The MDS Dataset

This is the Musical Distribution Shift (MDS) Dataset introduced and described in "Sound and Music Biases in Deep Music Transcription Models: A Systematic Analysis", published in EURASIP Journal on Audio, Speech, and Music Processing. The paper and its appendix contain various statistical analyses of the data. The code reproducing the analysis is available on GitHub.

Purpose

The MDS Dataset is not intended for training models. It is designed to serve as an evaluation benchmark to the Music Information Retrieval (MIR) community in research of Automatic Music Transcription (AMT) systems for solo piano music. As these systems have insofar been predominantly trained on the MAESTRO [1] dataset, the MDS dataset is curated to benchmark performance of such systems under various distribution shifts away from the MAESTRO data distribution.

Contents

The MDS dataset comprises of the following three subsets:

Genre: a subset of carefully sampled pieces from the ADL Piano MIDI dataset [2] -- recorded on our Disklavier piano -- spanning multiple genres beyond Classical music;

Random: a set of pseudo-randomly generated note sequences with carefully controlled polyphony and dynamics -- recorded on our Disklavier piano;

MAEtest: a subset of the MAESTRO [1] test set -- each MIDI performance is accompanied by two recordings: original audio recording from MAESTRO, and a new recording on our Disklavier piano -- reused from our prior work [3].

Structure

mds_dataset.zip/
│
├── Audio/
│ ├── 1_genre_dk/
│ ├── 2_random_dk/
│ ├── 3_maetest_dk/
│ └── 3_maetest_mae/
│
├── MIDI/
│ ├── 1_genre/
│ ├── 2_random/
│ └── 3_maetest/
│
└── metadata/
├── 1_genre_meta.csv
├── 2_random_meta.csv
└── 3_maetest_meta.csv

The names of subfolders in Audio and MIDI folders denote the relevant subsets. In Audio: the folders post-fixed with _dk contain audio recordings of our Disklavier piano, the _mae postfix marks the audio recordings originating from the MAESTRO dataset. The pairs of corresponding Audio and MIDI files have always matching base file names, up th the .mid and .wav extensions. The metadata files contain various annotations for the individual pieces of these subsets. Among other things, the 1_genre_meta.csv file contains paths to MIDI files in the originating ADL Piano MIDI (AMP) dataset under the path_in_adl_piano_midi column. Likewise, the 3_maetest_meta.csv file contains paths referring the selected pieces back to the originating MAESTRO [1] and as well as our mpteval [3] datasets under columns path_in_maestro and path_in_mpteval.

References

[1] C. Hawthorne, A. Stasyuk, A. Roberts, I. Simon, C.A. Huang, S. Dieleman, E. Elsen, J.H. Engel, D. Eck. In 7th International Conference on Learning Representations, (ICLR 2019), New Orleans, LA, USA, May 6-9, 2019. Enabling Factorized Piano Music Modeling and Generation with the MAESTRO Dataset (OpenReview.net, 2019). https://openreview.net/forum?id=r1lYRjC9F7

[2] L.N. Ferreira, L.H. Lelis, J. Whitehead. In Proceedings of the 16th AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, AIIDE’20 (2020). Computer-Generated Music for Tabletop Role-Playing Games. https://arxiv.org/abs/2008.07009

[3] P. Hu, L.S. Marták, C. Cancino-Chacón, G. Widmer. In Proceedings of the 25th International Society for Music Information Retrieval Conference, ISMIR 2024, San Francisco, CA, USA and online, November 10-14, 2024. Towards Musically Informed Evaluation of Piano Transcription Models. https://arxiv.org/abs/2406.08454