Four-Part Harmonization: Comparison of a Bayesian Network and a Recurrent Neural Network

Tatsuro Yamada, Tetsuro Kitahara, Hiroaki Arie, Tetsuya Ogata

    Research output: Chapter in Book/Report/Conference proceedingConference contribution

    Abstract

    In this paper, we compare four-part harmonization produced using two different machine learning models: a Bayesian network (BN) and a recurrent neural network (RNN). Four-part harmonization is widely known as a fundamental problem in harmonization, and various methods, especially based on probabilistic models such as a hidden Markov model, a weighted finite-state transducer, and a BN, have been proposed. Recently, a method using an RNN has also been proposed. In this paper, we conducted an experiment on four-part harmonization using the same data with both a BN and RNN and investigated the differences in the results between the models. The results show that these models have different tendencies. For example, the BN’s harmonies have less dissonance but especially the bass melodies are monotonous, while the RNN’s harmonies have more dissonance but especially bass melodies are smoother.

    Original languageEnglish
    Title of host publicationMusic Technology with Swing - 13th International Symposium, CMMR 2017, Revised Selected Papers
    EditorsMatthew E.P. Davies, Mitsuko Aramaki, Richard Kronland-Martinet, Sølvi Ystad
    PublisherSpringer-Verlag
    Pages213-225
    Number of pages13
    ISBN (Print)9783030016913
    DOIs
    Publication statusPublished - 2018 Jan 1
    Event13th international Symposium on Computer Music Multidisciplinary Research, CMMR 2017 - Matosinhos, Portugal
    Duration: 2017 Sep 252017 Sep 28

    Publication series

    NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
    Volume11265 LNCS
    ISSN (Print)0302-9743
    ISSN (Electronic)1611-3349

    Other

    Other13th international Symposium on Computer Music Multidisciplinary Research, CMMR 2017
    CountryPortugal
    CityMatosinhos
    Period17/9/2517/9/28

    Fingerprint

    Recurrent neural networks
    Recurrent Neural Networks
    Bayesian networks
    Bayesian Networks
    Hidden Markov models
    Transducer
    Probabilistic Model
    Markov Model
    Learning systems
    Transducers
    Machine Learning
    Model
    Experiment
    Experiments

    Keywords

    • Bayesian network
    • Harmonization
    • LSTM-RNN
    • Machine composition

    ASJC Scopus subject areas

    • Theoretical Computer Science
    • Computer Science(all)

    Cite this

    Yamada, T., Kitahara, T., Arie, H., & Ogata, T. (2018). Four-Part Harmonization: Comparison of a Bayesian Network and a Recurrent Neural Network. In M. E. P. Davies, M. Aramaki, R. Kronland-Martinet, & S. Ystad (Eds.), Music Technology with Swing - 13th International Symposium, CMMR 2017, Revised Selected Papers (pp. 213-225). (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 11265 LNCS). Springer-Verlag. https://doi.org/10.1007/978-3-030-01692-0_15

    Four-Part Harmonization : Comparison of a Bayesian Network and a Recurrent Neural Network. / Yamada, Tatsuro; Kitahara, Tetsuro; Arie, Hiroaki; Ogata, Tetsuya.

    Music Technology with Swing - 13th International Symposium, CMMR 2017, Revised Selected Papers. ed. / Matthew E.P. Davies; Mitsuko Aramaki; Richard Kronland-Martinet; Sølvi Ystad. Springer-Verlag, 2018. p. 213-225 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 11265 LNCS).

    Research output: Chapter in Book/Report/Conference proceedingConference contribution

