ClaveNet: Generating Afro-Cuban Drum Patterns through Data Augmentation

Daniel Flores García, Hugo Flores García, Matteo Riondato

Abstract: We present ClaveNet: a generative MIDI model for Afro-Cuban percussion. We adapt the Monotonic Groove Transformer (MGT) —originally trained on the Groove MIDI Dataset (GMD)— to generate Afro-Cuban-influenced MIDI drum grooves. As Afro-Cuban drum MIDI data is scarce in the GMD and overall, we devise a data augmentation scheme to enrich MIDI percussion datasets with Afro-Cuban-inspired drum grooves by mixing examples with “seed patterns” rudimentary to Afro-Cuban percussion. To validate the effectiveness of our data augmentation algorithm at creating drum grooves infused with Afro-Cuban patterns, we trained MGT models on variants of the Groove MIDI Dataset augmented with our algorithm, and compared them to a baseline model trained on a non-augmented dataset. Our results show that MGT models trained with our augmented datasets are able to generate drum grooves whose rhythmic features are cumulatively closer to those from an evaluation set of real Afro-Cuban examples. We explore the effects of different hyperparameters to our system, discuss individual generated samples of selected models, and assess their faithfulness to Afro-Cuban styles. We hope this project fosters more research on developing music co-creation systems that encompass diverse musical styles outside those found in publicly available datasets.