Chorale Harmonization in the Style of J.S. Bach, A Machine Learning Approach
Amanda Chilvers, Menno M. van Zaanen · Data Archiving and Networked Services (DANS) · 2008
In this article, we present a new approach to the task of automatic chorale harmonization. Taking symbolic musical transcripts as input, we use the TiMBL Machine Learning package to generate harmonizations of unseen chorale melodies. Extracting features from the soprano notes only, the system generates the other three melody lines in the chorale. We experiment with features denoting different properties of notes, such as pitch, note length, location in the bar or piece, and features of surrounding notes, but also with global features such as metre and key. The current system harmonizes 41.71% of the data exactly like J.S. Bach did (modulo octave).