Composing Music With Neural Networks and Probabilistic Finite-State Machines

Composing Music with Neural Networks and
Probabilistic Finite-State Machines
Tomasz Oliwa and Markus Wagner
Artificial Intelligence Center, University of Georgia

Institute for Computer Science, University of Koblenz-Landau
[email protected], [email protected]
Abstract. In this paper, biological (human) music composition systems

based on Time Delay Neural Networks and Ward Nets and on a prob-
abilistic Finite-State Machine will be presented. The systems acquire
musical knowledge by inductive learning and are able to produce com-
plete musical scores for multiple instruments and actual music in the
MIDI format. The quality of our approaches is analyzed in objective and
subjective manner with existing techniques.
Key words: Biological Inspired Music, Music Composition, Representa-

tion Techniques, Comparative Analysis, Time Delay Neural Networks,
Finite State Machines, Inductive Learning
1 Introduction
Artificial music composition systems have been created in the past using various
paradigms. Approaches using Recurrent Neural Networks [7] and Long-Short
Term Memory (LSTM) Neural Networks [3] architectures to learn from a dataset
of music and to create new instances based on the learned information have been
taken as well as approaches with genetic algorithms. The latter ones focus on
semi-objective [9], i.e. combined computational and human evaluation of the
songs, or fully objective fitness functions [8] to generate new songs. Associative
Memories have also been tried [5] using a context-sensitive grammar.
Classical Algorithm-based automatic music composition systems, which aim
at following predefined rules to construct music, stand or fall by human imple-
mentation of the underlying algorithms, which leaves the cumbersome task to
derive sets of musical creation rules completely to the human designer. Another
approach is to modify existing melodies by applying specific noise function to
create new melodies [4], thus possibly reducing the dominance of the human
factor. Heuristic Search Algorithms like Genetic Algorithms, on the other side,
suffer from the fitness bottleneck [2][6], a gigantic, and in terms of music mostly
unusable, search space.
Our machine learning systems extract important features/key elements from
a dataset of music (created by humans) and are able to produce new song ma-
terial which inherits these ideas. They compose music based on the extracted
information gained by inductive learning. In both of our following approaches, we
use machine learning techniques for the feature selection of musical information
from the music database.
1.1 Music Background

Western Music can be defined as the chronology of notes, a note itself by its
pitch and length (which we consider as our atomic unit, and with a “note”, we
always mean the combined information of note length and note pitch).
In classical music theory, a piece of music is written in a specific musical scale,
which defines a subset from the set of all possible notes. Our machine learning
systems use an unambiguous mapping of notes to our internal representation,
which means that every note can be learned, regardless of its pitch or length.
The music which we considered as our music database for the melody where
19 classical and folk songs 1 . As an addition, we included 77 drum patterns in
a drum training dataset from Metallica’s rock song “Creeping Death” for the
finite state machine (FSM) approach to come up with a multi-instrument song.
2 The Finite-State Machine approach

2.1 Stochastic uniform sampling with accumulated probabilities
Our system parsed all songs from the song database and constructed multidi-
mensional hashmaps (our knowledge base), which contain the probability of all
possible note sequences and their successors. Figure 1 shows the underlying al-
gorithm. The use of the accumulated probabilities simplifies the search for the
successor(s), based on the idea from the stochastic uniform sampling method.
The learned subsequent structures are similar to the “grouping structures” [1].
3 Music Representation
The 19 songs in our database were written in abc language [10]. Conversion
from and to the abc language format from the MIDI format can be done using
the abcMIDI package 2 . MIDI (Musical Instrument Digital Interface)3 defines a
communications protocol for instruments and computers to transfer data.
3.1 Feature Extraction

An illustrative example of assigning distinct integer values to notes, pauses,
triplets etc. of the beginning of the song ”Claret and Oysters” by Andy Ander-
son is shown in Figure 3.1 in the corresponding abc language and our integer
representation. The richness of this representation stands in contrast to other
approaches with a more restricted search space (like [9], which has no pauses or
triplets). We found more diversified music because of our richer representation
and thus bigger search space.
1
http://abc.sourceforge.net/abcMIDI/original/MIDI.zip
2
http://abc.sourceforge.net/abcMIDI/
3
http://www.midi.org/
Fig. 1. Example of our algorithm, a Probabilistic FSM for a possible successor (se-
quence). In practice up to 4 base nodes and 10 successor notes are possible.
E2 G2 G2 G F 10, 18, 18, 20, 16,

G2 A2 B2 c2 18, 22, 26, 30,
d2 e2 g e d B 34, 38, 48, 40, 36, 28,
G2 G F G2 A2 18, 20, 16, 18, 22
Fig. 2. Beginning of ”Claret and Oysters” in the abc language (left) and our integer
representation (right)
4 The Neural Network approach

