ï~~MELONET: Neural Networks that Learn HarmonyBased Melodic Variations
Johannes Feulner and Dominik Hornel
Institut fur Logik, Komplexitat und Deduktionssysteme,
Universitat Karlsruhe, D-76128 Karlsruhe, Germany
johannes@ira. uka. de
Abstract
MELONET, a system that can harmonize melodies and do melodic variations that are well bound to harmonic
contexts, is presented. MELONET comprises several neural networks that work on various subtasks. The topology reflects the fact that music is a phenomenon that occurs simultaneously at several time scales through neurones firing at different frequencies. This allows melodies to be learned more naturally since they can be treated
as sequences of (multi note) motifs instead of as sequences of single pitches. In order to let the motifs as they
evolve during the learning process (based on real music) correlate to musical intuition a representation scheme is
proposed that codes pitches in a distributed fashion relative to their harmonic context.
1 Introduction
There are several approaches to modelling melodies using neural networks. Most notably Todd
[1991], Mozer [1991] and Freisleben [1992] proposed models that are being based on the assumption
that melodies are sequences of pitches. However,
many melodies are composed out of little melodic
cells i.e. motifs. If a neural network is to learn
such melodies, its architecture should reflect the
type of melody under consideration.
Even where Freisleben goes on to model two
part pieces of music, he considers them as two simultaneously occurring pitch sequences, making it
difficult to capture the harmonic properties inherent
in any kind of tonal music. MELONET demonstrates that neural models can favourably model
melodies as sequences of harmony based motifs.
Combining the harmonic modelling of
HARMONET [Hild et al 92], [Feulner 93] together
with the SYSTEMA [Hornel 93] approach to motifs leads to a model that overcomes the aforementioned shortcomings.
2 Task Description
MELONET was built in order to learn melodic
variation from examples found in the literature. As
test cases the chorale variations of J. Pachelbel
I I I I I K I
A il
I VI W, Ad
X11
W, FW
I
I I 1 I
Figure 1
Partita super "Christus, der ist mein Leben" by Johann Pachelbel
a (top): Beginning of the four part chorale
b (bottom) Melodic variation of soprano voice
ICMC Proceedings 1994
121
Neural Nets
0