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

Tatsuro Yamada, Tetsuro Kitahara, Hiroaki Arie, Tetsuya Ogata*

*この研究の対応する著者

研究成果: Conference contribution

1 被引用数 (Scopus)

抄録

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.

本文言語English
ホスト出版物のタイトルMusic Technology with Swing - 13th International Symposium, CMMR 2017, Revised Selected Papers
編集者Matthew E.P. Davies, Mitsuko Aramaki, Richard Kronland-Martinet, Sølvi Ystad
出版社Springer Verlag
ページ213-225
ページ数13
ISBN(印刷版)9783030016913
DOI
出版ステータスPublished - 2018
イベント13th international Symposium on Computer Music Multidisciplinary Research, CMMR 2017 - Matosinhos, Portugal
継続期間: 2017 9 252017 9 28

出版物シリーズ

名前Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
11265 LNCS
ISSN(印刷版)0302-9743
ISSN(電子版)1611-3349

Other

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

ASJC Scopus subject areas

  • 理論的コンピュータサイエンス
  • コンピュータ サイエンス(全般)

フィンガープリント

「Four-Part Harmonization: Comparison of a Bayesian Network and a Recurrent Neural Network」の研究トピックを掘り下げます。これらがまとまってユニークなフィンガープリントを構成します。

引用スタイル