This academic article discusses a novel approach to multi-instrument music synthesis using performance conditioning. The authors propose a generative model that is conditioned on a specific performance and recording environment, allowing for better control of timbre and style. They implement this approach using state-of-the-art diffusion-based music generative models. The performance conditioning tool indicates the generative model to synthesize music with the style and timbre of specific instruments taken from specific performances. The authors’ prototype achieves high realism scores and offers novel control of timbre and style.
Publication date: 25 Sep 2023
Project Page: benadar293.github.io/midipm
Paper: https://arxiv.org/pdf/2309.12283
