VirtuosoNet: A Hierarchical Attention RNN for Generating Expressive Piano Performance from Music Score

Abstract

Interpreting and performing of music score is a challenging task for computers. We propose a musically structured hierarchical attention network to generate expressive piano performance in MIDI format given symbolic music scores such as musicXML. The network takes a sequence of input features extracted from each note in the score and returns performance parameters for the note. The model can render various expressive elements in music performance, including tempo change, dynamics, micro-timing of individual notes, and pedal control.