Turkish Makam Music Composition by Using Deep Learning Techniques

DOKUZ EYLÜL UNIVERSITY
GRADUATE SCHOOL OF NATURAL AND APPLIED SCIENCES
TURKISH MAKAM MUSIC COMPOSITION BY

USING DEEP LEARNING TECHNIQUES
by
İsmail Hakkı PARLAK
October, 2021
İZMİR
TURKISH MAKAM MUSIC COMPOSITION BY
USING DEEP LEARNING TECHNIQUES
A Thesis Submitted to the

Graduate School of Natural and Applied Sciences of Dokuz Eylül University
In Partial Fulfillment of the Requirements for the Degree of Doctor of
Philosophy in Computer Engineering
by
October, 2021
İZMİR
Ph.D. THESIS EXAMINATION RESULT FORM
We have read the thesis entitled “TURKISH MAKAM MUSIC COMPOSITION

BY USING DEEP LEARNING TECHNIQUES” completed by İSMAİL HAKKI
PARLAK under the supervision of PROF. DR. YALÇIN ÇEBİ and we certify that
in our opinion it is fully adequate, in scope and quality, as a thesis for the degree of
Doctor of Philosophy.
Prof. Dr. Yalçın ÇEBİ
Supervisor
Prof. Dr. Cihan IŞIKHAN Assoc. Prof. Dr. Derya BİRANT
Thesis Committee Member Thesis Committee Member
Assoc. Prof. Dr. Seda POSTALCIOĞLU Assoc. Prof. Dr. Ahmet Tuncay ERCAN
Examining Committee Member Examining Committee Member
Prof. Dr. Okan FISTIKOĞLU

Director
Graduate School of Natural and Applied Sciences
ii
ACKNOWLEDGMENT
I would like to thank my supervisor Prof. Dr. Yalçın ÇEBİ from the bottom of my
heart for his supervision, endeavor, and guidance, not just only throughout this study
but also in my personal life. I would like to offer my thanks and gratitude to Prof. Dr.
Cihan Işıkhan and Assoc. Prof. Dr. Derya Birant for their precious directives and
guidance. I was very privileged to have such skillful authorities in their respective
fields always guiding me positively throughout this work.
And finally, I would like to express my sincere gratitude to my wife, my son, and
my parents for their support, patience, and love. I could not be able to complete this
work without their invaluable support. Thank you.
iii
TURKISH MAKAM MUSIC COMPOSITION BY USING DEEP LEARNING
TECHNIQUES
ABSTRACT
Although music and other forms of fine arts have been accepted as a part of human
existence, people have developed various methods and algorithms throughout history
to create new creations in these domains. Especially in the last century, with the
invention of the computer and the exponential increase in its processing capacity, new
computational methods in artistic creativity have been tried and interesting results have
been obtained in related fields. Artificial composers, developed with Deep Learning
techniques, which is a sub-field of artificial intelligence, took part in interdisciplinary
studies and their artificial compositions began to be followed by those who are
interested in the subject with great curiosity and interest. However, the studies of
composing music using Deep Learning techniques were mostly performed on Western
Music, and Turkish Maqam Music remained untouched in this arena.
In the execution of this thesis, a system that can automatically compose Turkish
Makam Music using Deep Learning techniques and an easy-to-use web-browser based
graphical interface has been developed. The system, called Automatic Turkish Makam
Music Composer (ATMMC), takes 8 starting notes from its user and creates a
composition in Aksak or Düyek Usûls in one of Hicaz or Nihâvent Makams, depending
on the user's preference. Artificial compositions created by ATMMC can be stored in
the user’s computer to be opened with Mus2 application. Generated artificial
compositions were compared with the source data set according to various metrics and
it was seen that there was approximately 84% similarity between the source dataset
and artificial compositions. The developed system and its user interface are shared as
an open-source project.
Keywords: Deep learning, Turkish Makam Music, artificial intelligence, algorithmic

music composition, artificial composition.
iv
DERİN ÖĞRENME TEKNİKLERİ KULLANILARAK TÜRK MAKAM
MÜZİĞİ BESTELENMESİ
ÖZ
Müzik ve diğer güzel sanat dalları her ne kadar insan olmanın bir parçası olarak
kabul edilmişse de insanlar tarih boyunca bu alanlarda yeni yaratılar ortaya koymak
için çeşitli yöntem ve algoritmalar geliştirmişlerdir. Özellikle son yüzyılda
bilgisayarın icat edilmesi ve zaman içinde işlem kapasitesinin katlanarak artması ile
sanatsal yaratıcılıkta yeni yöntemler denenmiş ve ilgili alanlarda ilginç sonuçlar elde
edilmiştir. Yapay zekâ biliminin bir alt alanı olan derin öğrenme teknikleri ile
geliştirilen yapay besteciler, disiplinler arası çalışmalarda yer almış ve ortaya
çıkardıkları yapay besteler, konu ile ilgilenenler tarafından büyük bir merak ve ilgi ile
takip edilmeye başlanmıştır. Ancak derin öğrenme tekniklerini kullanarak müzik
besteleme çalışmaları çoğunlukla batı müziği üzerine yapılmış, Türk Makam Müziği
bu alanda yalnız kalmıştır.
Bu tez çalışması kapsamında derin öğrenme teknikleri kullanarak otomatik olarak

Türk Makam Müziği besteleyebilen bir sistem ve bu sistemin kolay kullanımı için web
tarayıcısında çalışan bir grafiksel arayüz geliştirilmiştir. Otomatik Türk Makam
Müziği Bestecisi adı verilen sistem, kullanıcısından 8 adet başlangıç notası alıp
kullanıcı tercihine göre Hicaz veya Nihâvent Makamlarından birinde Aksak veya
Düyek Usûllerde bir yaratı oluşturabilmektedir. Oluşturulan yapay beste Mus2 isimli
uygulama ile açılabilecek bir formatta oluşturulup kullanıcının bilgisayarına
kaydedilebilmektedir. Oluşturulan yapay besteler, çeşitli ölçütlere göre kaynak veri
seti ile karşılaştırılmış ve aralarında yaklaşık %84 benzerlik olduğu görülmüştür.
Geliştirilen sistem ve kullanıcı arayüzü açık kaynak olarak paylaşılmaktadır.
Anahtar kelimeler: Derin öğrenme, Türk Makam Müziği, yapay zekâ, algoritmik
besteleme, yapay besteleme.
v
CONTENTS
Page
Ph.D. THESIS EXAMINATION RESULT FORM .................................................. ii

ACKNOWLEDGEMENTS ...................................................................................... iii
ABSTRACT .............................................................................................................. iv
ÖZ ............................................................................................................................... v
LIST OF FIGURES.................................................................................................... vi
LIST OF TABLES.................................................................................................... vii
CHAPTER ONE - INTRODUCTION ..................................................................... 1
1.1 Overview .......................................................................................................... 1
1.2 Purpose ............................................................................................................. 2
1.3 Contribution to Literature................................................................................. 3
1.4 Organization of the Thesis ............................................................................... 3
CHAPTER TWO – LITERATURE REVIEW ....................................................... 5
2.1 Evolution of Turkish Music and Its Theory ..................................................... 5
2.2 Turkish Makam Music Concepts and Western Music ..................................... 8
2.3 Deep Learning and Automatic Music Composition ....................................... 11
2.4 Evaluation Methods for Artificial Music Compositions ................................ 13
vi
CHAPTER THREE - DATASET ........................................................................... 16
3.1 SymbTr Turkish Makam Music Symbolic Data Collection........................... 16
3.1.1 Statistical Review of SymbTr Dataset .................................................... 17
3.1.2 SymbTr Dataset N-gram Analysis .......................................................... 22
3.2 Dataset Preparation & Representation ........................................................... 25
CHAPTER FOUR – TECHNIQUES AND EVALUATION METHODS .......... 27
4.1 Long Short-Term Memory Networks ............................................................ 27
4.1.1 Activation Functions ............................................................................... 28
4.1.2 Long Short-Term Memory Cell Working Principles .............................. 29
4.2 Probability Distribution Analysis of Various Features .................................. 32
4.2.1 Pairwise Cross Validation ....................................................................... 34
4.2.2 Kernel Density Estimation ...................................................................... 34
4.2.3 Kullback-Leibler Divergence and Overlapped Area ............................... 35
4.2.4 Method Demonstration ............................................................................ 36
CHAPTER FIVE – AUTOMATIC COMPOSER ................................................ 40
5.1 Base Models ................................................................................................... 40
5.2 Specialist Models ........................................................................................... 42
5.3 Conductor Models .......................................................................................... 46
5.4 Full System Overview .................................................................................... 48
vii
CHAPTER SIX – GRAPHICAL USER INTERFACE FOR ATMMC ............. 51
6.1 Purpose of the Graphical User Interface ........................................................ 51
6.2 Graphical User Interface Decomposition ....................................................... 51
CHAPTER SEVEN – RESULTS & EVALUATION ........................................... 58
7.1 Subjective Evaluation ..................................................................................... 58
7.2 Objective Evaluation ...................................................................................... 61
7.2.1 Pitch Density Analysis ............................................................................ 62
7.2.2 ATMMC Detailed Analysis Results ....................................................... 64
CHAPTER EIGHT – CONCLUSION AND FUTURE WORK .......................... 72
8.1 Conclusion ...................................................................................................... 72
8.2 Future Work ................................................................................................... 73
REFERENCES ......................................................................................................... 74
viii
LIST OF FIGURES
Page
Figure 2.1 Comparison of TMM and WM divisions per whole tone interval. .......... 10
Figure 3.1 Inner structure of a Mu2 file from SymbTr dataset. ................................. 17
Figure 3.2 Percentages of Makams in SymbTr dataset. ............................................. 18
Figure 3.3 Scale of Hicaz Makam spanning 2 octaves. ............................................. 19
Figure 3.4 Illustration of note durations. .................................................................... 20
Figure 3.5 Hicaz merged octaves. .............................................................................. 20
Figure 3.6 Hicaz separated octaves. ........................................................................... 21
Figure 3.7 Nihâvent merged octaves. ......................................................................... 22
Figure 3.8 Nihâvent separated octaves....................................................................... 22
Figure 4.1 Internal structure of an LSTM cell (Olah, 2015). ..................................... 28
Figure 4.2 Sigmoid (left) and Tanh (right) function plots. ........................................ 29
Figure 4.3 Depiction of data flow through LSTM cell state. ..................................... 30
Figure 4.4 Data flow through the LSTM cell’s forget gate........................................ 30
Figure 4.5 LSTM cell state calculation. ..................................................................... 31
Figure 4.6 LSTM hidden state formation. .................................................................. 32
Figure 4.7 Workflow of Yang and Lerch’s method (Yang & Lerch, 2020). ............. 33
Figure 4.8 Kernel density estimation application (Vanderplas, 2007)....................... 35
Figure 4.9 Probability distributions for toy data set 1. ............................................... 37
Figure 4.10 Probability distributions for toy data set 2. ............................................. 39
Figure 5.1 Base Model’s layer architecture. .............................................................. 41
Figure 5.2 Base model diagram (a) and specialist model diagram (b)....................... 43
Figure 5.3 Operation scheme of Specialist Models. .................................................. 44
Figure 5.4 Specialist model prediction samples. ........................................................ 45
Figure 5.5 Conductor Model structural decomposition. ............................................ 47
Figure 5.6 Structural decomposition of ATMMC...................................................... 48
Figure 5.7 ATMMC composition process flowchart. ................................................ 50
Figure 6.1 Overview of ATMMC’s graphical user interface. .................................... 52
Figure 6.2 ATMMC graphical user interface left-hand side closeup......................... 53
Figure 6.3 ATMMC graphical user interface right-hand side closeup. ..................... 53
Figure 6.4 Before (a) and after (b) of automatic measure creation. ........................... 54
Figure 6.5 An erroneous measure. ............................................................................. 55
ix
Figure 6.6 Composition progress report pop-up. ....................................................... 56
Figure 6.7 Artificial composition download link. ...................................................... 57
Figure 7.1 Subjective evaluation responses for questions 1 and 2. ............................ 59
Figure 7.4 Subjective evaluation responses for question 7. ....................................... 61
Figure 7.5 Hicaz Makam pitch density graph. ........................................................... 63
Figure 7.6 Nihâvent Makam pitch density graph. ...................................................... 64
Figure 7.7 An example Hicaz composition by ATMMC. .......................................... 65
Figure 7.8 An example Nihâvent composition by ATMMC. .................................... 65
Figure 7.9 PDFs of pitch count (PC) feature.............................................................. 68
Figure 7.10 PDFs of note count (NC) feature. ........................................................... 69
Figure 7.11 PDFs of pitch count histogram (PCH) feature. ....................................... 69
Figure 7.12 PDFs of note length histogram (NLH) feature. ...................................... 70
Figure 7.13 PDFs of pitch range (PR) feature............................................................ 70
Figure 7.14 PDFs of average pitch interval (PI) feature. ........................................... 71
x
LIST OF TABLES
Page
Table 3.1 Tonal properties of Hicaz and Nihâvent Makams. .................................... 18
Table 3.2 4-gram frequencies for Hicaz Makam. ...................................................... 23
Table 3.3 4-gram frequencies for Nihâvent Makam. ................................................. 24
Table 3.4 5-gram frequencies for Hicaz Makam. ...................................................... 24
Table 3.5 5-gram frequencies for Nihâvent Makam. ................................................. 24
Table 4.1 Absolute measurement results for toy data 1. ............................................ 36
Table 4.2 Relative measurement results for toy data 1. ............................................. 37
Table 4.3 Absolute measurement results for toy data 2. ............................................ 38
Table 4.4 Relative measurement results for toy data 2. ............................................. 38
Table 7.1 Subjective evaluation responses for question 8. ........................................ 61
Table 7.2. Absolute metrics for Nihâvent Makam. .................................................... 66
Table 7.3. Absolute metrics for Hicaz Makam. ......................................................... 66
Table 7.4 Relative metrics for Hicaz and Nihâvent Makams. ................................... 67
xi
CHAPTER ONE
INTRODUCTION
1.1 Overview
Human beings have embraced artistic creativity as a faculty of their existence. The
human mind, which has evolved through time, has made aesthetic beauty a part of its
creations and has turned it into a symbol of advancement. Music, just like architecture,
literature, and other forms of fine arts, has become the stamp of nations and reflected
their history, culture, sentimental life, and collective experiences.
Throughout the history, great composers of various cultures either exploited

