Time-based Chart Partitioning: Improving Local Coherency in Rhythm Game Chart Generation

Abstract

Creating rhythm game charts can be a time-intensive process involving manually synchronizing gameplay elements with audio. Procedural Content Generation (PCG) techniques have been applied to automate this task, but current methods are limited by imprecise onsets and locally incoherent note patterning. In this paper, we introduce Time-based Chart Partitioning (TCP), a chart generation framework that combines neural network-based onset detection, symbolic beat snapping, and a time partitioning pattern matching algorithm. We evaluate TCP on osu!mania, comparing it against AutoOsu, a state-of-the-art chart generation system, and find that TCP achieves higher onset precision and improved local pattern coherence.