    Yamada, T, Kitahara, T, Arie, H & Ogata, T 2018, Four-Part Harmonization: Comparison of a Bayesian Network and a Recurrent Neural Network. in MEP Davies, M Aramaki, R Kronland-Martinet & S Ystad (eds), Music Technology with Swing - 13th International Symposium, CMMR 2017, Revised Selected Papers. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 11265 LNCS, Springer-Verlag, pp. 213-225, 13th international Symposium on Computer Music Multidisciplinary Research, CMMR 2017, Matosinhos, Portugal, 17/9/25. https://doi.org/10.1007/978-3-030-01692-0_15
    Yamada T, Kitahara T, Arie H, Ogata T. Four-Part Harmonization: Comparison of a Bayesian Network and a Recurrent Neural Network. In Davies MEP, Aramaki M, Kronland-Martinet R, Ystad S, editors, Music Technology with Swing - 13th International Symposium, CMMR 2017, Revised Selected Papers. Springer-Verlag. 2018. p. 213-225. (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)). https://doi.org/10.1007/978-3-030-01692-0_15
    Yamada, Tatsuro ; Kitahara, Tetsuro ; Arie, Hiroaki ; Ogata, Tetsuya. / Four-Part Harmonization : Comparison of a Bayesian Network and a Recurrent Neural Network. Music Technology with Swing - 13th International Symposium, CMMR 2017, Revised Selected Papers. editor / Matthew E.P. Davies ; Mitsuko Aramaki ; Richard Kronland-Martinet ; Sølvi Ystad. Springer-Verlag, 2018. pp. 213-225 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)).
    @inproceedings{da435ac1b74c43f09e30b3c258c57106,
    title = "Four-Part Harmonization: Comparison of a Bayesian Network and a Recurrent Neural Network",
    abstract = "In this paper, we compare four-part harmonization produced using two different machine learning models: a Bayesian network (BN) and a recurrent neural network (RNN). Four-part harmonization is widely known as a fundamental problem in harmonization, and various methods, especially based on probabilistic models such as a hidden Markov model, a weighted finite-state transducer, and a BN, have been proposed. Recently, a method using an RNN has also been proposed. In this paper, we conducted an experiment on four-part harmonization using the same data with both a BN and RNN and investigated the differences in the results between the models. The results show that these models have different tendencies. For example, the BN’s harmonies have less dissonance but especially the bass melodies are monotonous, while the RNN’s harmonies have more dissonance but especially bass melodies are smoother.",
    keywords = "Bayesian network, Harmonization, LSTM-RNN, Machine composition",
    author = "Tatsuro Yamada and Tetsuro Kitahara and Hiroaki Arie and Tetsuya Ogata",
    year = "2018",
    month = "1",
    day = "1",
    doi = "10.1007/978-3-030-01692-0_15",
    language = "English",
    isbn = "9783030016913",
    series = "Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)",
    publisher = "Springer-Verlag",
    pages = "213--225",
    editor = "Davies, {Matthew E.P.} and Mitsuko Aramaki and Richard Kronland-Martinet and S{\o}lvi Ystad",
    booktitle = "Music Technology with Swing - 13th International Symposium, CMMR 2017, Revised Selected Papers",

    }

    TY - GEN

    T1 - Four-Part Harmonization

    T2 - Comparison of a Bayesian Network and a Recurrent Neural Network

    AU - Yamada, Tatsuro

    AU - Kitahara, Tetsuro

    AU - Arie, Hiroaki

    AU - Ogata, Tetsuya

    PY - 2018/1/1

    Y1 - 2018/1/1

    N2 - In this paper, we compare four-part harmonization produced using two different machine learning models: a Bayesian network (BN) and a recurrent neural network (RNN). Four-part harmonization is widely known as a fundamental problem in harmonization, and various methods, especially based on probabilistic models such as a hidden Markov model, a weighted finite-state transducer, and a BN, have been proposed. Recently, a method using an RNN has also been proposed. In this paper, we conducted an experiment on four-part harmonization using the same data with both a BN and RNN and investigated the differences in the results between the models. The results show that these models have different tendencies. For example, the BN’s harmonies have less dissonance but especially the bass melodies are monotonous, while the RNN’s harmonies have more dissonance but especially bass melodies are smoother.

    AB - In this paper, we compare four-part harmonization produced using two different machine learning models: a Bayesian network (BN) and a recurrent neural network (RNN). Four-part harmonization is widely known as a fundamental problem in harmonization, and various methods, especially based on probabilistic models such as a hidden Markov model, a weighted finite-state transducer, and a BN, have been proposed. Recently, a method using an RNN has also been proposed. In this paper, we conducted an experiment on four-part harmonization using the same data with both a BN and RNN and investigated the differences in the results between the models. The results show that these models have different tendencies. For example, the BN’s harmonies have less dissonance but especially the bass melodies are monotonous, while the RNN’s harmonies have more dissonance but especially bass melodies are smoother.

    KW - Bayesian network

    KW - Harmonization

    KW - LSTM-RNN

    KW - Machine composition

    UR - http://www.scopus.com/inward/record.url?scp=85057426438&partnerID=8YFLogxK

    UR - http://www.scopus.com/inward/citedby.url?scp=85057426438&partnerID=8YFLogxK

    U2 - 10.1007/978-3-030-01692-0_15

    DO - 10.1007/978-3-030-01692-0_15

    M3 - Conference contribution

    AN - SCOPUS:85057426438

    SN - 9783030016913

    T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

    SP - 213

    EP - 225

    BT - Music Technology with Swing - 13th International Symposium, CMMR 2017, Revised Selected Papers

    A2 - Davies, Matthew E.P.

    A2 - Aramaki, Mitsuko

    A2 - Kronland-Martinet, Richard

    A2 - Ystad, Sølvi

    PB - Springer-Verlag

    ER -