This is the repository of SynMist, containing the synthesized mistake dataset as well as python scripts that generates the mistakes in a taxonomical way. simulate_mistakes.py and lowlvl.py contains functions regarding to the mid-level mistake scheduler and low-level deviation functions. region_classifier.py contains the simple texture \ technique region identifier.
python simulate_mistakes.py <input_midi_folder> <output_midi_folder> <run_id>
This processes all midi performance files in <input_midi_folder>, applies mistakes to them, and saves the files in <run_id>/<output_midi_folder>
The detailed analysis and annoation of individual mistakes from both datasets can be found here.
For accessing the burgmuller dataset, please refering to the original paper project page and find the download link.
For the augmented version of expert-novice dataset (containing transcribed MIDIs and error-annotation), please refer to the repository.
For interviews with piano teachers, we have put the evaluated samples into an online questionnaire. Their anonymized comments and rating are shown here.
Following the examples to adapt ASAP and AMAPS, create a similar one for whichever dataset as long as it is possible to extract a 1d array of timevalues representing the locations of the annotations.