Tuning Audio Encoders to be Piano Judges

Huan Zhang, Jinhua Liang, Simon Dixon, Queen Mary University of London

Pianism-Labeling Datasets (PLD)

A set of piano performance that labels different performance-related dimension and sourced from Youtube. The difficulty labeling (of Henle ratings) come from the CIPI dataset.

Expertise: novice_metadata advanced_metadata virtuoso_metadata

Difficulty: difficulty_metadata

Technique: technique_metadata

For the technique section, a subset of the recordings are sourced via self-playing (for specific technical passages that’s not available online) as well as editing, these wav files can be accessed here.

Expertise data demonstration

We acknowledge that it’s hard to align the performed pieces across three groups of expertise data, as the piece complexity naturally increase as people gets better playing piano. However, it’s still possible to find some repertoire from the dataset that can demonstrate the difference of playing across skill-levels, for example the Chopin Fantasie-Impromptu and Beethoven Moonlight Sonata, Mvt.1 in the following examples. Seeking a repertoire-overlapping subset with different expertise level will be a future work.