Extend pianist-style generation to structured musical material and broader settings

Develop artist-conditioned jazz piano generation that combines pianist identity embeddings with musical structure such as lead sheets, chord progressions, or melodic constraints; determine whether mixing artist embeddings enables interpolation between pianist styles; strengthen artist conditioning through classifier-free guidance; evaluate generated style through human listening studies modeled on the DownBeat Blindfold Test; and establish whether the approach generalizes beyond 12 artists and beyond solo piano.

Background

The paper introduces a gated cross-attention adapter that conditions a pretrained symbolic music transformer on pianist identity and evaluates stylistic fidelity primarily with pianist classifiers. The authors identify several unresolved extensions involving greater control over the musical material, interpolation between styles, stronger conditioning, perceptual evaluation, and broader coverage of artists and performance settings.

These directions are explicitly presented as open. In particular, classifier-based evaluations do not establish whether human listeners perceive the generated music as stylistically faithful, motivating a listening study; limiting the dataset to 12 selected solo pianists also leaves the generality of the approach unresolved.