PDF performance difficulty datasets

Predicting Performance Difficulty from Piano Sheet Music Images

This dataset has been introduced in the paper by Ramoneda et al., presented at the 24th International Society for Music Information Retrieval Conference in Milan, Italy, 2023. The dataset addresses the gap in annotated corpora for performance difficulty assessment in piano music.

Dataset Summary

The following table provides an overview of the datasets mentioned:

The collection comprises several datasets of varying sizes and characteristics. The Mikrokosmos Difficulty (MK) dataset contains 147 piano pieces, each classified into one of 147 difficulty classes, with an average imbalance ratio (AIR) of 0.78, including around 49.2 thousand noteheads, all from a single composer. The Can I Play It? (CIPI) dataset is more extensive, with 652 pieces spread across 9 difficulty classes, a lower AIR of 0.33, over 1.1 million noteheads, and music from 29 different composers. The Pianostreet-difficulty (PS) set provides a significant leap in volume, with 2,816 pieces and 9 difficulty classes, an AIR of 0.24, and a substantial 7.2 million noteheads from 92 composers. The Freescores-difficulty (FS) assortment is the largest, offering 4,193 pieces in 5 difficulty classes, an AIR of 0.37, around 5.8 million noteheads, and a diverse range of 747 composers. Lastly, the Hidden Voices (HV) collection, though small with only 17 pieces, has 4 difficulty classes, a perfect AIR of 1, comprises 21.5 thousand noteheads, and features works from 10 black female composers. Each dataset varies not just in size but also in the diversity of music difficulty, composers, and the amount of musical information as quantified by noteheads.

Note: The AIR (average imbalance ratio) is the mean of the individual ratios between each difficulty class and the majority label in each collection.

New Collections

To complement the existing data, three new collections have been compiled:

1. Pianostreet-difficulty (PS) Set

  - Retrieved from Pianostreet, featuring 2,816 works with 9 difficulty levels.

2. Freescores-difficulty (FS) Assortment

  - Contains 4,193 pieces with 5 difficulty levels from the Freescores platform.

3. Hidden Voices (HV) Collection

  - A set of 17 pieces by black female composers with 4-level difficulty labels.

And it is composed also from previous collections:

Mikrokosmos Difficulty Dataset (MK)

The MK dataset, as proposed by Ramoneda et al., is the first publicly available collection and includes:

- 147 piano pieces by Béla Bartók.

- Graded by the composer.

- Presented in symbolic format.

Can I Play It? Dataset (CIPI)

The CIPI dataset is a more extensive collection featuring:

-  652 piano works in various symbolic formats.

- Annotated across 9 different difficulty levels.

Citation

Please cite and refer to  the following paper when using the dataset

@inproceedings{ramoneda2023predicting,  title={Predicting performance difficulty from piano sheet music images},  author={Ramoneda, Pedro and Valero-Mas, J. J. and Jeong, Dasaem and Serra, Xavier},  booktitle={Proceedings of the 24th International Society for Music Information Retrieval Conference (ISMIR '23)},  year={2023},  location={Milan, Italy}}

.