Relevance of musical features for cadence detection. Bigo, L., Feisthauer, L., Giraud, M., & Levé, F. In Proceedings of 19th International Conference on Music Information Retrieval, Paris, 2018.
Relevance of musical features for cadence detection [link]Paper  abstract   bibtex   
Cadences, as breaths in music, are felt by the listener or studied by the theorist by combining harmony, melody, texture and possibly other musical aspects. We formalize and discuss the significance of 44 cadential features, correlated with the occurrence of cadences in scores. These features describe properties at the arrival beat of a cadence and its surroundings, but also at other onsets heuristically identified to pinpoint chords preparing the cadence. The representation of each beat of the score as a vector of cadential features makes it possible to reformulate cadence detection as a classification task. An SVM classifier was run on two corpora from Bach and Haydn totaling 162 perfect authentic cadences and 70 half cadences. In these corpora, the classifier correctly identified more than 75pct of perfect authentic cadences and 50pct of half cadences, with low false positive rates. The experiment results are consistent with common knowledge that classification is more complex for half cadences than for authentic cadences.
@InProceedings{    bigo.ea2018-relevance,
    author       = {Bigo, Louis and Feisthauer, Laurent and Giraud, Mathieu
                   and Lev{\'{e}}, Florence},
    year         = {2018},
    title        = {Relevance of musical features for cadence detection},
    abstract     = {Cadences, as breaths in music, are felt by the listener
                   or studied by the theorist by combining harmony, melody,
                   texture and possibly other musical aspects. We formalize
                   and discuss the significance of 44 cadential features,
                   correlated with the occurrence of cadences in scores.
                   These features describe properties at the arrival beat of
                   a cadence and its surroundings, but also at other onsets
                   heuristically identified to pinpoint chords preparing the
                   cadence. The representation of each beat of the score as a
                   vector of cadential features makes it possible to
                   reformulate cadence detection as a classification task. An
                   SVM classifier was run on two corpora from Bach and Haydn
                   totaling 162 perfect authentic cadences and 70 half
                   cadences. In these corpora, the classifier correctly
                   identified more than 75pct of perfect authentic cadences
                   and 50pct of half cadences, with low false positive rates.
                   The experiment results are consistent with common
                   knowledge that classification is more complex for half
                   cadences than for authentic cadences.},
    address      = {Paris},
    booktitle    = {Proceedings of 19th International Conference on Music
                   Information Retrieval},
    keywords     = {music analysis with computers},
    mendeley-tags= {music analysis with computers},
    url          = {https://hal.archives-ouvertes.fr/hal-01801060/}
}

Downloads: 0