ISMIR Paper: Unifying Drum Transcription and Stem Generation via Latent Diffusion

affige_yang · x · 2026-08-05

Automatic Drum Transcription (ADT) typically maps a music mixture directly to symbolic events, discarding useful acoustic stems for editing.

This upcoming ISMIR paper introduces a Separate-and-Detect architecture:

Evaluated on MDB Drums and ENST-Drums, the pipeline outperforms a strong U-Net separation baseline in overall transcription F1. It also beats an end-to-end ADT system on kick and snare F1 while providing editable separated audio stems.

Original post →

More from Research

Research channel →