conventional methods or invented new techniques to create new music. Canonic
composition in the late 15th century (Burkholder et al., 2010) and Mozart’s dice music
(Parlak & Kösemen, 2018) are examples of such inventions in music composition.
Especially in the last century, with the invention and rapid development of computers,
the number of algorithmic artistic creation experiments exploded. Hiller and
Isaacson’s string quartet the “Illiac Suite” reserved its place in history books as the
first computer composition that utilized musical functions and rules (Sandred et al.,
2009). Then between 1967 - 72, comes Xenakis with his stochastic music composition
methods (Luque, 2009). As an alternative to stochastic and rule-based systems, around
the 90s Cope’s EMI (Experiments in Musical Intelligence) software came out (Cope,
1989). EMI was an Artificial Intelligence (AI) composer system that learned the rules
of music composition from the dataset it was trained on.
Towards the 2000s, the availability of relatively low-cost general-purpose graphics

processing units (GPGPUs) and vast amounts of digital data provided Deep Learning
(DL), which is a subset of Artificial Neural Networks (ANN) based AI technology,
with a chance to take a huge leap and making its way to being one of the most popular
branches of today’s computer science research arena. Since then, DL has been
continuing to grow rapidly, focusing on automatic classification and prediction as well
as artistic computation (Briot et al., 2017).
1
As an example of artificial artistic creation, Google’s “Deep Dream” attracted very
broad attention in 2015 by creating psychedelic artistic images. In addition, Aiva
Technologies’s film score composer (2017) and Google’s Bach Doodle (2019) can be
pointed out as top-tier examples of recent DL-based artificial music composers (Huang
et al., 2019).
Most DL-based artificial composers are built in the domain of western music forms
like classical western music, jazz, rock, pop, etc. Unfortunately, studies focusing on
composing Turkish Makam Music (TMM) by DL-driven AI are very scarce. Western
music and TMM are very different, and they can be very easily distinguished even by
an untrained ear. This difference brings about the necessity to apply specialized
techniques to DL-based TMM composition, rather than directly copy-pasting DL-
based western music composition techniques. In this thesis, a sketch of such TMM
specialized DL technique is given, and results are discussed.
1.2 Purpose
The essential incentive of this work is investigating the complex and delicate matter
of artificially composing Turkish Makam Music (TMM) and providing a preliminary
solution to it by implementing a Deep Learning (DL) based artificial composer system.
The described system is called Automatic Turkish Makam Music Composer
(ATMMC), and its purpose is composing new TMM songs similar to past
compositions in the TMM repertoire. ATMMC operates on the domains of Hicaz and
Nihâvent Makams and its rhythmic domain comprises Aksak (9/8) and Düyek (8/8)
Usûls.
ATMMC’s compositions may induce new ideas for TMM composers, may help
conservatory students to quickly create pieces to exercise their instruments, or be a
source of art and entertainment. Also, this research may build a foundation for other
researchers who are willing to work in this domain.
2
Another purpose of this study is to provide a graphical user interface (GUI) of
ATMMC to non-programmers thus making the system available to anyone interested
in it. Through a web application, users can make ATMMC compose new pieces and
download the resulting song to their computers in .mu2 format. All source code,
results, and training set are shared on GitHub (Parlak, 2020).
1.3 Contribution to Literature
Computational research on TMM is not as abundant as its western music

counterpart and is mostly trying to solve classification problems, or more focused on
music information retrieval (MIR) aspects. This research contributes to filling the gap
of DL-based automatic TMM composition.
Another contribution is the web-based GUI that is developed to utilize ATMMC.

There are tens of musical notation desktop and web applications for western music.
However, the number of available musical notation applications for TMM is just 2,
which are “Nota” and “Mus2” (Yarman, 2010). Moreover, web-based TMM notation
applications do not exist yet. This study lays a foundation for a web-based TMM
notation application.
ATMMC backend AI system, combined with the web-based GUI, offers a complete
solution to the problem of creating artificial compositions for TMM with DL. Before
this study, such solutions were completely absent.
1.4 Organization of the Thesis
In this chapter, an overview and purpose of the study, and contribution to the
literature are given. The remaining sections of the thesis are given in the following
paragraphs.
In Chapter 2, a comprehensive literature review is given covering the topics of

evolution of the TMM and its theory, the differences between TMM and western
3
music, usage of DL techniques on artificial music composition, and how artificially
composed music can be evaluated.
In Chapter 3, the dataset used in this study is described as well as the methods for
preparing it for model training. Also, the expansion of the used dataset for transfer
learning is briefly given.
In Chapter 4, the details of ATMMC are given in terms of selection of the type of
used neural network, how models are put together and connected for accomplishing
different tasks of artificial composition, and graphical representations of the system’s
operational dynamics.
In Chapter 5, the design and implementation of the GUI is given as well as the
screen captures of a use case. Also, a brief user’s manual is given in this chapter
alongside related graphics.
In Chapter 6, artificial compositions of the system and evaluation of ATMMC’s

artistic capabilities are given and discussed in detail.
Finally, the summary and conclusion of the thesis are given in Chapter 7. Also,
plans, enhancement suggestions for the future, and new ideas are included in this
section.
4
CHAPTER TWO
LITERATURE REVIEW
2.1 Evolution of Turkish Music and Its Theory
A brief history of Turkish music, its origins, evolution, and development as well as
the progression of its theory is given in this section for laying the foundation to
understand its characteristics. This is also important for understanding today’s
widespread TMM theories and the reason for the absence of a fully accepted formal
theory.
Tıraşçı (Tıraşçı, 2019) gave voice to the history of Turkish music, its cornerstones,
and its theoreticians. According to his work, before the Huns, Turks were located in
the northern and southern regions of The Tian Shan (Tengri Mountains). Around 2000
BC, the Altai Mountains and Siberia became two significant sites for Turks. At that
time, music was performed only by the religious men, who were known as Shamans,
for protection, spiritual and healing purposes.
At the age of Huns (3rd century AD), Turks used the pentatonic scale. With one of
the oldest Turkish instruments Kopuz, Huns’ music traveled to Europe and left its
traces, especially in the Balkans and Hungary region (Aydoğan & Özgür, 2015). Also,
at the age of Huns, in Chinese sources, it is recorded that Turks used drums to hearten
their warriors (Özkan, 2006). Later, music became militarized and military music was
institutionalized; thus, the repertoire and the musical activity grew in return. At the age
of Göktürks (6th century AD), Turks became neighbors with cultural centers such as
China, Persia, Byzantine, and India which led Turkish music to progress in terms of
genre and form. Also, at the age of Göktürks, music was a part of Khan’s (The leader)
assemblies. At these assemblies, musicians paid greater attention to the artistic aspect
of the performed music which led to the separation of art and folk music. In that era,
Turkish music got rid of being used only for religious purposes and started to appeal
to perceptions such as pleasure and aesthetics (Tıraşçı, 2019).
5
Uygur Turks (8th - 9th century AD) used the 7 tone - diatonic scale and later, they
began using the 12 - tone chromatic scale. The oldest Turkish musical note system
belongs to Uygur Turks, in which every musical note was represented by a symbol
from the Uygur alphabet (Tıraşçı, 2019). By adopting Manichaeism and Buddhism,
Uygurs’ religious music got richer (Aydoğan & Özgür, 2015). According to Tıraşçı
(Tıraşçı, 2019), before adopting Islam, Turkish music genres were:
• Religious music: Shamans used to utter sacred words musically. They used
drums and various percussive instruments to accompany their ceremonies.
• Tuğ music: This genre was performed during military and official ceremonies.
Various percussion, cymbals, and horn instruments were used. It is believed to
be the ancestor of Mehter music.
• Heroic, epic music: This type of music was revolving around epic and heroic
events and stories. It was used to increase the mood of the community and
soldiers. It also served as transferring historical knowledge to the future
generation.
• Toy music: This genre was performed by the palace’s musicians at important
formal events such as receiving ambassadors or accession to the throne.
• Daily life music: This genre was performed by the folk who expressed their
feelings of love, pain, sorrow, or longing.
• Yuğ music: This genre was performed after events of the death of beloved ones
to express sorrow and grief.
• Hunting music: When presidents were going out for hunting, Turks used to pitch
tents and sing sacred words for the hunt’s abundance. This custom continued
even after Turks adopted Islam.
After Karakhanids met with Islam, onwards 9th century, Turkish music heavily
interacted with Arabic / Islamic music and changed significantly. Arabic quarter-tone
ornaments fused with Turkish music and today’s Turkish Makam feel began to emerge
(Aydoğan & Özgür, 2015). Al-Kindi (9th century) was the first to write on music theory
amongst Muslim philosophers. He used Pythagorean ratios in his work (Bozkurt et al.,
6
2009). He related musical notes to celestial bodies and systematized Islamic music. He
inspired Al-Farabi and Avicenna (Ibn-Sina) (Tıraşçı, 2019).
Al-Farabi (10th century), studied music through the works of Grecian philosophers
and Al-Kindi. He corrected missing and erroneous theoretical information of Greek
philosophers and made exceptional studies on the physics of music. Safi al-Din
Urmavi (13th century) solved the problem of temporal representation in music with
his musical notation system. Before him, there was no representation for temporal
information of music. He placed numbers below musical notation and solved the issue
of temporal representation. He also invented two musical instruments called Nüzhe
and Muğni. He was first to use the term Edvar (cycle) to represent various scales such
as Uşşak, Neva, Rast, Hicaz, etc. In addition, he proposed his 17-tone Pythagorean
scale by revising Al-Kindi’s work (Yarman, 2007). Safi al-Din Urmavi is one of the
most remarkable figures in the history of TMM theory. His work was accepted as the
fundamental TMM theory until 16th century (Uygun, 2008). Following Safi al-Din
Urmavi’s work and concept of Edvar, Mahmud Shirazi (14th century) was one of the
first who used the term Makam. In his works, he mentioned 17 Makams and their
scales (Tıraşçı, 2019).
Until the 15th century, there was no distinction between Turkish, Persian, and
Arabic music. But after the 15th century, Turkish artistic and cultural thought began
to find its place within the new and emerging theoretical studies. Yusuf bin
Nazimuddin wrote Risale-i Musiki, which is the first Ottoman musical theoretics. He
believed that the movement of the Universe created harmonious sounds which form
the basis of music. Inheriting Al-Farabi’s thoughts, he defined 12 Makams which relate
to 12 zodiacal constellations (Tıraşçı, 2019). At the 18th century, another important
figure in Ottoman music scene, Kutbü'n Nâyî (Lead of Ney Players) Osman Dede
developed a new musical notation system and created various writings on music
theory. He composed pieces in a wide range of structure and form, and he was the first
to give titles to pieces in Peşrev form (Erguner, 2007).
7
In the 20th century, Anatolia was housing three different mindsets of musical
groups. The first group was supporting Western music, whereas the second one was
standing by traditional Turkish music. And the last group was trying to combine the
two. Up until the 20th century, the innovations that emerged in matters such as the
sound system, pitches, and Makams could not be based on solid foundations. Rauf
Yekta Bey studied the theory of Turkish music and laid solid foundations of the system
used today (Tıraşçı, 2019).
In the Turkish Republic’s early years, Atatürk attached great importance to music
studies and music education, and he made effort to carry Turkish music on a par with
the contemporary world’s requirements. Hüseyin Sadeddin Arel, who was a student of
Rauf Yekta Bey, used the symbols that denote the intervals we use in written music
today. With his colleagues Dr. Suphi Ezgi and Prof. Dr. Salih Murat Uzdilek, they
created the Arel-Ezgi-Uzdilek (AEU) system which divided an octave into 24 non-
equidistant intervals (Aydoğan & Özgür, 2015). AEU system is being used and thought
of in today’s conservatoires as the official model (Bozkurt et al., 2009). Some may
argue that Arel’s system depends on Western music theory rather than Turkish music,
or it may lack representation of the practical musical performance, but nonetheless, it
is the most widely used system in Turkey today (Tıraşçı, 2019).
2.2 Turkish Makam Music Concepts and Western Music
In addition to Classical Indian Music and Chinese Music, two other traditions that
can be considered as civilization music today are Western Music (WM), which is the
most studied one amongst all computationally, and Turkish Makam Music (TMM)
(Barkçin, 2019). Understanding the similarities and differences between TMM and
WM is important for deciding how computational composition techniques developed
for WM can be applied to TMM. The main elements that distinguish music, which is
one of the most important components of the inclusive phenomenon called culture, are
the sound system, rhythmic structure, and style used by that culture (Karaosmanoğlu,
2017). Therefore, sound systems, rhythmic structures, and styles of TMM and WM
should be studied comparatively.
8
In almost all fundamental properties of their dynamics, TMM and WM are different
than each other (Abidin et al., 2017). Firstly, and most obviously, being a tonal genre,
WM revolves around harmony and chord progressions, whereas TMM, which is
modal, focuses on melodies (Özkan, 2006). In other words, WM is polyphonic, i.e., it
is based on multiple harmonious distinct pitches playing at the same time, whereas in
TMM, all orchestral entities play the same pitch in unison or in different octaves with
small ornamental differences (Bozkurt et al., 2014; Şentürk & Serra, 2016), which is
called heterophony (Yarman, 2008).
TMM is very rich in creating melodies evoking a vast variety of emotions in a wide
variety of musical modes which are called Makams (Barkçin, 2019). Very simply put,
Makams are modal structures, where melodies begin to form around an initial note and
end around a final note (Ederer, 2011). Makams are built on scales and perhaps, the
most significant difference between TMM and WM is the method of acquiring the
pitches in their various scales. WM divides an octave into 12 equidistant intervals
(Şentürk & Chordia, 2011), which are created by dividing a whole step into two equal
fractions, i.e., semi-tones (Uyar et al., 2014). But according to Arel-Ezgi-Uzdilek
(AEU) theory (Arel, 1968), which is the official TMM theory today (Wright & Turabi,
2001), as illustrated in Figure 2.1, a whole tone is divided into 9 equidistant intervals
each of which are called Koma (Şentürk & Serra, 2016). AEU theory divides an octave
into 53 equidistant fractions (Karaosmanoğlu, 2017). Within these 53 pitches, 24 are
used to describe the practical TMM tuning system (Karaosmanoğlu, 2012). This
phenomenon is illustrated in Figure 2.1, where F0, F1, ..., F5 denote the used
frequencies within a whole tone interval in AEU theory. Even though a whole step is
divided into 9 equidistant Komas, not all of them are used in practice. From Figure
2.1, it can be seen that western semi-tones (Ws) divide a whole tone interval exactly
by 2, which corresponds to 4.5 Koma, but even though 4 Koma and 5 Koma intervals
correspond to predetermined pitches in AEU theory, 4.5 Koma, which is Western
music semi-tone does not have any counterpart in AEU theory.
9
At this point, it should be noted that a whole tone in AEU theory roughly
corresponds to 204 cents and a western whole tone corresponds to 200 cents, where 1
cent is 1/1200th fraction of an octave (Karaosmanoğlu, 2017). To simplify the topic,
the 4 cents difference is omitted in the figure. Also, depending on the frequency range,
a 4-cent difference is not always discernable by human beings (Zarate et al., 2012).
Figure 2.1 Comparison of TMM and WM divisions per whole tone interval.
TMM is taught orally for centuries (Uyar et al., 2014) through a system called Meşk
(Tüfekçi, 2014). In Meşk system, master teaches the details of the studied subject to
the disciple orally then examines and corrects disciple’s abilities closely when needed.
As given in Section 2.1, since the efforts for the literalization of TMM theory coincided
with the beginning of the republican period, it was late compared to WM in this regard.
So even today, there are several different TMM theory suggestions, ranging from 17
to 79 pitches per octave (Yarman, 2007). Even though AEU theory is considered to be
imperfect by modern TMM theoreticians, it brought ease to teach and learn TMM
formally and thus provided great benefit (Özkan, 2006).
A key term for understanding Makams is Seyir. The term Seyir can be defined as
the rules that regulate the circulation of melodies in the Makam (Güvençoğlu &
Özgelen, 2020). Two distinct Makams may have different Seyirs while having the
exact same tone series. However, regardless of melodic movement, a tone series is
named by its dominant or initial note in WM.
10
Apart from Makam, another important concept of TMM is Usûl. Usûl can
superficially be translated to English as the meter. Usûl describes the temporal
properties of music in TMM (Şentürk & Chordia, 2011). Usûls are comprised of
percussion stroke sequences with assorted velocities in a fixed amount of time
(Bozkurt et al., 2014). In TMM, percussions play the composition’s Usûl continuously.
Whereas, in Classical Western Music, performances require a conductor for organizing
the temporal accord (Barkçin, 2019). Usûls can be as simple as a 2-beat Nim Sofyan
or a very complex 124-beat Cihar (Gönül, 2015).
The final significant concept in TMM is Form. Form is used to describe the scheme
of musical piece’s parts and their arrangement (Şenocak, 2012). The two main
branches in TMM form scheme are instrumental and vocal music (Özkan, 2006).
These branches are further divided into sub-branches. For example, Şarkı is a member
of the vocal music form. Şarkı form consists of Zemin, Nakarat, and Meyan sections
in Zemin, Nakarat, Meyan, and Nakarat order (Güvençoğlu & Özgelen, 2020). Zemin,
which is like an introduction, is the section that shows the characteristics of Şarkı’s
Makam. Nakarat is the section where melodies are diversified and finalize with
Makam’s tonic pitch. In the Meyan section, melodies travel between different Makams
usually in higher registers (Tüfekçioğlu, 2019).
2.3 Deep Learning and Automatic Music Composition
Definition of music can be reduced to “succession of harmonious frequencies in

certain durations”. By this definition, it can be deduced that music can be represented
as time-series data. Therefore, it would not be unreasonable to expect good results
from artificial music composition studies with Deep Learning models that are
specialized to work on time series data (Hananoi et al., 2016; Oord et al., 2016).
Recurrent Neural Networks (RNN) are a family of Artificial Neural Networks

(ANN) that are used for processing sequential data (Goodfellow et al., 2016).
Unfortunately, vanilla RNNs may suffer vanishing or exploding gradients when
learning long-time dependencies. This brings forth the hindrance of learning features
11
of the training set by deeper layers of the ANN. To overcome the vanishing gradient
problem Choi et al. (Choi et al., 2016) propose Long Short-Term Memory (LSTM)
systems for music composition. They show that using char-RNN and word-RNNs
alongside LSTMs can produce satisfactory results in generating jazz chord
progressions.
Oord et al. (Oord et al., 2016) introduce WaveNet in their work, which is a novel
study generating raw audio. WaveNets can produce state of the art results when applied
to the text-to-speech field and is also able to generate novel and realistic audio
waveforms when trained with piano performances. They state that WaveNet is based
on PixelCNN and it operates on 16000 samples per second audio files.
In their study of euphonious easy to follow pop music melody generation, Shin et
al. (Shin et al., 2017) again deployed LSTM NNs. It is easy to see that most researchers
use LSTMs on the task of automatic music generation due to their success in
forecasting time series data (Xu, 2020). However, there are various studies that use
different types of NN systems such as Generative Adversarial Networks (GANs). In
their paper, Shuyu et al. (Li et al., 2019) described such system to compose artificial
melodies. Using Hierarchical Recurrent Neural Networks (HRNN) is another modern
approach for artificial music composition. Wu et al. (Wu et al., 2020) used three LSTM
networks to construct an HRNN for creating symbolic melodies.
There are popular, open-source and well-maintained software and frameworks for
Deep Learning and machine learning. One of the most popular libraries used for Deep
Learning studies is TensorFlow, which is a Python interface capable of running on
both CPU and GPU, as well as on a wide variety of heterogeneous systems (Abadi et
al., 2015). Keras is a high-level wrapper Python API for TensorFlow (Chollet, 2015).
Keras makes it easier to work on Deep Learning and enables fast experimentation.
12
2.4 Evaluation Methods for Artificial Music Compositions
It is obvious that artificially composed music should be evaluated to determine the

strengths and weaknesses of the proposed systems and to compare the work with other
similar studies, thus gaining ideas and insights for future studies with better results.
However, this is not a very straightforward task. It is not very easy to formulate and
measure what makes a piece of music pleasant. For this reason, researchers use two
distinct evaluation strategies: subjective evaluation and objective evaluation.
Subjective evaluation is made by exposing the outcomes of composer system to human
evaluators and gathering their feedback about the system. Whereas objective
evaluation is made by implementing a series of statistical procedures.
An example of subjective evaluation is carried on by Shin et al. (2017). They

resorted to human evaluators for evaluation of their artificial composer’s outcomes.
They obtained information on human evaluators’ musical backgrounds and then asked
evaluators to choose the most organic and well-structured sounding sample within a
set of various pieces. This set contained samples from both authors’ and similar
studies’ results. Then as an additional step, they performed a Turing Test (Turing,
1950) on participants to find out whether they could differentiate authors’ machine-
made musical compositions from human-made ones or not.
Most research resorting to subjective evaluation strategies conduct their evaluation

process through surveys. Liang, Gotham, Johnson, & Shotton (2017) evaluated their
study by means of an online survey. In their survey, they collected participants’ age
and level of musical expertise information. Then they presented various synthetic, and
human-made music to participants without telling them their origin and asked them to
determine which is synthetic and which is organic. Another example is the study of
Chu, Urtasun & Fidler (2016). They conducted a survey amongst 27 participants. They
asked participants to compare their results with the results of Google’s Magenta (Brain
Team, 2020). They also collected participants’ commentaries on the reasoning behind
their decisions.
13
For objective evaluation there are different strategies. Marinescu (2019) performed
experiments with different types of neural networks and network configurations. To
compare and evaluate different generative models, they investigated training loss
values and validation accuracy percentages.
Trying to put forward a standard to objective evaluation of melody-based artificial

composers, Yang and Lerch (Yang & Lerch, 2020) propose a comprehensive method
comprising a collection of metrics. In their paper, they suggest calculating sets of pitch
and duration related features within and in-between training dataset and artificially
composed songs. They denote that, if a metric is only computed within a single set, it
is labeled as “absolute”. Absolute metrics deliver information related to the set from
which it is computed. They also propose “relative” metrics, which are acquired by
comparing training sets and generated sets. They suggest computing several pitch-
related and duration-related metrics. They suggest computing the below metrics for
extracting pitch-related features of the studied sets:
• Pitch Count (PC): The total number of distinct pitches disregarding duration
information per song (sample).
• Pitch Class Histogram (PCH): Histogram of pitches without octave information.
For example, Do4#4 and Do5#4 are accepted as equivalent pitches. The
computed histogram should be normalized.
• Pitch Class Transition Matrix (PCTM): Octave independent, transition matrix of
pitches. Again, duration information is disregarded, and the computed matrix
should be normalized.
• Pitch Range (PR): The spread between the highest and the lowest pitches per
sample.
• Average Pitch Interval (PI): Total pitch distances between successive notes per
sample divided by the total number of pitches.
Yang and Lerch also suggest computing rhythm-based features for obtaining further
information about studied datasets. These metrics are described as follows:
14
• Note Count (NC): Total number of distinct durations disregarding pitch
information per sample.
• Average Inter-Onset-Interval (IOI): The average time quanta in-between all
sequential notes.
• Note Length Histogram (NLH): Histogram of note durations. NLH should be
normalized.
• Note Length Transition Matrix (NLTM): Transition matrix between all durations
disregarding pitch information. NLTM should be normalized.
Yang and Lerch instruct that for obtaining absolute metrics, mean and standard
deviation of each feature should be computed. In addition, for obtaining relative
metrics, first, one should perform pairwise exhaustive cross-validation between
features and in result, get a histogram of each feature’s distances. Histograms are
computed by calculating the Euclidian distances between samples at each cross-
validation step (Yang and Lerch use the term “intra-set distances” for histograms
computed within a set. For histograms computed between different sets, they use the
term “inter-set distances”). After obtaining the histograms, the authors suggest
applying kernel density estimation for smoothing the results. Performing kernel
density estimation on histograms reveals probability distribution functions (PDF). And
finally, the authors suggest computing Kullback-Leibler divergence (KLD) and
overlapping area (OA) of inter-set and intra-set PDFs. Yang and Lerch argue that, for
having similar intra-set Gaussian distributions, the variance of 2 datasets should be
similar. They further state that mean values should be similar for having similar inter-
set distributions. Finally, Yang and Lerch point out that for displaying high similarity
between source dataset and generated dataset, KLD value between intra-generated
dataset and inter-sets should be small, and OA value between intra-generated dataset
and inter-sets should be large.
15
CHAPTER THREE
DATASET
3.1 SymbTr Turkish Makam Music Symbolic Data Collection
Having a well-formed, balanced, and large dataset is a key for success in Deep
Learning (DL), but available machine-readable sources for Turkish Makam Music
(TMM) are very rare. The largest and the best formatted machine-readable TMM
digital data source is SymbTr (Karaosmanoğlu, 2012). SymbTr data set contains 2,200
pieces from 155 distinct Makams encapsulating about 865,000 musical notes. In
addition, SymbTr scores are provided in the Text, MusicXML, PDF, MIDI, and Mu2
formats (Şentürk, 2017).
In this thesis, Mu2 format was preferred as the digital representation source. As
shown in Figure 3.1, the contents of Mu2 format in SymbTr collection can be viewed
by a text editor. Entities in Mu2 files are tab-separated which are distributed into rows.
Each row starts with an ID denoting the types of contents of that row. As shown in
Figure 3.1, “50” denotes the piece’s Makam; “51” denotes the piece’s Usûl; “52”
denotes the tempo etc. It can also be seen from Figure 3.1 that musical events in Mu2
files start with “9” followed by pitch name, duration’s numerator and denominator,
and several voicing-related features.
16
Figure 3.1 Inner structure of a Mu2 file from SymbTr dataset.
3.1.1 Statistical Review of SymbTr Dataset
There are more than 150 distinct Makams in SymbTr but not all of them have as
many representatives as others. As shown in Figure 3.2, the two most frequent Makams
in the SymbTr dataset are Hicaz (7.1% of the total set with 157 pieces) and Nihâvent
(5.9% of the total set with 130 pieces). Since DL models benefit from large datasets in
their training processes, target Makams for the thesis scope were chosen to be Hicaz
and Nihâvent.
17
Figure 3.2 Percentages of Makams in SymbTr dataset.
Hicaz and Nihâvent Makams sound very different from each other and naturally are
very different in terms of their scales and pitch characteristics. As shown in Table 3.1,
their tonic, leading, and dominant pitches are different. In addition, they have
completely different scales. As shown in Figure 3.3, Hicaz Makam can be defined as
the combination of Hicaz tetrachord in place (Dügâh) and Rast pentachord on Nevâ.
Whereas Nihâvent Makam is the combination of Bûselik pentachord on Rast and Kürdî
or Hicaz tetrachord on Nevâ (Özkan, 2006).
Table 3.1 Tonal properties of Hicaz and Nihâvent Makams.
Makam Tonic Dominant Leading

Hicaz A (Dügâh) D (Nevâ) G (Rast)
Nihâvent G (Rast) D / Bb (Kürdî) F# (Irak)
In twelve-tone equal temperament Western Music (12-TET WM) theory, the

frequency of any pitch is calculated by the Equation (3.1).
#&
𝐹! = 𝐹" × 2 $% (3.1)
18
where Fn is the frequency of the note to be calculated in Hz; Fb is the base
frequency, for example, 440 Hz for A4; and N is the number of semi-tones between Fn
and Fb. An octave consists of 12 semi-tones in WM. So, when N reaches 12, Fn
becomes 2 x Fb. In other words, the frequency of pitches just doubles when the octave
goes higher, and vice versa. The same case applies to TMM, except that the formula
used to calculate relative frequencies of pitches is different.
Figure 3.3 Scale of Hicaz Makam spanning 2 octaves.
The human brain perceives different octaves of the same pitch perfectly
harmonious. This is due to perfect alignments of pitches’ fundamental tones and their
harmonics. Thus, same pitches in different octaves can be accepted as equivalent.
According to the calculation shown in Figure 3.4, the proportions of pitches of Hicaz
Makam in terms of their durations within the SymbTr collection are shown in Figure
3.5 and Figure 3.6. In Figure 3.4, an example of a duration calculation case is given,
where the total duration of G notes (blue) is 1/4 + 1/4 + 1/8 +1/2 + 1/8 = 14/8 whereas,
B (green) notes have total duration of 1/8 + 1/1 + 1/4 = 11/8. So, the total duration of
G is longer than the total duration of B. In Figure 3.5, different octaves of the same
pitch are merged into the same space, i.e., A4 and A5 are just summed into A. Whereas,
in Figure 3.6, pitches are left as is without any merging or alterations.
19
Figure 3.4 Illustration of note durations.
As shown in Figure 3.5, where octaves are merged, the longest duration belongs to
A (Dügâh) which is the tonic pitch of Hicaz Makam. The second-longest duration
belongs to D (Nevâ) which is the dominant pitch of Hicaz.
Figure 3.5 Hicaz merged octaves.
As shown in Figure 3.6, where octaves are not merged, i.e., separated, the longest
duration belongs to D5 (Nevâ), which is the dominant pitch of Hicaz Makam. And the
second-longest duration belongs to A4 (Dügâh), which is the tonic pitch. Both Figure
3.5 and Figure 3.6 show that the total collection of Hicaz pieces in the SymbTr dataset
represents Hicaz Makam coherent with its formal definitions. Here it should be noted
that pitch names in Figures are given according to WM notation.
20
Figure 3.6 Hicaz separated octaves.
Similarly, in Figure 3.7 and Figure 3.8, merged and separated proportions of pitch
durations of Nihâvent pieces in the SymbTr collection are given. Again, it can be seen
that the longest durations belong to D5 (Nevâ), G4 (Rast), and Bb (Kürdî) pitches,
which are dominant and tonic pitches of Nihâvent Makam. This again shows that
pieces in Nihâvent Makam of SymbTr collection reflect the official definition of
Nihâvent Makam. In conclusion, it can be deduced that SymbTr reflects the tonal
characteristics of TMM Makams well, and therefore it is suitable for DL tasks.
21
Figure 3.7 Nihâvent merged octaves.
Figure 3.8 Nihâvent separated octaves.
3.1.2 SymbTr Dataset N-gram Analysis
N-grams are continuous sequences of N items within a sequential set (Kapadia,

2019). They are useful for information retrieval from the set they are extracted from,
as well as calculating probability of future elements of an existing sequence. Relative
22
frequency of a given N-gramx, which indicates the density of a given sequence of
length N among all possible sequences with equal lengths within a set, is simply
calculated according to Equation (3.2).
𝐶𝑜𝑢𝑛𝑡(𝑁𝑔𝑟𝑎𝑚' )
𝑅𝑒𝑙𝑎𝑡𝑖𝑣𝑒 𝐹𝑟𝑒𝑞𝑢𝑒𝑛𝑐𝑦(𝑁𝑔𝑟𝑎𝑚' ) =
𝐶𝑜𝑢𝑛𝑡(𝐴𝑙𝑙 𝑁𝑔𝑟𝑎𝑚𝑠) (3.2)
In Tables 3.2 thru 3.5, relative frequencies of 3-gram, 4-gram, and 5-gram values
for Hicaz and Nihâvent Makams are given. Calculated relative frequencies of 3, 4, and
5-grams are given in the last columns (Rel. Freq.) of each table. Here, note durations
are omitted, but only pitches are included in calculations.
The term Çeşni in TMM corresponds to short sequences of 4 or 5 pitches that evoke
a feeling of Makam in the listener (Altinköprü, 2018). In other words, they are mini
Makam identifiers. Thus, investigating 4-gram and 5-gram frequencies is especially
important in TMM analysis because they reveal practical Çeşnis of studied Makam. In
all N-gram relative frequency tables, note names are given according to the naming
convention of AEU theory. For example, do5#4 represents Do (C) in the 5th octave,
which is sharp by 4 Komas, which is the symbolic representation of the pitch “Nîm
Hicaz”.
Some of very characteristic 4-pitch runs for Hicaz and Nihâvent Makams such as
re5-do5#4-si4b4-la4 sequence which is the return-to-tonic motive for Hicaz, and a
similar purpose sequence do5-si4b5-la4-sol4 for Nihâvent Makam are listed in Table
3.2 and Table 3.3.
Table 3.2 4-gram frequencies for Hicaz Makam.
1st Pitch 2nd Pitch 3rd Pitch 4th Pitch Rel. Freq.
fa5 mi5 re5 do5#4 1.69
re5 do5#4 si4b4 la4 1.63
sol5 fa5 mi5 re5 1.28
mi5 re5 do5#4 si4b4 1.21
mi5 fa5 mi5 re5 1.09
23
Table 3.2 continues
mi5 re5 do5#4 re5 1.05

do5#4 si4b4 si4b4 la4 1.00
Table 3.3 4-gram frequencies for Nihâvent Makam.
1st Pitch 2nd Pitch 3rd Pitch 4th Pitch Rel. Freq.
re5 do5 si4b5 la4 1.60
mi5b5 re5 do5 si4b5 1.39
do5 si4b5 la4 sol4 1.17
re5 mi5b5 re5 do5 0.99
la4 si4b5 do5 re5 0.92
sol5 fa5 mi5b5 re5 0.82
do5 si4b5 la4 si4b5 0.74
In Table 3.4 and Table 3.5, very characteristic 5-pitch motives can be seen, such as
do5#4-si4b4-si4b4-la4-la4 for Hicaz Makam and re5-do5-si4b5-la4-sol4 for Nihâvent
Makam.
Table 3.4 5-gram frequencies for Hicaz Makam.
1st Pitch 2nd Pitch 3rd Pitch 4th Pitch 5th Pitch Rel. Freq.
do5#4 si4b4 si4b4 la4 la4 0.82
fa5 mi5 re5 do5#4 si4b4 0.79
mi5 re5 do5#4 si4b4 la4 0.73
sol5 fa5 mi5 re5 do5#4 0.72
la5 sol5 fa5 mi5 re5 0.65
fa5 mi5 re5 do5#4 re5 0.64
re5 mi5 fa5 mi5 re5 0.58
Table 3.5 5-gram frequencies for Nihâvent Makam.
1st Pitch 2nd Pitch 3rd Pitch 4th Pitch 5th Pitch Rel. Freq.
mi5b5 re5 do5 si4b5 la4 0.88
re5 do5 si4b5 la4 sol4 0.81
re5 mi5b5 re5 do5 si4b5 0.55
si4b5 la4 sol4 fa4#4 sol4 0.53
re5 do5 si4b5 la4 si4b5 0.51
sol5 fa5 mi5b5 re5 do5 0.45
re5 do5 si4b5 do5 re5 0.44
24
Tables 3.2 thru 3.5 clearly show that pieces of Hicaz and Nihâvent Makams in the
SymbTr dataset represent characteristic motives and Çeşnis of their belonging
Makams very well, consequently, SymbTr is proven to be a convenient dataset for
computational TMM research.
3.2 Dataset Preparation & Representation
There are various strategies to transform music data into a form suitable for DL
models such as timestep sampling, numerical values, binary vectors, textual
representation, etc. (Briot et al., 2017; Kumar & Ravindran, 2019; Wu et al., 2020).
Different types of music data conversions result in different success rates in DL
models’ training phases. Usually, the best strategy is chosen through a series of trials
and errors. It can be stated that the data preparation phase is the most important and
effort-consuming task in DL and data analysis (Barapatre & A, 2017).
A musical note consists of pitch and duration, i.e., vibration of air in a certain
frequency for a certain time. To clear up, “A3 ¼” in a 60 beats per minute (bpm)
context is a 220Hz oscillation that lasts for 1 second. From this point of view, a
monophonic TMM piece can be reduced to oscillations of various successive
durations. By this logic, all distinct pitch-duration tuples in the training dataset were
converted to distinct integer ids by a dictionary such as “Sol4 1/2” → 1, “Sol4 1/4” →
2, …, “Do6 1/32” → 401, etc. With such a dictionary, it is possible to convert pitch-
duration tuples into integers and vice versa.
In SymbTr, there are 405 unique pitch-duration tuples for Hicaz Makam and there
are 374 unique pitch-duration tuples for Nihâvent Makam. As the final step of data
preparation, each pitch-duration tuple’s ids were converted into one-hot encoded
vectors v, such that Equation (3.3) holds for Hicaz Makam.
25
𝑣!
#!#
𝑣"
𝑣=# $ 𝜔ℎ𝑒𝑟𝑒, 𝑣$ ∈ {0,1} ∧ 0 𝑣$ = 1
⋮
𝑣#!# $%! (3.3)
For Nihâvent Makam, the size of the encoding vector is 374, i.e., the last item in
the vector should be v373. By the data representation approach given in this section,
representing any pitch-duration as a DL model input or output reduces the artificial
music composition task into a one-hot encoded single-label classification problem.
26
CHAPTER FOUR
TECHNIQUES AND EVALUATION METHODS
Deep Learning is a machine learning area that came up around 2006, which utilizes
many layers of non-linear processing techniques for feature extraction and
classification tasks over investigated datasets (Deng, 2014). Automatic Turkish
Makam Music Composer (ATMMC) is based on a collection of Deep Learning (DL)
models which utilize Long Short-Term Memory Networks (LSTM) that interact with
each other for accomplishing various tasks of structured artificial TMM composition.
4.1 Long Short-Term Memory Networks
LSTMs emerged for curing vanishing or exploding error signals of the conventional
Back-Propagation Through Time (BPTT) algorithms used in Recurrent Neural
Networks (RNN), and they succeed in learning very long dependencies exceeding
1000 timesteps (Hochreiter & Schmidhuber, 1997). A closer look at LTSMs’ working
principles is given in following sections.
An LSTM cell’s inner structure and its connections to other cells are illustrated in
Figure 4.1. As shown in the figure, LSTM cells consist of pointwise vector addition
and multiplication operations, as well as several gates made up of vectors flowing
through various activation functions. Definitions of LSTM cells’ activation functions,
which are shown with red and green circles in Figure 4.1, are given in Section 4.1.1.
27
Figure 4.1 Internal structure of an LSTM cell (Olah, 2015).
4.1.1 Activation Functions
Neural networks multiply and add vectors in their layers sequentially. For
regulating outputs of consequent multiplications, some forms of activation functions
are used. Activation functions should map a vide possible range of inputs into a limited
domain, and they should be differentiable for applications of Back-Propagation (Rojas,
1996).
In LSTMs, two different activation functions are utilized which are namely Sigmoid
and Tanh. Sigmoid activation function (σ), as given in Equation (4.1) and shown in
Figure 4.2, maps its input into (0, 1) interval (Szandała, 2020).
1
𝜎(𝑥) = (4.1)
1 + 𝑒 ('
28
Tanh function has a similar shape to the Sigmoid function as displayed in Figure
4.2, but it maps the (-∞, ∞) interval into (-1, 1). The formal definition of the Tanh
function is given in Equation (4.2) (Szandała, 2020).
𝑒 ' − 𝑒 ('
𝑡𝑎𝑛ℎ(𝑥) = (4.2)
𝑒 ' + 𝑒 ('
Both Sigmoid and Tanh activation functions are highly utilized in a wide range of
neural network types. They both introduce non-linearity to the system and allow for
classification in complex spaces. Even though the Sigmoid activation function is
computationally less expensive, Tanh can represent mappings in negative domain,
hence can save the neural network from getting stuck during the training phase.
Figure 4.2 Sigmoid (left) and Tanh (right) function plots.
4.1.2 Long Short-Term Memory Cell Working Principles
The most important component of an LSTM cell is the cell carrying the information
in the network chain like a conveyor belt (Olah, 2015). As shown with the blue line in
Figure 4.3, data flows from an LSTM cell to another through cell state.
29
Figure 4.3 Depiction of data flow through LSTM cell state.
At any time step (t), xt represents the current input, ht-1 represents the previous
hidden state, and Ct-1 represents the previous cell state. As shown in Figure 4.4, the
previous hidden state and current input are concatenated into a new vector, and this
new vector is mapped to the activation value of the forget gate, ft, as shown in Equation
(4.3) (van Houdt et al., 2020).
𝑓) = 𝜎(𝑊* . [ℎ)($ , 𝑥) ] + 𝑏* ) (4.3)
where Wf is the weight and bf is the bias associated with the forget gate. At this step,
if ft is calculated to be close to 1, that means the previous cell state will be strongly
remembered. Whereas if ft is close to 0, the previous cell state is going to be forgotten.
Figure 4.4 Data flow through the LSTM cell’s forget gate.
30
In the next phase, as shown in Figure 4.5, cell state Ct gets updated by the input
gate and gets its final value. Ct is calculated according to Equation (4.4), Equation
(4.5), and Equation (4.6) (Olah, 2015).
Figure 4.5 LSTM cell state calculation.
𝑖) = 𝜎(𝑊+ . [ℎ)($ , 𝑥) ] + 𝑏+ ) (4.4)
CK ) = 𝑡𝑎𝑛ℎ(𝑊, . [ℎ)($ , 𝑥) ] + 𝑏, ) (4.5)
𝐶) = 𝑓) ∗ 𝐶)($ + 𝑖) ∗ CK ) (4.6)
Finally, ht, which is the current hidden state, is calculated according to Equation
(4.7) and Equation (4.8) by using the output gate (van Houdt et al., 2020). Output gate
decides what the hidden state for the next LSTM cell will be. The hidden state contains
information about previous inputs (x) and is used for predictions. An illustration of
hidden state calculation is shown in Figure 4.6.
o) = 𝑡𝑎𝑛ℎ(𝑊- . [ℎ)($ , 𝑥) ] + 𝑏- ) (4.7)
ℎ) = 𝑜) ∗ 𝑡𝑎𝑛ℎ(𝐶) ) (4.8)
31
Figure 4.6 LSTM hidden state formation.
4.2 Probability Distribution Analysis of Various Features
Detailed objective evaluations regarding Usûl and form features were performed
according to the methods proposed by Yang and Lerch (2020). An overview of Yang
and Lerch’s methodology is given in Section 2.4. In this section, the details of Yang
and Lerch’s method’s workflow are given.
The general workflow of Yang and Lerch’s method is illustrated in Figure 4.7. As
described in Section 2.4, Yang and Lerch’s method computes both absolute and
relative metrics. Absolute metrics are computed within a given set to give insight about
its characteristics whereas relative metrics show how related to sets are (Yang & Lerch,
2020).
32
Figure 4.7 Workflow of Yang and Lerch’s method (Yang & Lerch, 2020).
Absolute metrics are computed by calculating the mean and standard deviation
(STD) of proposed features. However, relative metrics require a number of additional
steps. Details of some techniques involved in relative metrics computation are given
in the following sub-sections.
33
4.2.1 Pairwise Cross Validation
Pairwise cross validation is finding the distance of features both within and between
sets. If there are two sets such as set_1 = [1, 2, 3] and set_2 = [5, 0, 7] the distance
between sets are {dist([1], [5, 0, 7]), dist([2], [5, 0, 7]), dist([3], [5, 0, 7])}, where dist
denotes Euclidian distance. Since the distance is computed between two sets in this
example, it is called inter-set distance (Yang & Lerch, 2020).
When the distance is computer within the same set it is then called intra-set distance.
As an example intra-set distance for set_2 would be {dist([5], [0, 7]), dist([0], [5, 7]),
dist([7], [5, 0])}.
4.2.2 Kernel Density Estimation
Kernel density estimation is applied to inter-set and intra-set distance histograms

for smoothing out the results into probability density functions (PDF). PDF is one of
the best probability distribution functions to show how the whole mass is distributed
over the x-axis (Węglarczyk, 2018).
An illustration of kernel density application is given in Figure 4.8. It can be seen

from the figure that the density estimation with a Gaussian kernel smooths out
segmented histograms.
34
Figure 4.8 Kernel density estimation application (Vanderplas, 2007).
4.2.3 Kullback-Leibler Divergence and Overlapped Area
The final step of relative metrics calculation is finding the Kullback-Leibler

Divergence (KLD) and Overlapped Area (OA) values. KLD quantifies the difference
between probability distributions (Brownlee, 2019). Definition of KLD between two
continuous probability densities P(x) and Q(x) is given in Equation (4.9) (Zhang et al.,
2021).
.
𝑃(𝑥)
𝐾𝐿(𝑃(𝑥) || 𝑄(𝑥)) = S 𝑃(𝑥)𝑙𝑜𝑔 𝑑𝑥
(. 𝑄(𝑥) (4.9)
It is easy to see from Equation (4.9) that KL(P(x) || Q(x)) ≠ KL(Q(x) || P(x)), i.e.,
KLD is unbounded. Thus, Yang and Lerch (Yang & Lerch, 2020) suggest further
calculating OA, to provide a bounded measure.
35
In contrast to KLD, OA is a measure of similarity between two PDFs. It is the area
of intersection between multiple PDF (Pastore, 2018). Calculating both KLD and OA,
hence gives the difference and similarity of the two datasets.
4.2.4 Method Demonstration
To further clarify Yang and Lerch’s method (Yang & Lerch, 2020), two toy
examples are given in this section. The first set of toy data are:
• base_set = [[18], [15], [18], [17], [19]]

• gen_set_1 = [[19], [18], [14], [18], [17]]
• gen_set_2 = [[1], [2], [1], [0], [2]]
The above sets represent 3 datasets. base_set represents the dataset of 5 pieces
(samples) from which gens_set_1 and gen_set_2 are artificially created. The numbers
represent pitch count per sample. Computed absolute metrics are given in Table 4.1.
From both mean and STD values in Table 4.1, it is obvious that gen_set_1 is much
similar to base_set than gen_set_2.
Table 4.1 Absolute measurement results for toy data 1.
Mean STD
base_set 17.4 1.35
gen_set_1 17.2 1.72
gen_set_2 1.2 0.74
Relative measurement results are shown in Table 4.2 and Figure 4.9. Results in
Table 4.2 show that KLD metric between intra-baset_set and inter base_set –
gen_set_1 is 0.1 / 0.04 = 2.5 times smaller than KLD metric between intra-baset_set
and inter base_set – gen_set_2, which can be interpreted as gen_set_1 is 2.5 times less
different to base_set compared to gen_set_2. But the OA difference is much higher.
OA of the base_set and gen_set_1 is 0.76 / 2.35e-10 ≈ 3.23e+9 times larger than
base_set and gen_set_2’s OA.
36
Table 4.2 Relative measurement results for toy data 1.
KLD OA
base_set & gen_set_1 0.04 0.76
base_set & gen_set_2 0.10 2.35e-10
A visual representation of base_set, gen_set_1, and gen_set_2 difference PDFs is

shown in Figure 4.9. Even though intra-base_set (blue), intra-gen_set_1 (green), and
intra-gen_set_2 (purple) distributions have different density peaks, they align on the
x-axis closely. But when PDFs of inter-set distances are investigated, base_set &
gen_set_2 distance PDF lies on the right-hand side of the graph on the x-axis far away
from the intra-set PDFs. This graph proves that PDF plots of inter-set distances can
show similar and different datasets clearly.
Figure 4.9 Probability distributions for toy data set 1.
The feature metrics in the first toy example had obvious differences. As shown in
Table 4.2, both Mean and STD values for gen_set_1 and gen_set_2 were greatly
different. For further investigation, another toy example is given below:
37
• base_set = [[18], [19], [18], [17], [19]]
• gen_set_1 = [[19], [18], [19], [18], [17]]
• gen_set_2 = [[1], [2], [1], [0], [2]]
Computed absolute metrics related to the second toy data are given in Table 4.3. In
this example, the mean values are different, but the STDs are the same. Thus, merely
looking for absolute metrics may not provide a distinction as clear as the first toy set.
Table 4.3 Absolute measurement results for toy data 2.
Mean STD
base_set 18.2 0.74
gen_set_1 18.2 0.74
gen_set_2 1.2 0.74
Relative measurement metrics and plots of PDFs for toy data 2 are given in Table
4.4 and Figure 4.10. Table 4.4 shows that just looking at KLD values, gen_set_2 seems
to be more similar to base_set than gen_set_1. But when OA values are examined,
gen_set_1 is found to have 0.67 / 1.98e-47 ≈ 3.38e+46 times larger OA with base_set,
compared to gen_set_2.
Table 4.4 Relative measurement results for toy data 2.
KLD OA
base_set & gen_set_1 0.02 0.67
base_set & gen_set_2 0.01 1.98e-47
In Figure 4.10, intra-base_set (blue), intra-gen_set_1 (green), and intra_gen_set_2

(purple) PDFs are shown to be aligned nearly perfectly that they are overlapping and
covering each other up. However, again examining the inter-set differences show that
gen_set_2 is much more similar to base_set than gen_set_2. From both toy set 1 and 2
figures, it can be deduced that placement of inter-set distance PDFs on the x-axis is
effective in showing how similar or different the datasets are.
38
Figure 4.10 Probability distributions for toy data set 2.
39
CHAPTER FIVE
AUTOMATIC COMPOSER
Automatic Turkish Makam Music Composer (ATMMC) is a collection of Deep

Learning (DL) and N-gram probability models. In this chapter, its building blocks and
their inner structures are given in detail.
5.1 Base Models
The Base Model (BM) is the most important building block of ATMMC. It is a
Deep Neural Network (DNN) from which all other DNNs of ATMMC are derived. As
illustrated in Figure 5.1, BM consists of a 600-unit LSTM layer followed by a 50%
dropout layer followed by again 600 units LSTM and 50% dropout layers, and finally,
a fully connected dense layer and a SoftMax activation layer (Parlak et al., 2021). BM
is compiled with categorical cross-entropy loss function and the RMSprop (root mean
square prop) optimizer. The learning rate of RMSprop optimizer is 0.001.
40
Figure 5.1 Base Model’s layer architecture.
As described in Section 3.2, musical notes are represented as pitch-duration tuples

encoded into one-hot vectors. BMs’ training input and output data are created by
matching 8 pitch-duration pairs with 1 following pitch-duration pair. As an example,
if the processed piece is [pd0, pd1, pd2, …, pdk+8], then the training input set becomes
[[pd0, pd1, …, pd7], [pd1, pd2, …, pd8], …, [pdk, pdk+1, …, pdk+7]] and training output
set becomes [[pd8], [pd9], …, [pdk+8]], where each pdi is a one-hot encoded vector
representing a pitch-duration tuple.
Trained BMs learn the conditional probability P(pdt+1 | pdt-7, ..., pdt) of the next
pitch-duration tuple pdt+1 at any given time t with respect to the previous 8 pitch-
duration tuples. Thus, in fact, they learn to compose the next note of a musical piece
according to the most recent 8 notes.
41
ATMMC is designed to compose musical pieces with certain Usûl and forms. But
BMs are not aware of any compositional structure. Since they are trained on the whole
dataset of their target Makams and their datasets consist of mixed Usûl and forms, they
learn the general feel of their target Makams. On their own, they output pleasant
melodies in their target Makams’ scales, but their compositions lack structure.
For each Hicaz and Nihavent Makams, multiple BMs with various hyperparameters
were trained and 2 subjectively most musical sounding models were chosen to form
the basis for each Makam’s Specialist Models.
5.2 Specialist Models
Specialist Models (SM) are the models that specialize in composing sections of
Şarkı form in certain Usûls for Hicaz and Nihâvent Makams. An SM can either
compose just one of the Zemin, Nakarat, or Meyan sections.
In SymbTr collection, there are 17 pieces within Hicaz Makam having Aksak Usûl
and Şarkı form, and the number of Nihâvent pieces in Düyek Usûl and Şarkı form is
21. The number of available pieces in the SymbTr collection is not sufficient for
training DL models that are capable of producing structured TMM pieces. To solve
this problem, 88 additional pieces were collected from various archives and manually
converted to Mu2 formats. Also, as well as the available pieces in SymbTr dataset,
their Zemin, Nakarat, and Meyan sections are manually labeled.
SMs are deep LSTM networks that are derived from BMs inheriting their weights.
As shown in Figure 5.2, BMs and SMs have the same architecture. Also, as shown in
Figure 5.2b, SMs first LSTM layer is frozen, and their other layers are left trainable.
The technique of inheriting other models’ weights is called Transfer Learning and it is
especially useful when the amount of training data is limited (Weiss et al., 2016).
With transfer learning, a DL model inherits its initial weights from an already
trained model on a similar domain. During the training phase, having knowledge in a
42
similar domain, the derived model learns new relations about its training data in
addition to its past expertise gathered from its base model. This way, knowledge is
transferred between models operating on similar domains. The difference between
frozen layers and trainable layers shown in Figure 5.2b is weights in trainable layers
are updated during the training phase, but weights in the frozen layers are kept
unchanging.
Figure 5.2 Base model diagram (a) and specialist model diagram (b).
For each Zemin, Nakarat, and Meyan sections of target Makam’s Şarkı forms, two
SMs are trained, adding up to a total of 6 SMs per Makam. As shown in Figure 5.3,
Zemin SM is trained for composing Zemin section from 8 initial notes (pitch-duration
tuples), Nakarat SM is trained for composing the Nakarat section by using the last 8
notes of the Zemin section, and finally, Meyan SM is trained for composing the Meyan
section by using the last 8 notes of Nakarat section. By passing the last 8 notes of each
section to the SMs of next section, a connection and harmony between Zemin, Nakarat,
and Meyan sections are established.
43
Figure 5.3 Operation scheme of Specialist Models.
The final layer of SMs, the SoftMax layer, outputs a probability distribution as
given in Equation (5.1), where zi represents each element of SM’s output vector, and
k is the number of unique notes.
𝑒 /&
𝜎(𝑧⃗)+ =
∑012$ 𝑒 /' (5.1)
From SM’s output, the prediction for the next note is extracted. Two possible
outputs of SMs are shown in Figure 5.4. If SM is confident with its prediction, it will
favor a single note, which is a strong candidate, above all other possible notes as shown
in Figure 5.4a. But if SM is not confident with its prediction, as shown in Figure 5.4b,
it will output several low probability peaks, i.e., multiple weak candidates.
44
Figure 5.4 Specialist model prediction samples.
Only 1 note should be chosen from SMs candidates. For choosing the candidate
from the probability distribution, high and low threshold values are determined. In
Figure 4.10, the high threshold is 0.70, and it is shown with a green line. Whereas the
low threshold value, which is shown with the red line, is 0.18. In the case of an existing
peak above the high threshold, it is chosen to be the next note. But, if there is no peak
above the high threshold, peaks above the low threshold are inspected, and the one
with the strongest 4-gram probability is chosen to be the next note as given in Equation
(5.2). Where pdt represents the candidate whose 4-gram score is being calculated, and
the sequence of {pdt−3, pdt−2, pdt−1} represents the last 3 pitch-duration tuples fed into
the SM.
𝐶𝑜𝑢𝑛𝑡(𝑝𝑑)(3 , 𝑝𝑑)(% , 𝑝𝑑)($ , 𝑝𝑑) )

𝑃(𝑝𝑑) | 𝑝𝑑)(3 , 𝑝𝑑)(% , 𝑝𝑑)($ ) =
𝐶𝑜𝑢𝑛𝑡(𝑝𝑑)(3 , 𝑝𝑑)(% , 𝑝𝑑)($ ) (5.2)
After 4-gram probability calculation, if a tie occurs between multiple candidates,

the next note is chosen randomly from the candidate list.
The artificial compositions are greatly affected by tuning the high and low threshold
values. Playing with the threshold values may cause more candidates to pass barriers
45
causing more adventurous or traditional synthetic compositions depending on their
values. Moreover, since Zemin, Nakarat, and Meyan sections are composed by
separate SMs, conventionality of each section can be separately controlled by related
SM’s high and low threshold values.
5.3 Conductor Models
As described in Section 5.2, two different SMs are trained per Zemin, Nakarat, and
Meyan sections for each Makam. This results in two candidates for the next note to be
composed, predicted by each SM. In order to choose one candidate over the other
Conductor Models (CoMo) are trained. Again, as illustrated in Figure 5.5, CoMos are
deep LSTM networks, which have 3 layers of 100 LSTM units and a fully connected
final layer. Between their layers, CoMos have 50% dropout factors. CoMos are
compiled with RMSprop (root mean square prop) optimizers and categorical cross-
entropy loss functions.
Similar to the training process of SMs, CoMos are first trained over the whole
dataset for their target Makam. This way they learn the general characteristics of their
Makam without explicit Usûl and form information. Then the trained CoMos are again
trained on Zemin, Nakarat, and Meyan sections only in their target Usûls.
46
Figure 5.5 Conductor Model structural decomposition.
As shown in Figure 5.6, CoMos receive the 8 previous pitch duration tuples shown
as pdi, pdi+1, …, pdi+7 as well as candidates from two SMs shown as pdi+8(A) and
pdi+8(B). Then, they output a vector 𝑣 = [𝑣4 , 𝑣" ] ∶ 𝑣4 + 𝑣" = 1 denoting the
probability of SM(A)’s and SM(B)’s candidates. If |𝑣4 − 𝑣" | ≥ 0.2, the candidate
with greater probability is determined to be the next note. But, if the probability
difference between the candidates is smaller than 0.2, this means CoMo is not very
confident with its output, thus, the next note is determined by a random selection
between pdi+8(A) and pdi+8(B).
47
Figure 5.6 Structural decomposition of ATMMC.
5.4 Full System Overview
ATMMC is a collection of SMs and CoMos. There are 6 SMs and 3 CoMos for
each of Hicaz and Nihavent Makams, adding up to 12 SMs and 6 CoMos for the whole
system.
As illustrated in Figure 5.7, ATMMC system works as follows: after choosing the
target Makam and Usûl, a user enters 8 initial pitch-duration tuples (pd0, pd1, …, pd7)
to the system. Then ATMMC composes the 9th note (next_note) according to user
input and appends it to the end of the notes list (Song). Next, ATMMC picks the last
8 notes from the notes list and composes the 10th note (next_note) according to the
selected 8-note set. This process of picking the last 8 notes and composing the next
note repeats until a 4-bar long Zemin section is completed.
When the Zemin section is completed, ATMMC picks the last 8 notes of the Zemin
section and begins composing the Nakarat section. Here, it should be noted that
ATMMC changes its SMs and CoMos from ATMMC_Zemin to ATMMC_Nakarat
which are specialized in the Nakarat sections. After composing the 4-bar long Nakarat
section, ATMMC picks the last 8 notes from the Nakarat section and composes the
48
Meyan section similarly. Again, while composing the Meyan section, only Meyan
section associated SMs and CoMos (ATMMC_Meyan) are utilized.
Once composition processes of Zemin, Nakarat, and Meyan sections are completed,
ATMMC merges the composed sections into a single Şarkı (Mus2_File) and writes
that Şarkı to the disk in Mu2 format.
49
Figure 5.7 ATMMC composition process flowchart.
50
CHAPTER SIX
GRAPHICAL USER INTERFACE FOR ATMMC
6.1 Purpose of the Graphical User Interface
Automatic Turkish Makam Music Composer (ATMMC) is developed in Python

programming language as an open-source project and the source of the application is
publicly accessible at https://github.com/ihpar/TMMDLFT. Users can install Python
and the necessary libraries on their computers, run the application with the terminal,
and create Mus2 files via ATMMC according to the input parameters. However, users
who want to use ATMMC via terminal should be familiar with computer
programming, albeit at a basic level.
A graphical user interface that can be accessed with an internet browser has been
published at http://music.cs.deu.edu.tr/tmmgui to eliminate this programming
knowledge requirement and make ATMMC available to users who do not have any
computer programming experience (Parlak et al., 2021). With this interface, users can
compose artificial creations in ATMMC and save the result file on their computers.
The recorded files can then be viewed and played back with the Mus2 application.
6.2 Graphical User Interface Decomposition
The graphical user interface developed for ATMMC is shown in Figure 6.1. Due to
the large size of the interface, the picture is split into two, and the details of the left-
hand side of the interface, denoted as “L” in Figure 6.1, is given in Figure 6.2.
Likewise, the details of the right-hand side denoted as “R” in Figure 6.1, is given in
Figure 6.3 for better visibility.
51
Figure 6.1 Overview of ATMMC’s graphical user interface.
As shown in Figure 6.2, the voicing feature of the interface can be turned on or off
with the button indicated with “A”. While the voicing feature of the graphical interface
is active, users can hear the notes they add. In addition, when the space key is pressed,
users can listen to all notes they have inserted, one after the other. The frequencies of
the pitches used in the vocalization were determined according to the Arel-Ezgi-
Uzdilek (AEU) system using the TuneJS library (Bernstein & Taylor, 2003).
Makam and Usûl selection functionalities are shown in Figure 6.2 “B” and “C”
respectively. From Figure 6.2 “B”, the user can switch between Hicaz and Nihâvent
Makams. Likewise, from Figure 6.2 “C”, the user can switch between Düyek and
Aksak Usûls. When Hicaz Makam is selected, Usûl automatically switches to Aksak.
Similarly, when Nihâvent Makam is selected Usûl automatically switches to Düyek.
Also changing the target Usûl and Makam updates the time signature and key signature
automatically.
From the section shown in Figure 6.2 “D”, the user can pick pitch alteration symbols
if needed. The pitch alteration symbols given in this section are again from AEU
theory.
The section shown in Figure 6.2 “E” is where the user enters and modifies the 8
initial notes required for executing ATMMC.
52
Figure 6.2 ATMMC graphical user interface left-hand side closeup.
In Figure 6.3 “A”, duration symbols are shown. The user can click and select any
duration from this section and add the next note accordingly. Silence symbols are
shown in Figure 6.3 “B”. Just like inserting notes, the user can enter silence symbols
for various durations. And finally, the button for triggering ATMMC’s composition
process is shown in Figure 6.3 “C” (Bestele; Compose).
Figure 6.3 ATMMC graphical user interface right-hand side closeup.
As a general guideline for the graphical user interface, the user can select any of the
note durations, pitch alteration symbols, or silence symbols from the menu by mouse
clicks and then click to the desired area onto the section marked in Figure 6.2 “E” for
inserting the selected element. If the user wants to delete the inserted item, highlighting
53
the unwanted item by clicking on it then pressing the delete key on the computer’s
keyboard will do the task.
As shown in Figure 6.4, a new measure is created automatically when the old
measure is fully occupied. In Figure 6.4a, total duration within the measure is 1/8 +
1/4 + 1+8 + 1/4 = 6/8. When a 1/4 note is inserted as shown in Figure 6.4b, the total
duration reaches 6/8 + 1/4 = 8/8 and an empty measure is automatically created
allowing for new note insertions.
Figure 6.4 Before (a) and after (b) of automatic measure creation.
If the total duration in the current measure is incorrect, for example, higher than 8/8
for Düyek Usûl, as shown in Figure 6.5, the notes in the measure turn red, indicating
the occurrence of an error in the measure. When the user creates a faulty measure, the
button that triggers the composition process gets disabled and the system does not
allow the automatic composition process to start.
54
Figure 6.5 An erroneous measure.
After the user enters 8 initial notes correctly, the button indicated with Figure 6.3
“C” becomes active and it becomes possible for the user to send the notes he has
written to ATMMC.
When the user initiated the composition process, the status window shown in Figure
6.6 pops up and the progress of the composition processes running in the ATMMC
server is displayed to the user in real-time.
55
Figure 6.6 Composition progress report pop-up.
The status steps in pop-up window shown in Figure 6.6 are in Turkish and they can
be translated as:
• Başlangıç: Start
• Besteci başlatıldı: Composer started
• Zemin besteleniyor: Zemin is being composed
• Nakarat bestelenecek: Nakarat is going to be composed
• Meyan bestelenecek: Meyan is going to be composed
• Bitiş: Finish
• İndirme hazırlanacak: Download is going to be prepared
Depending on the ATMMC server load, the composition process can take between
30 seconds and 3 minutes. When the composition process is finished, the composition
progress pop-up transitions to its final state as shown in Figure 6.7. Finally, the user
can save the ATMMC’s artificial composition as a Mus2 compatible file, by clicking
the "download ready" link.
56
Figure 6.7 Artificial composition download link.
The status steps in pop-up window shown in Figure 6.7 are in Turkish and they can
be translated as:
• Başlangıç: Start
• Besteci başlatıldı: Composer started
• Zemin bestelendi: Zemin is composed
• Nakarat bestelendi: Nakarat is composed
• Meyan bestelendi: Meyan is composed
• Bitiş: Finish
• İndirme hazır: Download is ready
57
CHAPTER SEVEN
RESULTS & EVALUATION
Both subjective and objective evaluation methods were performed for evaluating
the ATMMC's effectiveness. For evaluation, 20 Hicaz and 20 Nihâvent pieces were
composed by ATMMC, where each piece consisted of 4 bars of Zemin, Nakarat, and
Meyan sections. Details and results of the evaluations are given in the following
sections.
7.1 Subjective Evaluation
Subjective evaluations were performed by asking a series of questions in the form

of a questionnaire to artists and researchers who are accepted as domain experts in
Turkish Makam Music. A total of 5 artists participated in the survey and evaluated
ATMMC’s artificial compositions according to the questions listed below:
• Question 1: To what extent do the ATMMC compositions you have listened to

represent the Hicaz Makam?
• Question 2: Do the pieces you have listened to have the characteristics of Aksak
Usûl?
• Question 3: Do the pieces you have listened to have the characteristics of Şarkı
form?
• Question 4: How original are the pieces you have listened to?
• Question 5: If you hadn't been told that they were composed by a computer,
would you think that the songs you've heard were composed by a human being?
• Question 6: Did the songs you have listened to hold artistic beauty?
• Question 7: Would your thoughts change if you listened to the songs you just
heard from a real-life professional performer and not from a MIDI instrument?
• Question 8: For what purpose can this artificial composer be used?
The responses of the participants to the first and second questions are given in
Figure 7.1. Participants replied that ATMMC’s artificial compositions reflect their
58
target Makams perfectly and reflect their target Usûls above average. It should be
noted that ATMMC’s ability to reflect its targeted Usûls is lower than its ability to
reflect its targeted Makams. This is a consequence related to the quantities in training
sets. In training sets, the number of pieces in a certain Makam highly surpasses the
number of pieces in discrete Usûls. Having much higher amounts of representatives,
Makams are better leaned compared to Usûls.
Question 1 Responses Question 2 Responses
40%
60%
100%
None Very little Moderately Perfectly No Moderately Perfectly
Figure 7.1 Subjective evaluation responses for questions 1 and 2.
Responses to question 3, shown in Figure 7.2, exhibit that the participants think
ATMMC’s artificial compositions are moderately coherent with the characteristics of
Şarkı form. Again, this result can be interpreted similarly to the results of the second
question, which is related to the fewer representatives of Şarkı form in the training set.
Responses to question 4, shown in Figure 7.2, reveal that the participants found the
ATMMC’s compositions moderately original. From the answers given to question 4,
it can be concluded that the ATMMC has learned the dataset instead of overfitting it.
59
20%
80%
100%
No Moderately Perfectly None Partially Perfectly
It can be seen from Figure 7.3 that the participants would think ATMMC was a
human being if they were not told that it was a machine, and they moderately find its
compositions artistically pleasing. This result indicates that the system can moderately
mimic the composers who composed the dataset it was trained on.
20% 20%
80% 80%
No Maybe Yes No Partially Yes
The participants would rate ATMMC higher if the synthetic compositions were
played back to them from a professional real instrument recording rather than MIDI
sounds, as shown in Figure 7.4. This result is particularly important for future studies
to record the synthetic compositions with real instruments in a professional
environment.
60
Question 7 Responses
20%
80%
No Yes, positively
Figure 7.4 Subjective evaluation responses for question 7.
Finally, responses to question 8 are given in Table 7.1. A small percentage of the
participants found ATMMC to be useless. Whereas most participants thought that
ATMMC could help human composers in composing new musical pieces by
introducing new ideas. At the same time, some participants thought that the system
can be used to quickly generate etudes for music students in conservatories.
Table 7.1 Subjective evaluation responses for question 8.
Response Percentage
It is useless. 20%
May assist human composers in composing new pieces. 60%
It can be used to create etudes for students. 20%
In general, when all the subjective evaluations are considered, it can be said that
ATMMC can compose above-average pieces in terms of pitch, rhythm, and form
representation abilities. Also, it can be deduced that ATMMC cannot be used to
replace human composers but to assist them.
7.2 Objective Evaluation
Objective evaluations on ATMMC's compositions were carried out in two

branches: pitch density analysis on the overall data set and detailed analysis on song
61
forms with Usûl information. The details of performed analysis and findings are given
in following sections.
7.2.1 Pitch Density Analysis
Pitch density analysis was performed by finding the total duration of all pitches in
both the SymbTr and ATMMC-compositions datasets. While calculating the total
duration, the tempos of the processed pieces were accepted to be the same. In other
words, note lengths are not measured with seconds but with note length fractions like
1/2, 1/4, 1/32, etc. Pitches with lower total percentages than %5 were omitted for
avoiding the cluttering of results’ readability. Finally, the density values were
calculated by normalizing the whole sets into [0, 1] intervals.
The pitch density graph of Hicaz Makam is shown in Figure 7.5. It can be seen from
the graph that the most emphasized peaks in the Hicaz set are near re5 and la4 pitches,
which are the tonic and the dominant of the Hicaz Makam.
The similarity of the pitch densities of SymbTr and ATMMC-compositions sets is

calculated as shown in Equation (7.1).
#
2 min (𝑆+ , 𝐴+ )
𝑆𝑖𝑚𝑖𝑙𝑎𝑟𝑖𝑡𝑦(𝑆, 𝐴) = × ]
𝑁 𝑆+ + 𝐴+ (7.1)
+2$
where S and A are pitch histograms of SymbTr, and ATMMC-compositions sets

