This page provides additional materials accompagnying our
Late-Breaking Result paper “Expressive Piano Performance Rendering from Unpaired Data”.
Overview
Full architecture of our proposed approach for expressive performance rendering.
The performance rendering model (in green) modifies the note properties of a score (timing, duration and velocity) in order to make an expressive rendition of the composition.
The model is trained without providing aligned pairs of scores and associated performances:
instead, its is trained to make performances realistic enough to deceive a discriminator fed
with real performances, while preserving the identity of the input score.
Visualization
The MIDI scores and performances are taken from the test set of the ASAP dataset and were synthesized into audio using the Arturia Piano V3 software.
The presented systems include:
Deadpan: the plain MIDI score rendered directly.
Human: a real human performance.
VirtuosoNet: a supervised model, using MusicXML input score files.
