sound chip. The collected game music is played with a wider variety of instruments than the existing Nintendo game music dataset, NES-MDB. We collected the high-quality score music data by converting commands of the YM2413 sound chip and analyzed the generated music using the collected dataset with the Transformer model. After conducting case studies, quantitative and qualitative evaluations, we found that the model could create music that has the characteristics of game music in the 80s, with this small dataset. We plan to add high-level labels such as in-game events and scenes to the dataset, and will conduct a conditional music generation study using that labels as a future study.