Cross-Composer Dataset

The dataset presented on this website served as basis for studying the composer identification task for Western classical music recordings in the PhD dissertation [1]. It is compiled from commercial audio recordings, totalling 1100 tracks, where a track refers to the movement level of a piece. For 11 different composers, the dataset contains each 100 tracks comprising different musical forms, keys, and tempi. We provide annotations including composer- and piece-specific information as well as album information. Furthermore, chroma-based audio features and automatically computed chord labels are available.

See also Cross-Era Dataset

Details

For the experiments in [1], we were interested in the typical repertoire of Western classical music. Therefore, we focused on composers whose works frequently appear in concerts and on classical radio programs. To allow for a comparison to state-of-the-art algorithms, we considered an 11-composer setting similar to the MIREX Audio Classical Composer Identification scenario, an annual evaluation contest of the Music Information Retrieval (MIR) community. Our datasets comprises 100 pieces by each of the 11 composers as shown in the following table:

Class Tracks Composer Lifetime
01_bach 100 Bach, Johann Sebastian 1685–1750
02_beethoven 100 Beethoven, Ludwig van 1770–1827
03_brahms 100 Brahms, Johannes 1833–1897
04_dvorak 100 Dvorak, Antonin 1841–1904
05_handel 100 Handel, George Frideric 1685–1759
06_haydn 100 Haydn, Franz Joseph 1732–1809
07_mendelssohn 100 Mendelssohn Bartholdy, Felix 1809–1847
08_mozart 100 Mozart, Wolfgang Amadeus 1756–1791
09_rameau 100 Rameau, Jean–Philippe 1683–1764
10_schubert 100 Schubert, Franz 1797–1828
11_shostakovich 100 Shostakovich, Dmitri 1906–1975
Total 1100 11

We included a large variety of instrumentations including orchestral works, piano pieces, and solo concertos as well as compositions for choir, organ, and harpsichord. The pieces stem from commercial recordings on 94 different albums and are played by 68 different interpreters. The following table provides more detailed information about the instrumentations in the dataset.

Instruments Fraction of Pieces
Orchestra 38.7 %
Piano 38.6 %
Ensemble 19.5 %
Choir 6.6 %
Organ 6.3 %

Annotations

If you publish results obtained using these annotations, please cite [1].

We provide detailed annotations to the dataset comprising composer- and piece-related information (title, instrumentation) as well as performance-specific information (album name). To study the influence of the "artist effect" [3], we also provide a numerical artist identifier to be used as a filter. The annotations are given as a with delimiter "," (comma) comprising with the following fields:

Column Content Example
A Class 01_bach
B Filename CrossComp-0001_01_bach_ouverture_no._1_in_c_major_bwv_1066__boure_iii.mp3
C CrossComp-ID CrossComp-0001
D Composer Bach; Johann Sebastian
E Performer Cologne Chamber Orchestra; Muller-Bruhl
F Album BACH; J.S.: Orchestral Suites Nos. 1-4; BWV 1066-1069
G Artist_filter_no 1
H Instrumentation orchestra

Download Annotations

Chroma Features

If you publish results obtained using these features, please cite [1].

Since the dataset consists of commercial recordings, we cannot make the audio files publicly available. In order to allow reproducibility of some of our experiments, we provide chroma features of the pieces. We use the NNLS chroma algorithm as published in [5], which is freely available as a VAMP plugin. Concerning the parameters, we used a window size of 8192 samples and a step size of 4410 samples leading to a chromagram resolution of 10 Hz. We use the NNLS approximate transcription and no normalization. The features are provided as a .zip file containing one .csv file for each of the 11 classes. The columns are used in the following order where the first column is only filled when a new file begins:

Column Content Example
A Class/Filename "01_bach/CrossComp-0001_01_bach_ouverture_no._1_in_c_major_bwv_1066__boure_iii.mp3"
B Time (seconds) 0.10000
C-N Chroma A-G# 0.0006, 0.0193, 0.0019, 2.6477, 0.0101, 1.2693, 0.0268, 0.4974, 0.4409, 0.1174, 2.5127, 0.1873

Download NNLS Chroma Features (zip, 127 MB)

Chord Features

If you publish results obtained using these features, please cite [1].

