piano-hand-assignment

Weights for piano-hand-assignment: assign every note of a piano MIDI file to the right or the left hand.

Files

File	Architecture	Parameters
lstm.pt	2-layer bidirectional LSTM, d=192	1.58M
transformer.pt	4-layer Transformer encoder, d=192, 4 heads	1.75M

Both are PyTorch checkpoints saved as {"arch": ..., "state": state_dict, "args": ...}. Inference averages the two probability distributions.

Usage

from piano_hand_assignment import HandAssigner

assigner = HandAssigner()                 # downloads these files on first run
assigner.assign_file("input.mid", "output.mid")

Training data

30,382 two-hand piano scores from PDMX (MuseScore-derived MusicXML), 20.4M notes. Hand labels come from the staff number in the score.

Evaluation

Test set	Accuracy	Hand-switch F1
Held-out PDMX split (2.0M notes)	94.30	0.918
Yamaha Disklavier pieces (unseen domain)	95.91	0.927