Learning a Latent Space of Multitrack Measures

Ian Simon, Adam Roberts, Colin Raffel, Jesse Engel, Curtis Hawthorne, Douglas Eck

Paper

https://arxiv.org/abs/1806.00195

Interpolations (Better Synths)

Latent space interpolations between two measures. These samples were rendered with manually-chosen synths faithful to the General MIDI programs. All other samples on this page were rendered with FluidSynth and a free SoundFont.

Interpolation 1
(two model samples)
Interpolation 2
(two model samples)
Interpolation 3
(Billie Jean to Don't Stop Believin')

8-Bar Interpolations (Figure 2)

Latent space interpolations between two sampled measures.


Attribute Transformations (Figure 3)

Latent space transformations via attribute vector arithmetic.
Original Measure
Reconstructed Measure
Increased Range
Strings Only
More Tracks


Chord Conditioning (Figure 4)

Several points in the latent space decoded over the chord progression: Cm → F → Bdim → G → C


Chord Progressions (Figure 5)

Several points in the latent space decoded over 8-bar chord progressions.

Am Dm G C
F Bdim E E
C G Bb F
Ab Eb D G
C C+ Am E
F Fm C G

Longer Interpolations

32-bar interpolations over repeating chord progressions.
C G Am F
Dm F Am G


Individual Measure Samples

Individual measures sampled from three models with varying numbers of free bits τ and concatenated.

τ = 0
τ = 64
τ = 256