MusicVAE

MusicVAE implements several configurations of Magenta's variational autoencoder model, including melody and drum "loop" models, 4- and 16-bar "trio" models, and chord-conditioned "multi-track" models. In the demos below, we encode the input NoteSequences, and call sample() to obtain a new NoteSequence that matches them. Each section uses a different model checkpoint (melody, drums or trio) to generate the samples.

Drums OneHot

Interpolation (3x3 steps)

Input:

Result:

It took:

Sampling

It took:

Samples:

Drums NADE

Interpolation(3x3 steps)

Input:

Result:

It took:

Sampling

It took:

Samples:

Melody OneHot

Interpolation (5 steps)

Input:

Result:

It took:

Sampling

It took:

Samples:

Chord-Conditioned Melody OneHot

Interpolation (5 steps)

Input:

Result:

It took:

Sampling

It took:

Samples:

Melody 16-bar Conductor OneHot

Interpolation (5 steps)

Input:

Result:

It took:

Sampling

It took:

Samples:

Trio 2-bar (flat)

Reconstruction

Input:

Result:

It took:

Sampling

It took:

Samples:

Trio 4-bar (hierarchical)

Reconstruction

Input:

Result:

It took:

Sampling

It took:

Samples:

Total Leaked Memory: