181
DOI: 10.4312/mz.61.2.181-199
UDK: 781”17”Tartini G.(0.032)=131.1:004.352.242
A Digital Approach to Tartini’s Theoretical 
Works: Training and Testing a Custom AI Model 
for Handwritten Text Recognition*617
Jerneja Umer Kljun
University of Ljubljana
ABSTRACT
The article describes the digitisation of Giuseppe Tartini’s manuscripts with the help of hand-
written text recognition via the Transkribus app. It explains the basics concepts, describes 
the training process for a customised model and the tests run on Tartini’s scientific texts. 
The results of private and public AI models are compared in terms of accuracy, readability 
and efficiency of the editing process.
Keywords: Giuseppe Tartini, eighteenth century, HTR, Transkribus, digital humanities, 
digital musicology
IZVLEČEK
Članek prinaša podroben opis procesa digitalizacije rokopisov Giuseppeja Tartinija z uporabo 
programske opreme Transkribus, namenjene samodejnemu prepoznavanju in transkribiranju 
rokopisnega gradiva. V članku so predstavljena osnovna načela tehnologije HTR, opisana 
sta postopek učenja namenskega modela UI za transkribiranje ter preizkus dela z različnimi 
modeli pri transkribiranju Tartinijevih znanstvenih besedil. Sledi primerjava rezultatov, pri-
dobljenih z dvema zasebnima in enim javno dostopnim modelom za transkripcijo italijanskih 
besedil, in sicer z vidika natančnosti, berljivost in učinkovitosti, pri čemer je upoštevan tudi 
čas urejanja in revizije samodejno generirane transkripcije.
Ključne besede: Giuseppe Tartini, 18. stoletje, HTR, Transkribus, digitalna humanistika, 
digitalna muzikologija
* The author acknowledges the financial support from the Slovenian Research and Innovation 
Agency (research core funding no. P6-0446).
muzikološki zbornik • musicological annual lxi/2
182
Introduction
The past decade has seen a renewed interest in Giuseppe Tartini, especially 
due to two consecutive Interreg project (tARTini1 and TartiniBIS2) dedicated 
to the legacy of the composer as well as the preservation and promotion of 
his cultural heritage. However, even after the publication of Tartini’s letters in 
2020,3 and the launch of a digital collection4 and a thematic catalogue5 in 2025, 
much more remains to be uncovered.
The rapid advancements in artificial intelligence have created novel oppor-
tunities to access historical records. The article, meant as an account of an on-
going research project, presents a possible further step in the digitisation of 
Tartini’s scientific texts kept at the Piran branch of the Regional Archives Ko-
per (Pokrajinski arhiv Koper – Archivio regionale di Capodistria, hereinafter 
SI-PAK) through the integration of Handwritten Text Recognition (HTR) 
technology for automatic or assisted transcription provided by the Transkribus 
platform (READ Coop).6
Although an increasing number of researchers and institutions have begun 
exploring the use of HTR technology in various contexts,7 there is still room 
to further investigate its practical implementation and potential, as the field is 
constantly evolving. This article contributes to that growing body of scholarly 
work by providing a detailed account of the HTR software used, namely the 
Transkribus platform, including the training process for a tailored text reco-
gnition model, i.e. a trained AI algorithm, able to detect the most probable 
sequence of characters for each text line,8 its advantages and limitations, and 
its possible applications in future interdisciplinary research and teaching, espe-
cially in the field of digital musicology.
To this end, the article first gives a brief introduction to the general concepts 
of HTR and describes the main features of the Transkribus app. It then focuses 
1 “TARTini”, Interreg Italia – Slovenija, accessed September 5, 2025, https://2014-2020.ita-slo.eu/
it/tartini.
2 “TartiniBIS”, Interreg Italia – Slovenija, accessed September 5, 2025, https://www.ita-slo.eu/it/
tartini-bis.
