TY - GEN
T1 - Four-Part Harmonization
T2 - 13th international Symposium on Computer Music Multidisciplinary Research, CMMR 2017
AU - Yamada, Tatsuro
AU - Kitahara, Tetsuro
AU - Arie, Hiroaki
AU - Ogata, Tetsuya
N1 - Funding Information:
Acknowledgments. This work has been supported by MEXT Grant-in-Aid (No. 16H05878, 16K16180, 16H01744, 26280089, 26240025, 16KT0136, 17H00749).
Publisher Copyright:
© 2018, Springer Nature Switzerland AG.
PY - 2018
Y1 - 2018
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
Y2 - 25 September 2017 through 28 September 2017
ER -