Accepted at the ICML 2026 Workshop on ML for Audio (Oral)
10 scores from ASAP-MAESTRO dataset and generated performances.
For CFG Bridge, we synthesize performances using deadpan scores of fixed duration (11 s).
For PianoKontext, we synthesize performances using variable-length deadpan scores and durations derived from target human performances.
| Score | Deadpan | Human | CFG Bridge (Baseline) | PianoKontext (Ours) |
|---|---|---|---|---|
| Liszt — Transcendental Etude no. 3 | ||||
| Schubert — Piano Sonata in A Major, D 664 | ||||
| Chopin — Etude Op. 10 No. 12 'Revolutionary' | ||||
| J. S. Bach — Prelude and Fugue no. 13 in F-sharp major, BWV 858 | ||||
| J. S. Bach — Prelude and Fugue in G-sharp minor, BWV 863 | ||||
| Beethoven — Piano Sonata No. 16 in G Major, Op. 31 No. 1 | ||||
| Beethoven — Piano Sonata No. 18 In E Flat Major, Op. 31, No. 3 | ||||
| Beethoven — Piano Sonata No. 3 in C Major, Op. 2 No. 3: II. Adagio | ||||
| Beethoven — Piano Sonata No. 18 In E Flat Major, Op. 31, No. 3: IV. Presto con fuoco | ||||
| Rachmaninoff — Prelude Op. 23 No. 6 in E flat major |
PianoKontext is capable of generating performances with predefined target duration. We provide an example of generated performances with durations set relative to the score.
| Condition | Deadpan | Human | Duration factor=0.8 | Duration factor=1 | Duration factor=1.2 |
|---|---|---|---|---|---|
| Debussy — Pour le piano |