Human-In-The-Loop Chord Progression Generator With Generative Adversarial Network

This paper proposes a framework for chord progression generation that uses human perception as a discriminator in a generative adversarial network model. By incorporating human perception as a discriminator, chords that do not exist in the distribution of training data can be treated as correct answers. However, since symbolic chord progressions cannot be calculated as numerical and continuous values, it is not clear how to compute the perturbation of a chord progression for human evaluation. Therefore, we formalized the perturbations of chord progressions in the embedding space based on existing music theory and devised a pairwise comparison task interface to collect human feedback for training the generative model. To verify the effectiveness of the proposed framework, we experimented to generate chord progressions based on crowd workers' evaluations as a discriminator and then asked musicians to evaluate the generated chord progressions. Consequently, the model generates significantly more natural and more diverse chord progressions, compared to the case where human perception is not incorporated.