This research addresses the challenge of automating the creation of rhythm game content, which is traditionally hand-crafted by human designers. The study focuses on the OSU! rhythm game, specifically the mania 4k mode, and proposes a deep learning-based framework to generate game content—known as Beatmaps—directly from audio files. The primary objective is to improve the accuracy of timestamp prediction and the naturalness of generated action sequences to better match human-designed standards.
The methodology employs a two-stage generation process: first, a timestamp generation model identifies the timing of actions, and second, an action type generation model determines the specific inputs required. To enhance performance, the researchers introduce a C-BLSTM (Convolutional Bidirectional Long Short-Term Memory) model, which effectively captures both past and future musical context. Additionally, a fuzzy label method is implemented to address data sparsity and class imbalance, transforming binary action labels into a continuous probability distribution based on Gaussian functions.
Key findings indicate that the proposed C-BLSTM model combined with fuzzy labels significantly improves timestamp prediction, raising the F-Score from 0.8159 to 0.8430. Furthermore, the study addresses the inherent subjectivity in rhythm game design, noting that even human-created Beatmaps for the same song can vary significantly. Human subject testing confirmed that the generated content is perceived as more natural than that produced by previous research methods, although it remains slightly less natural than professional human-crafted maps. The research demonstrates that integrating density-based difficulty levels and melody-informed feature extraction allows for the successful, automated generation of rhythm game content across varying difficulty tiers.