representing pitch densities, and Si and Ai represent each bin in corresponding
histograms. By applying Equation (7.1) to pitch density distribution shown in Figure
7.5, ATMMC-compositions for Hicaz Makam is found to be 67% similar to SymbTr
in terms of pitch densities.
62
Figure 7.5 Hicaz Makam pitch density graph.
A similar result is shown in Figure 7.6 for Nihâvent Makam. It can be seen that the
peaks are located around re5, si4b5, and sol4, which are the tonic and dominant pitches
of Nihâvent Makam. Again, it can be seen from the graph that both SymbTr and
ATMMC pitch densities are distributed coherently. Applying Equation (7.1),
ATMMC-compositions for Nihâvent Makam is found to be 73% similar to SymbTr in
terms of pitch densities. From Figure 7.6, it can be deduced that Both SymbTr and
ATMMC sets represent Nihâvent Makam in terms of pitch densities very well and
ATMMC is efficient in Nihâvent compositions in terms of adequate pitch density
distribution.
63
Figure 7.6 Nihâvent Makam pitch density graph.
7.2.2 ATMMC Detailed Analysis Results
For assessing ATMMC’s capabilities of reflecting its training set several metrics
and their absolute and relative measures are calculated as suggested by Yang and Lerch
(Yang & Lerch, 2020). For assessment, 20 pieces for Hicaz Makam and 20 pieces for
Nihâvent Makam were artificially composed by ATMMC. Then the composed pieces
were compared with the pieces in SymbTr. Comparisons were performed only within
Şarkı form and Aksak Usûl for Hicaz Makam and Düyek Usûl for Nihâvent Makam.
Two example ATMMC composition samples are shown in Figure 7.7 and Figure 7.8.
64
Figure 7.7 An example Hicaz composition by ATMMC.
Figure 7.8 An example Nihâvent composition by ATMMC.
Both absolute and relative metrics regarding pitch count (PC), pitch count per bar
(PC/Bar), note count (NC), note count per bar (NC/Bar), pitch-class histogram (PCH),
pitch-class histogram per bar (PCH/Bar), note length histogram (NLH), pitch-class
transition matrix (PCTM), pitch range (PR), average pitch interval (PI), average inter-
onset-interval (IOI), and note length transition matrix (NLTM) features are given in
Table 7.2, Table 7.3, and Table 7.4.
65
The metrics given in Table 7.2 and Table 7.3 show that ATMMC has similar
average means both for Nihâvent and Hicaz Makams with SymbTr dataset. When
average means are compared, 100 × (1 − (5.26 − 4.47) / 5.26) ≈ 84.98% average
similarity for Nihâvent Makam and 100 × (1 − (8.64 − 7.24) / 8.64) ≈ 83.79% average
similarity for Hicaz Makam to SymbTr dataset are evaluated. It can also be seen from
Table 7.2 and Table 7.3 that Hicaz and Nihâvent sets generated by ATMMC have less
diversity in terms of pitch and duration variations. Reduction in diversity for measured
features is an expected outcome since ANN models generally exclude outliers and
learn the general structure of relations over the set they are trained.
Table 7.2. Absolute metrics for Nihâvent Makam.
SymbTr ATMMC
Mean STD Mean STD
PC 14.05 1.62 11.60 1.20
PC/Bar 4.18 1.25 4.76 1.26
NC 6.60 1.65 4.40 1.01
NC/Bar 2.69 0.75 2.09 0.70
PCH 0.05 0.07 0.05 0.07
PCH/Bar 0.05 0.13 0.05 0.11
NLH 0.07 0.13 0.07 0.17
PCTM 0.00 0.01 0.00 0.01
PR 31.85 2.97 27.35 2.57
PI 3.59 0.47 3.34 0.30
IOI 0.08 0.03 0.04 0.01
NLTM 0.00 0.02 0.00 0.04
Average 5.26 0.75 4.47 0.62
Table 7.3. Absolute metrics for Hicaz Makam.
SymbTr ATMMC
Mean STD Mean STD
PC 14.60 2.95 10.85 1.38
PC/Bar 5.35 1.45 5.29 1.48
NC 6.45 1.82 5.05 1.11
66
Table 7.3 continues
NC/Bar 3.12 1.03 2.35 0.85

PCH 0.04 0.06 0.04 0.07
PCH/Bar 0.04 0.09 0.04 0.09
NLH 0.06 0.14 0.06 0.15
PCTM 0.00 0.01 0.00 0.01
PR 67.00 8.98 56.50 6.61
PI 7.00 0.52 6.77 0.44
IOI 0.05 0.01 0.03 0.01
NLTM 0.00 0.03 0.00 0.03
Average 8.64 1.42 7.24 1.01
KLD and OA metrics between intra-SymbTr set and inter-ATMMC set for
Nihâvent and Hicaz Makams are given in Table 7.4. Both Nihavent and Hicaz sets of
ATMMC have the largest OA values for pitch count per bar (PC/Bar) feature. This
result can be interpreted as the best feature of ATMMC is its ability in producing
diverse pitches per bar.
It can be seen from Table 7.4 that, on average, ATMMC’s Hicaz composer slightly
outperforms the Nihavent composer both by larger OA and smaller KLD values. This
outcome might be related to the larger size of the Hicaz set in SymbTr compared to
the Nihâvent collection of SymbTr.
Table 7.4 Relative metrics for Hicaz and Nihâvent Makams.
Hicaz Nihâvent
KLD OA KLD OA
PC 0.243 0.533 0.043 0.604
PC/Bar 0.037 0.898 0.029 0.886
NC 0.045 0.701 0.115 0.566
NC/Bar 0.065 0.640 0.010 0.618
PCH 0.056 0.638 0.250 0.811
PCH/Bar 0.051 0.511 0.165 0.595
NLH 0.012 0.814 0.165 0.603
67
Table 7.4 continues
PCTM 0.194 0.573 0.114 0.630

PR 0.154 0.669 0.249 0.551
PI 0.023 0.855 0.102 0.820
IOI 0.166 0.358 0.340 0.548
NLTM 0.025 0.732 0.153 0.541
Average 0.089 0.660 0.144 0.647
For further demonstration of ATMMC’s effectiveness, PDFs (Probability

Distribution Functions) of various features of Hicaz and Nihavent Makam for both
inter and intra sets are given in Figures 7.9 thru 7.14. Intra-SymbTr sets are represented
with blue lines, whereas intra-ATMMC sets are represented with green lines.
Distances between SymbTr and ATMMC sets' intra-distances are represented by the
orange lines (inter sets).
It can be seen from Figure 7.9 that inter-set distributions of pitch count (PC) feature
for both Hicaz and Nihâvent Makams align with intra-SymbTr and intra-ATMMC sets
on the x-axis. As demonstrated in Section 4.2.4, this alignment shows that the two
datasets are harmonious.
Figure 7.9 PDFs of pitch count (PC) feature.
68
It can be seen from Figure 7.10 and 7.11 that inter-set distributions of note count
(NC) and pitch class histogram (PCH) features for both Hicaz and Nihâvent Makams
align with intra-SymbTr and intra-ATMMC sets on the x-axis. Again, this alignment
shows that the two datasets are coherent with each other.
Figure 7.10 PDFs of note count (NC) feature.
Especially in Figure 7.11 Nihâvent part, it is visible that the three distributions align
both in the x-axis and y-axis very well. This well-established alignment is also
supported by the data from Table 7.4.
Figure 7.11 PDFs of pitch count histogram (PCH) feature.
69
Again, in Figure 7.12 it is obviously visible that the three distributions align both
in the x-axis and y-axis very well, especially for Hicaz Makam. This alignment can be
interpreted as ATMMC’s ability to mimic note lengths (NLH) in SymbTr is sufficient.
Figure 7.12 PDFs of note length histogram (NLH) feature.
However, in Figure 7.13 pitch range (PR) features are not as well-aligned as other
features. This means ATMMC squeezes the highest and lowest notes into a smaller
interval than the samples in the SymbTr dataset. Since the highest and the lowest
pitches may be considered as outliers, this result can be interpreted as ATMMC’s
lacking in representing the source dataset’s outliers.
Figure 7.13 PDFs of pitch range (PR) feature.
70
Finally, it can be seen from Figure 7.14 that the distributions for average pitch
interval (PI) are aligning very well for both Makams. As also supported by the data in
Table 7.4, this alignment can be interpreted as ATMMC’s ability to manage the
distance of jumps between consecutive pitches very well.
Figure 7.14 PDFs of average pitch interval (PI) feature.
71
CHAPTER EIGHT
CONCLUSION AND FUTURE WORK
8.1 Conclusion
Deep Learning (DL) based music composition has become a popular field of
research due to the availability of large digital music datasets, state-of-the-art DL
technologies, and high-power computer systems. Although the majority of DL-based
music research is carried on Western Music, there are also few studies on Turkish
Music. However, the majority of these DL-based Turkish Music research is in the field
of music information retrieval (MIR), and the number of studies on automatic music
composition is negligible.
In the scope of this thesis, a DL-based symbolic music generation system, Artificial
Turkish Makam Music Composer (ATMMC) was developed by training Long Short-
Term Memory (LSTM) networks on SymbTr, which is the most comprehensive open-
source dataset, compiled for computational Turkish Music research.
As the first of its kind and novel system, ATMMC can create artificial compositions
in Hicaz and Nihâvent Makams in Şarkı Forms. Also, for the system to be used by
users without programming knowledge, a graphical user interface that can be accessed
via the internet browser at http://music.cs.deu.edu.tr/tmmgui/ has been developed. Via
the graphical user interface, users can enter 8 initial notes and command ATMMC to
complete their compositions. When the automatic composition process ends, users can
download the resulting composition to their computers in Mu2 format, to be viewed
and played by Mus2 software later.
The effectiveness of the ATMMC system has been evaluated both subjectively
through a survey and objectively with various metrics. As the result of subjective
evaluations, the system was found to be in between above-average and good segments.
Objective evaluations show that ATMMC is around 84% similar to the dataset on
which it was trained. In addition, due to the availability of a slightly larger dataset,
72
artificial compositions in Hicaz Makam are found to be slightly better than the artificial
compositions in Nihavent Makam.
When subjective and objective evaluation results are addressed as a whole, it can
be concluded that ATMMC can be used to assist human composers in creating new
TMM pieces. It also can help composers to overcome writer’s block by quickly
offering composition drafts. Finally, ATMMC has been found to be useful in quickly
creating copyright-free etudes for conservatory students.
8.2 Future Work
Automatic TMM composition is mostly an untouched research field and is widely

open to experimentation. Plans for future studies can be listed as follows, but not
limited to:
1. Training different DL models on TMM data and compare their results to find the
most suitable one for TMM.
2. Widening the scope of ATMMC by training it on other Makams.
3. Creating a fully functional web application for TMM notation.
4. Creating a new open-source digital TMM dataset compiled from the saved pieces
revealed by the use of the to-be-developed web application.
5. Optimizing the future ATMMC models into JavaScript for front-end web.
6. To provide suggestions to users through DL models in the web application to be
developed.
73
REFERENCES
Abadi, M., Agarwal, A., Barham, P., Brevdo, E., Chen, Z., Citro, C., et al. (2016).
TensorFlow: Large-scale machine learning on heterogeneous distributed systems.
Arxiv Preprint Arxiv, 1603.04467.
Abidin, D., Öztürk, Ö., & Özacar Öztürk, T. (2017). Klasik Türk müziğinde makam
tanıma için veri madenciliği kullanımı. Gazi Üniversitesi Mühendislik-Mimarlık
Fakültesi Dergisi, 32(4), 1221–1232. https://doi.org/10.17341/gazimmfd.369557
Altinköprü, H. (2018). Maqam modulations in traditional Turkish art music. Eurasian

Journal of Music and Dance, 12, 1–11.
Arel, H. S. (1968). Türk musikisi nazariyatı dersleri. İstanbul: Hüsnütabiat Matbaası.
Aydoğan, S., & Özgür, Ü. (2015). Gelenekten geleceğe makamsal Türk müziği.
Ankara: Arkadaş Yayıncılık.
Barapatre, D., & A, V. (2017). Data preparation on large datasets for data science.
Asian Journal of Pharmaceutical and Clinical Research, 10(13), 485.
https://doi.org/10.22159/ajpcr.2017.v10s1.20526
Barkçin, S. (2019). 40 Makam 40 anlam (D. Yabul, Ed.; 1st ed.). İstanbul: Ketebe.
Bernstein, A., & Taylor, B. (2003). TuneJS. Retrieved September 16, 2020, from
https://github.com/abbernie/tune
Bozkurt, B., Ayangil, R., & Holzapfel, A. (2014). Computational analysis of Turkish
makam music: Review of state-of-the-art and challenges. Journal of New Music
Research, 43(1), 3–23. https://doi.org/10.1080/09298215.2013.865760
74
Bozkurt, B., Yarman, O., Karaosmanoğlu, M. K., & Akkoç, C. (2009). Weighing
diverse theoretical models on Turkish maqam music against pitch measurements:
A comparison of peaks automatically derived from frequency histograms with
proposed scale tones. Journal of New Music Research, 38(1), 45–70.
https://doi.org/10.1080/09298210903147673
Brain Team, G. (2020). Magenta. Retrieved January 12, 2019, from

https://magenta.tensorflow.org
Briot, J.-P., Hadjeres, G., & Pachet, F. (2017). Deep learning techniques for music
generation - a survey. ArXiv Preprint ArXiv: 1709.01620.
Brownlee, J. (2019). How to calculate the KL divergence for machine learning.

