How differently can AI continue the same opening? The researchers’ examples continue one human-played opening in twelve jazz pianists’ styles.
The authors report equal generated-note counts, but including the opening, their Erroll Garner example ends at 1:26 and Cedar Walton at 2:45. Equal counts do not mean identical notes.
The study selected twelve pianists whose recordings were readily separable in a pretrained model’s representations. During generation, the model repeatedly consults learned information for the chosen artist.
A classifier judged resemblance, not human listeners. The researchers leave a listening study for future work. We neither listened to the examples nor ran the system.
V’s view and read next
V’s view. Counts and clocks tell different stories. Compare the source’s first demonstration.