Published June 12, 2023 | Version v1
Conference paper Open

Exploring Polyphonic Accompaniment Generation using Generative Adversarial Networks

  • 1. School of ECE, National Technical University of Athens
  • 2. ROR icon Institute for Language and Speech Processing
  • 3. ROR icon Athena Research and Innovation Center In Information Communication & Knowledge Technologies


Recently, various neural network architectures have shown capability of achieving compelling results in the field of automatic music generation. Motivated by this, in this work we design a generative framework that is structurally flexible and adaptable to different musical configurations and practices. At first, we examine the task of multi-track music generation without any human input, by modifying and proposing improvements to the MuseGAN architecture, an established GAN-based system, which we use as our baseline. Afterwards, we extend our developed framework to a cooperative human-AI setup for the generation of polyphonic accompaniments to user-defined tracks. We experiment with multiple structural variants of our model, and two different conditional instruments, namely piano and guitar. For both unconditional and conditional cases, we evaluate the produced samples objectively, using a set of widely used musical metrics, as well as subjectively, by conducting a listening test across 40 subjects. The experimental results, using the Lakh Pianoroll Dataset, reveal that our proposed modifications lead to improvements over the baseline from an auditory perspective in the unconditional case, and also provide useful insights about the properties of the produced music in the conditional setup, depending on the utilized configuration.



Files (1.7 MB)

Name Size Download all
1.7 MB Preview Download