Understanding Performance Limitations in Automatic Drum Transcription

Recent advancements in Automatic Drum Transcription (ADT) have improved overall transcription performance. However, state-of-the-art (SOTA) models still struggle with certain drum classes, particularly toms and cymbals, and the specific factors limiting their performance remain unclear. This paper addresses this gap by leveraging the Separate-Tracks-Annotate-Resynthesize Drums (STAR Drums) dataset to create multiple dataset versions that systematically eliminate potential performance constraints. We conduct experiments using three common ADT deep neural network (DNN) architectures to identify and quantify these limitations. For drum transcription in the presence of melodic instruments (DTM), the primary limiting factor is interference from melodic instruments and singing. Aside from this, performance improves by approximately five percent when training and testing use the same single drum kit, only strong onsets are present, or notes are not played simultaneously. For drum transcription of drum-only recordings (DTD), nearly error-free transcription is achieved when simultaneous onsets are removed. This confirms that overlapping drum hits are the main performance constraint. By identifying key ADT challenges, we provide insights to enhance SOTA models and improve overall transcription accuracy.