Multiple Predominant Melody Annotations

This dataset addresses a central evaluation problem in predominant melody extraction: human experts may provide different, yet equally valid, reference annotations for the same jazz recording. Instead of relying on a single ground truth, the dataset makes multiple expert annotations available for selected jazz recordings.

The annotations support a more nuanced evaluation of melody estimation algorithms. They make it possible to study voice activity detection, fundamental frequency estimation, and annotator agreement under musically realistic conditions, especially in jazz contexts where melodic foreground and accompaniment can be ambiguous.

The dataset accompanies our ISMIR 2016 paper on evaluating multiple predominant melody annotations in jazz recordings.

Website: Multiple Predominant Melody Annotations for Jazz Recordings

Associated Publications

2016

Towards Evaluating Multiple Predominant Melody Annotations in Jazz Recordings

Stefan Balke, Jakob Abeßer, Jonathan Driedger, Christian Dittmar, and Meinard Müller

In Proceedings of the International Conference on Music Information Retrieval (ISMIR), 2016

Bib PDF

@inproceedings{BalkeADDM16_FundFreq_ISMIR,
  author = {Balke, Stefan and Abe{\ss}er, Jakob and Driedger, Jonathan and Dittmar, Christian and M{\"u}ller, Meinard},
  title = {Towards Evaluating Multiple Predominant Melody Annotations in Jazz Recordings},
  booktitle = {Proceedings of the International Conference on Music Information Retrieval ({ISMIR})},
  address = {New York, USA},
  year = {2016},
  pages = {246--252},
}