Enhanced Automatic Drum Transcription via Drum Stem Source Separation

2025-09-29

AI Summary

Current automatic drum transcription suffers from coarse-grained class labeling (only five drum types) and the absence of velocity modeling, resulting in low MIDI realism. To address this, we propose a two-stage approach integrating drum source separation and transcription: first, we apply an open-source drum separation model to decompose mixed drum audio into individual sources, enabling fine-grained distinction between crash and ride cymbals; second, we feed the separated mono-source signals into the ADTOF transcription system and introduce a lightweight post-processing module to estimate MIDI velocity values. Our method expands drum classification from five to seven classes (eight in benchmark evaluation) while maintaining high detection accuracy and significantly enhancing MIDI expressiveness. Experiments demonstrate that the generated drum scores exhibit superior realism and practical utility in music information retrieval and production tasks. To the best of our knowledge, this is the first work to end-to-end integrate a general-purpose drum separation model with the ADTOF system while supporting velocity estimation.

Technology Category

• Planning, Routing, and Scheduling: Activity and Plan Recognition
• Computer Vision: Motion & Tracking
• Machine Learning: Mixture of Experts (MoE)

Application Category

• Graph Algorithms and Modeling for the Web: Representation, reconstruction, and subgraph or motif discovery in Web-related graphs
• Responsible Web: Machine-in-the-loop, human agency and autonomy
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Automatic Drum Transcription (ADT) remains a challenging task in MIR but recent advances allow accurate transcription of drum kits with up 5 classes - kick, snare, hi-hats, toms and cymbals - via the ADTOF package. In addition, several drum kit emph{stem} separation models in the open source community support separation for more than 6 stem classes, including distinct crash and ride cymbals. In this work we explore the benefits of combining these tools to improve the realism of drum transcriptions. We describe a simple post-processing step which expands the transcription output from five to seven classes and furthermore, we are able to estimate MIDI velocity values based on the separated stems. Our solution achieves strong performance when assessed against a baseline of 8-class drum transcription and produces realistic MIDI transcriptions suitable for MIR or music production tasks.

Problem

• Expanding drum transcription from five to seven distinct classes
• Estimating MIDI velocity values using separated drum stems
• Improving realism of drum transcriptions for MIR applications

Innovation

• Combines drum stem separation with transcription
• Expands transcription classes from five to seven
• Estimates MIDI velocity from separated stems