Listening Examples

MIDI samples from our listening study. Within a clip, the three model versions share the same prompt and the same given notes; only what each model writes differs. Every clip of the study is listed, in a fixed order, without selection.

Playback is synthesized in your browser (GeneralUser GS SoundFont, about 8 MB, loaded once), with every note at the same loudness.

H-Net
hierarchical model with temporal realignment
H-Net w/o realignment
the same model without temporal realignment
Vanilla
flat Transformer on the same representation, with temporal realignment
Vanilla w/o realignment
the same flat Transformer without temporal realignment
Ground truth
the original clip

Highlight fades part of the piano roll: Condition is the prompt and the given notes, Output is what the model wrote (for the ground truth, the original notes in its place). Playback always includes both.

Highlight: