Coarse-To-Fine Framework For Music Generation via Generative Adversarial Networks

Abstract

Automatic music generation is highly related to Natural Language Processing (NLP). A current note in melody always depends on its context, just like a word in NLP. Yet the difference is that music is built upon a set of special chords that formulates the skeleton of the melody. To enhance automatic music generation, we propose a two-step adversarial procedure: Step 1 learns to generate chords via a chord generative adversarial networks (GANs); and step 2 trains a melody GAN to generate music for which the input is conditioned on the chords produced through the first step. Under such a two-step procedure, the chords generated in the first step formulate a basic framework of the music, which can theoretically and practically improve the performance of melody generation in the second step. Experiments demonstrate that such a cascading process is able to generate high-quality music samples with both acoustical and music theoretical guarantees.