An artificial neural network (NN) is a layered network of artificial neurons and
is able to learn real-valued functions from examples (in this case, the subsequent
nodes). The temporal characteristics of music are exploited with a Time Delay
Neural Network (TDNN) architecture, where the input patterns are successively
delayed in time. The best result we had was using a Ward Net architecture, as
implemented in the NeuroShell 24 package and modified it into a TDNN-Ward
Net, which is shown in Figure 3. The entire song database was used as the
training set.
5 Experimental Results
The reference song “Claret and Oysters” and one song made by the FSM are
visualized in Figure 4 with our integer representation, with the integer numbers
on the x-axis and the time (sequential numbers) on the y-axis. As can be seen,
there exist repeating patterns in the graph, the “landscape” of the song shares
similar “hills”, for example the notes 50-57 and 80-87.
4
http://www.wardsystems.com/products.asp?p=neuroshell2
Fig. 3. TDNN illustration using the Ward Net, process of continuously generating four
notes, based on the last ten notes, with Ti indicating the i-th note of the song
Fig. 4. Visualization of the internal representation
5.1 Results with the FSM
Several smaller and longer learned patterns can be recognized by ear and eye,
not only from “Claret and Oysters” in Figure 4, but from different songs of the
entire database as well. It is noteworthy that the overall quality of the song
does not change over time (in contrast to the NN songs, described in the next
section). The beginning of this song is shown in Figure 5 as musical score.
5.2 Results with the NN
In Figure 6, the result from a TDNN-Ward Net, which was trained over 72000
epochs with an average error of 0.0006714 on the training set (the song “Claret
and Oysters”), is shown.
In the second half of the song, after a given starting seed, the NN is oscillating
more often between extreme notes than in the beginning of the song, it can not
predict the song any more. Including multiple songs to the knowledge database
did not significantly change the result. That means that even with a large number
of epochs the network architecture was not able to memorize the whole song.
Fig. 5. Sample song generated by the FSM in its musical score
Fig. 6. Sample song generated by the NN, trained on one song (internal representation)
6 Comparison with other approaches

In total, three songs (which feature drums violas, electric guitars played by FSM
and TDNN Ward Nets) created by our systems have been made public 5 and we
encourage the reader to form his/her own opinion about the music’s quality.
6.1 Musical Intervalls - Consonances and Dissonances

Statistics of music intervals, the difference in pitch between two notes, can be
provided. In Figure 7, it can be seen that a randomly generated song does not
have any preference for any interval. Although higher intervals were used by the
FSM occasionally, preferences for medium and low size intervals were observed.
“Claret and Oysters” has a preference for lower intervals as well.
7 Conclusion and Future Work

When listening to artificially composed music, critics tend to describe the cre-
ations as “..compositions only their mother could love..” [7]. While the local
contours normally make sense, the pieces are not musically coherent.
5
http://www.uni-koblenz.de/˜wagnermar/evomusart
Fig. 7. Frequencies of the intervals of a random generated song, a FSM song and
”Claret and Oysters”
Our two advantages are that (a) we learn key elements from analyzed songs
and features like distinct scale, arpeggios, musical style expressed through fre-
quent musical patterns and subpatterns are identified and used to create new
songs without an explicit human modeling and (b) a bias is induced to produce
short coherent sequences at once.
Future research need to recognize and use different coherent structures of a
song (like refrain, chorus, solo etc.) A newly composed song would then be struc-
tured by the learned musical structures and inherit features which are present
only in specific parts of songs, such as the refrain.
References
1. A. Baratè, G. Haus, and L. A. Ludovico. Music analysis and modeling through
petri nets. In CMMR, volume 3902 of Lecture Notes in Computer Science, pages
201–218. Springer, 2005.
2. J. A. Biles. Genjam: A genetic algorithm for generating jazz solos, June 15 1994.
3. D. Eck and J. Schmidhuber. Finding temporal structure in music: Blues improvi-
sation with lstm recurrent networks, 2002.
4. Y.-W. Jeon, I.-K. Lee, and J.-C. Yoon. Generating and modifying melody using
editable noise function. In CMMR, volume 3902 of Lecture Notes in Computer
Science, pages 164–168. Springer, 2005.
5. T. Kohonen. A self-learning musical grammar, or ”associative memory of the
second kind”. In IJCNN, volume I, pages I–1–I–6, Washington DC, 1989. IEEE.
6. E. R. Miranda and J. A. Biles, editors. Evolutionary Computer Music. Springer,
March 2007.
7. M. Mozer. Neural network music composition by prediction: Exploring the benefits
of psychoacoustic constraints and multi-scale processing, 1994.
8. J. Schoenberger. Genetic algorithms for musical composition with coherency
through genotype. Spring 2002.
9. M. Unehara and T. Onisawa. Construction of music composition system with
interactive genetic algorithm. Oct, 2003.
10. C. Walshaw. The abc notation system. http://abcnotation.org.uk, 1993.

Composing Music With Neural Networks and Probabilistic Finite-State Machines

Uploaded by

Copyright:

Available Formats

Composing Music With Neural Networks and Probabilistic Finite-State Machines

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Composing Music With Neural Networks and Probabilistic Finite-State Machines

Uploaded by

Copyright:

Available Formats

Composing Music with Neural Networks and

Probabilistic Finite-State Machines

Tomasz Oliwa and Markus Wagner

Artificial Intelligence Center, University of Georgia

Abstract. In this paper, biological (human) music composition systems

Key words: Biological Inspired Music, Music Composition, Representa-

1.1 Music Background

2 The Finite-State Machine approach

3.1 Feature Extraction

E2 G2 G2 G F 10, 18, 18, 20, 16,

4 The Neural Network approach

Fig. 4. Visualization of the internal representation

5.1 Results with the FSM

5.2 Results with the NN

6 Comparison with other approaches

6.1 Musical Intervalls - Consonances and Dissonances

7 Conclusion and Future Work

You might also like