Latent rhythm transformation of drum recordings
Audio examples and figures accompanying the paper.
Abstract
A method is proposed for rhythm style transfer of multitimbral drum recordings via conditioning a VAE on rhythm and timbral features. Modulation and estimation of latent parameters and a novel resequencing process for reconstruction loss result in an end-to-end transformation circumventing manual segmentation and alignment.
Audio Examples
Examples below contain source (original with timbre), target (rhythm), and transformation (model output) recordings. Figures depict waveforms, ADT conditioning* and attention head activations associated with examples. Download available in above link for more examples.
Example 2
Example 3
Example 4
Example 5
Example 6
Conference Information
Organization: Audio Engineering Society
Event: AES International Conference on Artificial Intelligence and Machine Learning for Audio
Date: September 8–10, 2025
Location: London, UK
BibTeX
@inproceedings{hockman2025latent,
title = {Latent rhythm transformation of drum recordings},
author = {Hockman, Jason and Drysdale, Jake},
booktitle = {Late-Breaking Demo at the AES International Conference on Artificial Intelligence and Machine Learning for Audio},
year = {2025},
address = {London, UK},
month = {September 8--10},
organization = {Audio Engineering Society}
}