We also provide chord sequences extracted from the audio files using the Chordino plugin based on NNLS chroma features [5]. The tool is part of the Chordino VAMP plugin. Concerning the parameters, we used a window size of 16384 samples and a step size of 4410 samples leading to a resolution of 10 Hz. We use the NNLS approximate transcription but do not make use of the bass chroma. The dictionary file for our chord analysis can be found here. The features are provided as a .zip file containing one .csv file for each of the 11 classes. The columns are used in the following order:

Column Content Example
A Class/Filename "01_bach/CrossComp-0001_01_bach_ouverture_no._1_in_c_major_bwv_1066__boure_iii.mp3"
B Time (seconds) 21.00000
C Chord Label "B_dim_dim7"

The first column is only filled when a new file begins. New lines are only written at chord changes. In our nomenclature, the triad type (maj, min, dim, aug) is specified after the first underscore. If existing, the seventh type (maj7, min7, dim7) is specified after the second underscore. "N" indicates No-Chord regions.

Download Chordino Chord Features (zip, 1 MB)

Literature

This is an accompanying website to the PhD thesis [1], where further details on the dataset, the annotation process, and the applications are discussed.

  1. Christof Weiß
    Computational Methods for Tonality-Based Style Analysis of Classical Music Audio Recordings
    PhD Thesis, Ilmenau University of Technology, 2017. PDF Presentation
    @PhdThesis{Weiss17_StyleAnalysis_PhD,
    author           = {Christof Wei{\ss}},
    title            = {Computational Methods for Tonality-Based Style Analysis of Classical Music Audio Recordings},
    school           = {Ilmenau University of Technology},
    address          = {Ilmenau, Germany},
    year             = {2017},
    url              = {http://www.db-thueringen.de/receive/dbt_mods_00032890},
    url-pdf          = {http://www.db-thueringen.de/servlets/MCRFileNodeServlet/dbt_derivate_00039054/ilm1-2017000293.pdf},
    url-presentation = {http://www.audiolabs-erlangen.de/fau/assistant/weiss/publications/2017_Weiss_PhD-Defense_TUIlmenau.pdf}
    }
  2. Matthias Mauch and Simon Dixon
    Approximate Note Transcription for the Improved Identification of Difficult Chords
    In Proceedings of the 11th International Society for Music Information Retrieval Conference (ISMIR): 135–140, 2010. PDF
    @inproceedings{MauchD10_DifficultChords_ISMIR,
    author    = {Matthias Mauch and Simon Dixon},
    title     = {Approximate Note Transcription for the Improved Identification of Difficult Chords},
    booktitle = {Proceedings of the 11th International Society for Music Information Retrieval Conference ({ISMIR})},
    year      = {2010},
    address   = {Utrecht, The Netherlands},
    pages     = {135--140},
    url-pdf   = {http://ismir2010.ismir.net/proceedings/ismir2010-25.pdf}
    }
  3. Arthur Flexer
    A Closer Look on Artist Filters for Musical Genre Classification
    In Proceedings of the 8th International Society for Music Information Retrieval Conference (ISMIR): 341–344, 2007. PDF
    @inproceedings{Flexer07_GenreClassification_ISMIR,
    author    = {Arthur Flexer},
    title     = {A Closer Look on Artist Filters for Musical Genre Classification},
    pages     = {341--344},
    booktitle = {Proceedings of the 8th International Society for Music Information Retrieval Conference ({ISMIR})},
    address   = {Vienna, Austria},
    year      = {2007},
    url-pdf   = {http://ismir2007.ismir.net/proceedings/ISMIR2007_p341_flexer.pdf}
    }

Acknowledgements

This dataset was created at Fraunhofer Institute for Digital Media Technology in Ilmenau, Germany as part of the PhD dissertation [1] by Christof Weiß, which was supported by the Foundation of German Business (Stiftung der Deutschen Wirtschaft). We thank Robert Gräfe for contributing to the dataset and to the annotations.