These are the soundfonts used to render the midi files. We use different soundfonts for training and testing to see whether the models generalize.
- FluidR3_GM (here we use the two presents ElectricPiano, YamahaGrandPiano)
- acoustic_piano_imis_1.sf2 (http://zenvoid.org/audio/)
- grand-piano-YDP-20160804.tar.bz2 (https://musescore.org/de/handbook/soundfonts-und-sfz-dateien)
Note that we only align the default score to one performance. However, because the performances are synthesized, this does not matter: they contain exactly the same note events, with exactly the same simultaneities, just the absolute timing changes different (because of tempo augmentations).
(This stops being fine with live MIDI that has mistakes.)