Abstract
Discovering and exploring the underlying structure of multi-instrumentalmusic using learning-based approaches remains an open problem. We extend therecent MusicVAE model to represent multitrack polyphonic measures as vectors ina latent space. Our approach enables several useful operations such asgenerating plausible measures from scratch, interpolating between measures in amusically meaningful way, and manipulating specific musical attributes. We alsointroduce chord conditioning, which allows all of these operations to beperformed while keeping harmony fixed, and allows chords to be changed whilemaintaining musical "style". By generating a sequence of measures over apredefined chord progression, our model can produce music with convincinglong-term structure. We demonstrate that our latent space model makes itpossible to intuitively control and generate musical sequences with richinstrumentation (see https://goo.gl/s2N7dV for generated audio).