3 Giorgia Malagò (ed.), Giuseppe Tartini / Pisma in Dokumenti / Letters and Documents, vol. 1, transl. 
Jerneja Umer Kljun and Roberto Baldo (Trieste: EUT Edizioni Università di Trieste, 2020).
4 “Archives and Catalogue”, Discover Tartini, accessed September 5, 2025, https://www.
discovertartini.eu/archivi?type=1&lang=en.
5 “Thematic Catalogue”, Discover Tartini, accessed September 5, 2025, https://www.discovertartini.
eu/archivi?type=5&lang=en.
6 “Transkribus”, Transkribus, accessed August 20, 2025, https://www.transkribus.org/.
7 Joe Nockels et al., “Understanding the Application of Handwritten Text Recognition Technology 
in Heritage Contexts: A Systematic Review of Transkribus in Published Research”, Archival 
Science 22, no. 3 (2022): 367–392.
8 “Training Text Recognition Model”, Transkribus, accessed September 5, 2025, https://help.
transkribus.org/training-text-recognition-models.
Jerneja Umer Kljun: A Digital Approach to Tartini’s Theoretical Works
183
on the archival sources selected for a test run with Transkribus, describing the 
main features of Tartini’s theoretical manuscripts. In addition, an experiment 
was conducted to test the effectiveness of the app by comparing the results 
of automatic transcription with two custom AI models with the results of a 
publicly available AI model for handwritten Italian, in terms of readability, the 
ability of the model to accurately recognize numbers, abbreviations, ratios and 
fractions, as well as the estimated time needed for editing a single automati-
cally recognized page.
A Brief Introduction to Handwritten Text Recognition
Handwritten text recognition, i.e. the technology that generates machine- 
readable transcriptions from images of manuscripts, is certainly not a novelty 
– according to Muehlberger et al., it is “an active research area in the computa-
tional sciences dating back to the mid-twentieth century”.9 Originally closely 
linked to the development of optical character recognition (OCR), which is 
used to convert scanned images of printed text into machine-encoded text, it 
quickly became “a research area in its own right”10 due to the computational 
complexity involved in dealing with the specificity of handwriting, with its 
frequent deviations from the norm, as well as the variability of different hand-
writing styles.11 And while the OCR problem has been solved rather early and 
effectively, as there is (only) a finite number of fonts that can be used in print, 
which allows the software to be programmed to recognize and read them all, 
the nearly infinite number of possibilities rendered the same task impossible 
when dealing with handwriting.12
The HTR technology made considerable leaps forward with the advent and 
improvement of artificial neural networks.13 With the use of AI models and 
Deep Learning, HTR software is now able to learn to recognize hand writing 
in an image, using “the model’s knowledge to transcribe that handwriting into 
digital text”.14 But as the topic of deep neural network approaches exceeds the 
scope of this article, readers should refer to the cited sources as well as the 
numerous user-friendly resources available on the READ-COOP platform 
9 Guenter Muehlberger et al., “Transforming Scholarship in the Archives through Handwritten 
Text Recognition: Transkribus as a Case Study”, Journal of Documentation 75, no. 5 (2019): 995.
10 Ibid., 955.
11 Ibid.; Jože Glavič, “Primer uporabe programa Transkribus in izdelava modela za avtomatsko 
optično prepoznavanje znakov za poenostavljeno transkribiranje ročno pisane gotice”, Moderna 
arhivistika 3, no. 1 (2020): 87.
12 “What Is HTR and How Does It Work”, Transkribus blog, accessed September 5, 2025, https://
blog.transkribus.org/en/what-is-handwriting-recognition-and-how-does-it-work.