Retrieved February 8, 2021, from
https://machinelearningmastery.com/divergence-between-probability-
distributions/
Burkholder, J. P., Grout, D. J., & Palisca, C. (2010). A history of Western music (10th
ed.). New York City: W. W. Norton.
Choi, K., Fazekas, G., & Sandler, M. (2016). Text-based LSTM networks for
automatic music composition. ArXiv Preprint ArXiv:1604.05358
Chollet, F. (2015). Keras: The Python deep learning library. Retrieved June 11, 2019,
from https://keras.io/
Chu, H., Urtasun, R., & Fidler, S. (2016). Song from PI: A musically plausible network
for pop music generation. 5th International Conference on Learning
Representations, ICLR 2017 - Workshop Track Proceedings, 1–9.
75
Cope, D. (1989). Experiments in musical intelligence (EMI): Non‐linear linguistic‐
based composition. Interface, 18(1–2).
https://doi.org/10.1080/09298218908570541
Deng, L., & Yu, D. (2014). Deep learning: Methods and applications. Foundations
and Trends in Signal Processing 7(3–4) 197-387.
Ederer, E. B. (2011). The theory and praxis of makam in classical Turkish music 1910-
2010. Santa Barbara: University of California.
Erguner, S. (2007). Osman Dede, nayi. TDV İslâm Ansiklopedisi, 33, 461–462.
Gönül, M. (2015). Türk mûsı̇ kîsi usûllerı̇ nı̇ n gösterı̇ mı̇ , ifadesı̇ ve tasnı̇ fı̇ ne bı̇ r bakış.
İSTEM, 13(25), 31–46.
Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep learning. In MIT Press.
Cambridge: MIT Press.
Güvençoğlu, Ş., & Özgelen, O. Z. (2020). Türk makam müzikçisinin seyir defteri - 1.
İstanbul: Pan Yayıncılık.
Hananoi, S., Muraoka, K., & Kiyoki, Y. (2016). A music composition system with
time-series data for sound design in next-generation sonification environment.
2016 International Electronics Symposium (IES), 380–384.
Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural

Computation, 9(8), 1735–1780. https://doi.org/10.1162/neco.1997.9.8.1735
76
Huang, C.-Z. A., Hawthorne, C., Roberts, A., Dinculescu, M., Wexler, J., Hong, L., &
Howcroft, J. (2019). The Bach doodle: Approachable music composition with
machine learning at scale. Proceedings of the 20th International Society for Music
Information Retrieval Conference, ISMIR 2019, 793–800.
http://arxiv.org/abs/1907.06637
Kapadia, S. (2019). Language models: N-gram. A step into statistical language.

Retrieved March 22, 2020, from https://towardsdatascience.com/introduction-to-
language-models-n-gram-e323081503d9
Karaosmanoğlu, M. K. (2012). A Turkish makam music symbolic database for music

information retrieval: Symbtr. Proceedings of the 13th ISMIR Conference, Porto,
Portugal, ISMIR, 223–228.
http://compmusic.upf.edu/system/files/static_files/KemalKaraosmanoglu_Ismir20
12.pdf
Karaosmanoğlu, M. K. (2017). Müzik aritmetiği ve ses sistemleri (1st ed.). İstanbul:

İTÜ Vakfı.
Kumar, H., & Ravindran, B. (2019). Polyphonic music composition with LSTM neural
networks and reinforcement learning. ArXiv Preprint ArXiv: 1902.01973.
Li, S., Jang, S., & Sung, Y. (2019). Automatic melody composition using enhanced
GAN. Mathematics, 7(10), 883. https://doi.org/10.3390/math7100883
Liang, F., Gotham, M., Johnson, M., & Shotton, J. (2017). Automatic stylistic
composition of Bach chorales with deep LSTM. Ismir 2017.
Luque, S. (2009). The stochastic Synthesis of Iannis Xenakis. Leonardo Music

Journal, 19, 77–84.
77
Marinescu, A.-I. (2019). Bach 2.0 - generating classical music using recurrent neural
networks. Procedia Computer Science, 159, 117–124.
https://doi.org/10.1016/j.procs.2019.09.166
Olah, C. (2015). Understanding LSTM networks. Retrieved March 11, 2021, from
http://colah.github.io/posts/2015-08-Understanding-LSTMs/
Oord, A. van den, Dieleman, S., Zen, H., Simonyan, K., Vinyals, O., Graves, A.,
Kalchbrenner, N., Senior, A., & Kavukcuoglu, K. (2016). WaveNet: A generative
model for raw audio. ArXiv, 21(3), 793–830.
https://doi.org/10.1162/neco.2008.04-08-771
Özkan, İ. H. (2006). Türk musikisi nazariyatı ve usulleri kudüm velveleleri (18th ed.).
İstanbul: Ötüken Neşriyat.
Parlak, İ. H. (2020). TMMDL. Retrieved April 5, 2021, from

https://github.com/ihpar/TMMDLFT
Parlak, İ. H., & Kösemen, C. (2018). Automatic music generation by true random
numbers for Turkish makams. 2018 4th International Conference on Computer
and Technology Applications (ICCTA), 64–68.
https://doi.org/10.1109/CATA.2018.8398657
Parlak, İ. H., Çebi, Y., & Işıkhan, C. (2021). A Graphical User Interface for Deep
Learning-Based Automatic Turkish Makam Music Composer. 2021 11th
International Hisarli Ahmet Symposium, 125–126.
Parlak, İ. H., Çebi, Y., Işıkhan, C., & Birant, D. (2021). Deep learning for Turkish
makam music composition. Turkish Journal of Electrical Engineering & Computer
Sciences. https://doi.org/10.3906/elk-2101-44
78
Pastore, M. (2018). Overlapping: A R package for estimating overlapping in empirical
distributions. Journal of Open Source Software, 3(32).
https://doi.org/10.21105/joss.01023
Rojas, R. (1996). The backpropagation algorithm. In Neural Networks (149–182).

Berlin: Springer.
Sandred, Ö., Laurson, M., & Kuuskankare, M. (2009). Revisiting the Illiac suite - a
rule-based approach to stochastic processes. Sonic Ideas/Ideas Sonicas, 2, 42–46.
https://www.researchgate.net/publication/260791942
Şenocak, E. (2012). Tarı̇ hı̇ süreç içı̇ nde Türk müzı̇ ğı̇ nde şarki formu. Retrieved April
4, 2020, from http://earsiv.halic.edu.tr/xmlui/bitstream/handle/
20.500.12473/1658/342441.pdf?sequence=1&isAllowed=y
Şentürk, S. (2017). SymbTr. Retrieved February 7, 2018, from

https://github.com/MTG/SymbTr
Şentürk, S., & Chordia, P. (2011). Modeling melodic improvisation in Turkish folk
music using variable-length Markov models. 12th International Society for Music
Information Retrieval Conference (ISMIR 2011), 269–274.
Şentürk, S., & Serra, X. (2016). Composition identification in Ottoman-Turkish

makam music using transposition-invariant partial audio-score alignment. 13th
Sound and Music Computing Conference (SMC 2016), 434–441.
Shin, A., Crestel, L., Kato, H., Saito, K., Ohnishi, K., Yamaguchi, M., Nakawaki, M.,
Ushiku, Y., & Harada, T. (2017). Melody generation for pop music via word
representation of musical properties. ArXiv, 1–9. http://arxiv.org/abs/1710.11549
79
Szandała, T. (2020). Review and comparison of commonly used activation functions
for deep neural networks. In Bio-inspired Neurocomputing (203-224). Singapore:
Springer.
Tıraşçı, M. (2019). Türk musikisi nazariyatı tarihi. İstanbul: Kayıhan.
Tüfekçi, A. (2014). Exploring Ney Techniques (1st ed.). İstanbul: Pan Yayıncılık.
Tüfekçioğlu, S. (2019). Osmanlıda yenilikçi hareketlerle birlikte Türk mûsikîsine

eklenen yeni türler ve bu türlerin teori-pratik ikilemi. Güzel Sanatlar Enstitüsü
Dergisi, 173–182. https://doi.org/10.32547/ataunigsed.523556
Turing, A. (1950). Computing machinery and intelligence. Mind, 59(236), 433–460.

https://doi.org/10.1093/mind/LIX.236.433
Uyar, B., Atlı, H. S., Şentürk, S., Bozkurt, B., & Serra, X. (2014). A Corpus for
computational research of Turkish makam music. Proceedings of the 1st
International Workshop on Digital Libraries for Musicology - DLfM ’14, 1–7.
https://doi.org/10.1145/2660168.2660174
Uygun, M. N. (2008). Safiyüddin el-Urmevî. TDV İslâm Ansiklopedisi 35, 479–480.
Van Houdt, G., Mosquera, C., & Nápoles, G. (2020). A review on the long short-term
memory model. Artificial Intelligence Review, 53(8), 5929–5955.
https://doi.org/10.1007/s10462-020-09838-1
Vanderplas, J. (2007). Density estimation — scikit-learn 0.24.2 documentation.

Retrieved February 21, 2021, from https://scikit-
learn.org/stable/modules/density.html
Węglarczyk, S. (2018). Kernel density estimation and its application. ITM Web of
Conferences, 23, 00037. https://doi.org/10.1051/itmconf/20182300037
80
Weiss, K., Khoshgoftaar, T. M., & Wang, D. (2016). A survey of transfer learning.
Journal of Big Data, 3(1), 9. https://doi.org/10.1186/s40537-016-0043-6
Wright, O., & Turabi, A. H. (2001). Klasik Türk mûsikîsi’nde Çârgâh: Tarih-teori
çelişkisi. M.Ü. İlahiyat Fakültesi Dergisi, 21(1977), 81–104.
Wu, J., Hu, C., Wang, Y., Hu, X., & Zhu, J. (2020). A hierarchical recurrent neural
network for symbolic melody generation. IEEE Transactions on Cybernetics,
50(6), 2749–2757. https://doi.org/10.1109/TCYB.2019.2953194
Xu, X. (2020). LSTM networks for music generation. 1–7.

http://arxiv.org/abs/2006.09838
Yang, L.-C., & Lerch, A. (2020). On the evaluation of generative models in music.
Neural Computing and Applications, 32(9), 4773–4784.
https://doi.org/10.1007/s00521-018-3849-7
Yarman, O. (2007). A comparative evaluation of pitch notations in Turkish makam

music: Abjad scale & 24-tone Pythagorean tuning – 53 equal division of the octave
as a common grid. Journal of Interdisciplinary Music Studies, 1(1), 51–62.
https://doi.org/10.13140/RG.2.2.14971.72483
Yarman, O. (2008). 79-tone tuning & theory for Turkish maqam music as a solution
to the non-conformance between current model and practice. Doktora Tezi,
İstanbul Teknik Üniversitesi, İstanbul.
Yarman, O. (2010). Türk makam müziğini bilgisayarda temsil etmeye yönelik başlıca
yazılımlar. Müzikte Temsil Müziksel Temsil Sempozyumu II, 320–327.
https://doi.org/10.13140/RG.2.2.24566.19526
81
Zarate, J. M., Ritson, C. R., & Poeppel, D. (2012). Pitch-interval discrimination and
musical expertise: Is the semitone a perceptual boundary? The Journal of the
Acoustical Society of America, 132(2), 984–993.
https://doi.org/10.1121/1.4733535
Zhang, Y., Liu, W., Chen, Z., Li, K., & Wang, J. (2021). On the properties of
Kullback-Leibler divergence between gaussians. http://arxiv.org/abs/2102.05485
82

Turkish Makam Music Composition by Using Deep Learning Techniques

Uploaded by

Copyright:

Available Formats

Turkish Makam Music Composition by Using Deep Learning Techniques

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

Turkish Makam Music Composition by Using Deep Learning Techniques

Uploaded by

Copyright:

Available Formats

DOKUZ EYLÜL UNIVERSITY

GRADUATE SCHOOL OF NATURAL AND APPLIED SCIENCES

TURKISH MAKAM MUSIC COMPOSITION BY

A Thesis Submitted to the

We have read the thesis entitled “TURKISH MAKAM MUSIC COMPOSITION

Prof. Dr. Yalçın ÇEBİ

Prof. Dr. Cihan IŞIKHAN Assoc. Prof. Dr. Derya BİRANT

Thesis Committee Member Thesis Committee Member

Examining Committee Member Examining Committee Member

Prof. Dr. Okan FISTIKOĞLU

İsmail Hakkı PARLAK

Keywords: Deep learning, Turkish Makam Music, artificial intelligence, algorithmic

Bu tez çalışması kapsamında derin öğrenme teknikleri kullanarak otomatik olarak

Ph.D. THESIS EXAMINATION RESULT FORM .................................................. ii

CHAPTER ONE - INTRODUCTION ..................................................................... 1

1.1 Overview .......................................................................................................... 1

1.2 Purpose ............................................................................................................. 2

1.3 Contribution to Literature................................................................................. 3

1.4 Organization of the Thesis ............................................................................... 3

CHAPTER TWO – LITERATURE REVIEW ....................................................... 5

2.1 Evolution of Turkish Music and Its Theory ..................................................... 5

2.2 Turkish Makam Music Concepts and Western Music ..................................... 8

2.3 Deep Learning and Automatic Music Composition ....................................... 11

2.4 Evaluation Methods for Artificial Music Compositions ................................ 13

3.1 SymbTr Turkish Makam Music Symbolic Data Collection........................... 16

3.1.1 Statistical Review of SymbTr Dataset .................................................... 17

3.1.2 SymbTr Dataset N-gram Analysis .......................................................... 22

3.2 Dataset Preparation & Representation ........................................................... 25

CHAPTER FOUR – TECHNIQUES AND EVALUATION METHODS .......... 27

4.1 Long Short-Term Memory Networks ............................................................ 27

4.1.1 Activation Functions ............................................................................... 28

4.1.2 Long Short-Term Memory Cell Working Principles .............................. 29

4.2 Probability Distribution Analysis of Various Features .................................. 32

4.2.1 Pairwise Cross Validation ....................................................................... 34

4.2.2 Kernel Density Estimation ...................................................................... 34

4.2.3 Kullback-Leibler Divergence and Overlapped Area ............................... 35

4.2.4 Method Demonstration ............................................................................ 36

CHAPTER FIVE – AUTOMATIC COMPOSER ................................................ 40

5.1 Base Models ................................................................................................... 40

5.2 Specialist Models ........................................................................................... 42

5.3 Conductor Models .......................................................................................... 46

5.4 Full System Overview .................................................................................... 48

6.1 Purpose of the Graphical User Interface ........................................................ 51

6.2 Graphical User Interface Decomposition ....................................................... 51

CHAPTER SEVEN – RESULTS & EVALUATION ........................................... 58

7.1 Subjective Evaluation ..................................................................................... 58

7.2 Objective Evaluation ...................................................................................... 61

7.2.1 Pitch Density Analysis ............................................................................ 62

7.2.2 ATMMC Detailed Analysis Results ....................................................... 64

CHAPTER EIGHT – CONCLUSION AND FUTURE WORK .......................... 72

8.1 Conclusion ...................................................................................................... 72

8.2 Future Work ................................................................................................... 73

Throughout the history, great composers of various cultures either exploited

Towards the 2000s, the availability of relatively low-cost general-purpose graphics

1.3 Contribution to Literature

Computational research on TMM is not as abundant as its western music

Another contribution is the web-based GUI that is developed to utilize ATMMC.

1.4 Organization of the Thesis

In Chapter 2, a comprehensive literature review is given covering the topics of