The paper proposes HyperGANStrument, a novel neural synthesizer that enhances the generation capability of GANStrument by introducing a pitch-invariant hypernetwork. This hypernetwork modulates the weights of a pre-trained GANStrument generator, improving the reconstruction fidelity and generation diversity of the generator. The model is lightweight and efficient, making it suitable for real-world applications as musical instruments. It enhances not only the generation of realistic and novel instrument timbres but also significantly improves the editability of synthesized sounds.
Publication date: 11 Jan 2024
Project Page: https://noto.li/MLIuBC
Paper: https://arxiv.org/pdf/2401.04558
