MAST melody dataset

The Musical Aptitude Standard Test (MAST) melody dataset is designed and shared to facilitate comparison of algorithms in the field of automatic music performance assessment.

The dataset includes melodic pattern reproduction performances by students (singing) together with the reference melodic pattern played on piano and assessment results. All recordings are collected during entrance exams (in 2015 and 2016) of Istanbul Technical University (ITU). 

The recordings are annotations in 2022 within the context of another research project supported by TUBITAK with grant number 121E198 as a part of the Scientific and Technological Research Projects Funding Program (1001). 

Annotations are performed via blind listening of individual performances after listening to a few renditions of  the melodic pattern by the experts. The files were presented in random order (after grouping samples in terms of melodic patterns) (i.e. the expert annotated all samples of a melodic pattern in random order and moved to the next group of samples for the next melodic pattern). 

A 4-level grading system was used during the evaluations of the data set; 
1-Completely Off, 2-Major Mistakes, 3-Minor Mistakes, and 4-Perfect. 

Annotations were carried by 3 experts; a professor of musicology who has taken part as a jury member in entrance 
exam auditions, and two music conservatory students of graduate-level programs. The last two annotators re-annotated
all collection a few months after the first annotation task. The csv file contains 5 annotations in 5 columns where
two of these columns are for the repeated annotations. To facilitate analysis, we also added columns that carry a flag
if all annotations match (column: 'fullAgree'), the score/grade all annotations agreed on (column: 'fullAgree_score') 
and the majority score.

In addition to audio files and annotations, two commonly used features are also included: 1) f0-series extracted using Crepe Pitch Tracker (https://github.com/marl/crepe), 2) chroma features computed using Librosa library's chroma_stft function (https://librosa.org/doc/main/generated/librosa.feature.chroma_stft.html)

The directory structure is:

annotations: Contains 5 distinct annotations by 3 experts in a scale 1-4 in a csv file. 

f0data_crepe: MASTmelody dataset, latest version (f0 series extracted using Crepe Pitch Tracker)

audioFiles: Audio files sampled at 8kHz

chroma: chroma features computed using Librosa library's chroma_stft function using its default settings except n_chroma which is set to 24. 

If you use this dataset, please refer to the following paper which announced its original version:

Bozkurt, B., Baysal, O., Yuret, D. A Dataset and Baseline System for Singing Voice Assessment, 13th Int. Symposium on Computer Music Multidisciplinary Research, Porto, Sept. 25-28, 2017.

@inproceedings{inproceedings,
author={Bozkurt, B., Baysal, O., Yuret, D.},
title={A Dataset and Baseline System for Singing Voice Assessment},
year={2017},
booktitle={13th Int. Symposium on Computer Music Multidisciplinary Research, CMMR 2017}
}