These examples illustrate the diversity of sequence styles and drum kit styles available in the dataset. The full dataset contains 1,059 unique sequences synthesized using 43 drum kits for a total of 444 hours of examples, all with ground truth MIDI labels.
These are a few example questions from the Listening Test comparing our model (OaF-Drums) with others. Raters were given an example drum performance and asked to rate which synthesized transcription better captured the contents of the original. Transcriptions from the different models were synthesized using a standard SoundFont.
Notice how the addition of velocity (how hard the drum was hit) adds significantly to the perceptual similarity of OaF-Drums to the original.
Raters saw only pairwise comparisons, but for simplicity, here we just list all models at once.