Overview

This research line focuses on removing reverberation from drum recordings, aiming for cleaner transients, improved rhythmic definition, and better control in music production and audio restoration workflows. In practice, this is useful when drum recordings need further processing or reuse, but room tails mask the original impact and articulation of the hits.

Drum dereverberation is particularly challenging because the boundary between the “clean” signal and reverberant components is inherently ambiguous. Resonances from the shell and membrane, sympathetic vibrations across kit elements, and early reflections may be perceived as part of the instrument itself. For this reason, the research targets room-induced reverberation and explicitly addresses the distinction between early reflections and diffuse reverberant tails, while algorithmic/analog reverbs (e.g., plate, spring, gated effects) are treated as a different problem setting.

Methodologically, we adopt the cold diffusion scheme on stereo spectrograms (real/imaginary representations) at 44.1 kHz. In practical terms, the model learns a step-by-step restoration process: instead of trying to remove reverberation in one pass, it progressively refines the signal toward a drier and clearer version while preserving the character of the drums. The research includes multiple backbone architectures, such as UNet and Diffusion Transformer (DiT), in order to study both restoration quality and generalisation to unseen conditions (e.g., rooms, and reverberation profiles).