Reproduce “DrumGAN: Synthesis of Drum Sounds With Timbral Feature Conditioning Using Generative Adversarial Networks” on OpenResearch

About this paper

J. Nistal, S. Lattner, G. Richard

DRUMGAN, developed by researchers at Sony CSL Paris and Télécom Paris, synthesizes high-quality, diverse drum sounds using Progressive Growing Wasserstein GANs. The model provides intuitive timbral control by conditioning generation on continuous perceptual features, achieving superior audio fidelity and closer distributional similarity to real data compared to previous methods.

Key insights

• Combining Progressive Growing Wasserstein GANs with continuous perceptual conditioning allows for the generation of high-fidelity audio while providing fine-grained, intuitive control over timbral attributes.
• An auxiliary regression loss within the GAN discriminator is effective for enforcing feature coherence, guiding the generator to align output features with conditioning inputs without compromising overall audio fidelity or data distribution approximation.
• The learned latent space of DRUMGAN facilitates perceptually smooth and structured interpolations between sounds, indicating a meaningful organization of diverse drum timbres.

Summary by alphaXiv.