This is the augmented Expert-Novice dataset from paper Expert and Novice Evaluations of Piano Performances by Dr. Jiang, with original dataset in zenodo. This augmentation includes:

1. Transcription into MIDI files by the Bytedance high-resolution model (original wav not included in this repo).
   • Manually check each transcription file to ensure the onset are correct. Overall the quality is pretty good as the pieces are simple.
2. GPT4 cleaned instructor and rater feedback with question-answers in evaluation_qa.csv
3. Annotated the error regions in the same fashion as the Burgmuller set (Separate MIDI track with pitch 0 covering the error region), according to the rater feedback. New tracks are *-annot.mid, ones without annotations are free of salient mistake regions. The type of mistake is briefly discussed in mistake_notes.pdf

@inproceedings{morsi2024SimulatingContexts, 
    title = {Simulating and Validating Piano Performance Mistakes for Music Learning Contexts},
    publisher={Proceedings of the Sound and Music Computing Conference (SMC)}, 
    author={Morsi, Alia and Zhang, Huan and Maezawa, Akira and Dixon, Simon and Serra, Xavier}, 
    year={2024}
}