FMAKv2: A Dataset of Key and Mode Annotations for the Free Music Archive

We present the second version of FMAKv1, FMAKv2, a dataset containing song-level key and mode annotations of 5489 songs, spread across 17 genres, used in the paper STONE: Self-supervised Tonality Estimator, accpeted at ISMIR 2024.

About the version 1 of FMAK:

FMAKv1 is a an expert-labeled dataset for the evaluation of key detection, created by Stella Wong and Gandalf Hernandez. The FMAK metadata is made freely available for public use under a Creative Commons Attribution 4.0 International License.

The DOI of FMAK is 10.5281/zenodo.10719860 and the link is https://zenodo.org/records/10719860

The difference between FMAKv1 and FMAKv2 is the modification of around 200 songs' annotations. Other annotations remain the same, therefore annotated by Stella Wong. FMA track id and Spotify URI remain unchanged from FMAKv1.

Authors of FMAKv1 did not verify the modifications of annotations of FMAKv2 and should not be held liable for potential mislabelings in FMAKv2.

For each song, we provide information of:

• FMA track id (6 digits)
• Spotify URI (when available)
• Key and mode

All the audios in FMAKv2 are identical as FMAKv1, and can be downloaded from FMAKv1 repository.

If you use annotations from fmakv2, please cite the following paper:

@article{kong2024stone,
  title={STONE: Self-supervised Tonality Estimator},
  author={Kong, Yuexuan and Lostanlen, Vincent and Meseguer-Brocal, Gabriel and Wong, Stella and Lagrange, Mathieu and Hennequin, Romain},
  journal={International Society for Music Information Retrieval Conference (ISMIR 2024)},
  year={2024}
}