Datasets:
| license: cc-by-4.0 | |
| task_categories: | |
| - audio-classification | |
| tags: | |
| - music | |
| - piano | |
| - midi | |
| - error-detection | |
| - disklavier | |
| pretty_name: ScoreInformed Piano Transcription Dataset | |
| size_categories: | |
| - n<1K | |
| # ScoreInformed: Disklavier Piano Performance MIDI | |
| 7 Disklavier performances with per-note mistake annotations. | |
| Each piece has 3 MIDI files: | |
| - `*_correct.mid` : Correctly played notes (matching the score) | |
| - `*_missed.mid` : Notes that should have been played but were not | |
| - `*_fa.mid` : Extra/false alarm notes (played but not in score) | |
| The actual performance = correct + fa. | |
| The intended score = correct + missed. | |
| ## Structure | |
| Disklavier_XX_{group}_{piece}_correct.mid | |
| Disklavier_XX_{group}_{piece}_missed.mid | |
| Disklavier_XX_{group}_{piece}_fa.mid | |
| ## Usage | |
| For token-level evaluation: only `fa` notes can be labeled (missed notes have no tokens). | |
| For frame-level evaluation: both `fa` and `missed` time regions indicate problems. | |
| ## Reference | |
| From "Score-informed transcription for automatic piano tutoring" (Benetos et al.). | |
| Part of [EVPMR](https://github.com/anusfoil/eval-piano-midi-repr) benchmark. | |