13 Ibid., 79.
14 “What Is HTR and How Does It Work”.
muzikološki zbornik • musicological annual lxi/2
184
for a detailed overview of the technology behind Transkribus.15 From a user’s 
perspective, it is sufficient to note that HTR: “is now a mature machine learn-
ing tool capable of producing accurate, machine-processable text from images 
of historical manuscripts […] speeding up the transcription of primary sources 
and facilitating full text searching and analysis of historic texts at scale.”16
Working with Transkribus
Transkribus is a comprehensive AI-powered, user-friendly HTR software, de-
veloped within the framework of the Recognition and Enrichment of Archival 
Documents (READ) European Union Horizon 2020 project, which is now 
run by the READ-COOP cooperative.17 The Transkribus app allows users 
to quickly transcribe handwritten documents using one of the many publicly 
available text recognition models (at the moment of writing this article, there 
are 325 public models trained on a variety of scripts and languages). Howe-
ver, it is much more effective if it is trained to recognise a specific hand. The 
workflow for training a custom text recognition model in Transkribus consists 
of several steps, which can be roughly summarised as follows:
i. Layout recognition (segmentation):
after uploading an image of a manuscript, the automatic layout recognition 
function segments the image into text regions and lines that connect the 
text and the image and serve as a reference point for text recognition;
ii. manual transcription or text-image matching and editing (data preparation):
the editing software shows the layout editor and the text editor simulta-
neously, allowing the user to manually transcribe the page line by line. The 
transcription can also be enriched by using textual tags to mark-up words 
(e.g. names, dates, places) or to add other information about the manuscript. 
Once completed, the transcription can be downloaded in different formats 
(image, searchable PDF, docx, text file and XML). It is now also possible to 
align pre-existing transcriptions with images of manuscripts;18
iii. training:
accurately transcribed pages serve as ground truth, i.e. accurate and veri-
fied data used for training and testing AI models.19 To train a solid HTR 
model, it is advisable to manually transcribe at least 10,000 words for each 
15 Transkribus Blog, accessed September 5, 2025, https://blog.transkribus.org/en.
16 Nockels et al., “Understanding the Application of Handwritten Text Recognition Technology in 
Heritage Contexts”, 368.
17 Muehlberger et al., “Transforming Scholarship in the Archives through Handwritten Text 
Recognition”.
18 “Text-Image Matching”, Transkribus, accessed September 5, 2025, https://help.transkribus.org/
import-existing-transcriptions-with-text-image-matching.
19 “What Is Ground Truth?”, Transkribus, accessed September, 2025, https://help.transkribus.org/
import-existing-transcriptions-with-text-image-matching.
Jerneja Umer Kljun: A Digital Approach to Tartini’s Theoretical Works
185
hand. After choosing the training data, i.e. ground truth data on which 
the knowledge of the model is based, and the validation data, i.e. a set of 
examples from the same collection that serves as evaluation of the mo-
del, the users can proceed to set up and train their model.20 During the 
training process, the validation data is set aside to serve to test the model’s 
performance. By comparing the automatically transcribed pages against 
the accurate manually transcribed data, the model counts all the incor-
rectly recognized characters, giving a “score” to the overall performance. 
The metric used is the character error rate or CER, i.e. the number of errors 
ex pressed as a percentage of the total number of transcribed characters.21
This exact process has been applied to a selection of documents described in 
the next section to train a custom model for the transcription of Tartini’s the-
oretical work.
Towards a Digitised Tartini Collection
The Tartini family collection is succinctly described in Pucer’s introduction to 
the catalogue:22
[The Tartini collections] was previously stored at the “Sergej Mašera” Ma-
ritime Museum in Piran, where some of the documents were also displayed 
as exhibits. The collection was organised in a makeshift and non-archival 
manner. The documents were numbered; however, they were not organised 
chronologically or by subject. An inventory of sorts was included in the form 
of a manuscript written by Igor Cvetko. In 1986, the collection was acquired 
by the Piran branch of PAK.23
Pucer started working on the collection in 1991, compiling a detailed inven-
tory of the material. The collection, covering the period from 1654 to 1951, 
comprises 11 archival units, i.e. 1.1 linear meter, organised in seven thematic 
sections, indexed with the letters A through G, namely A – Family records; 
B – Scientific texts; C – Copies of Tartini’s work; D – Commemoration of 
the 200th anniversary of Tartini's birth in Piran; E – Print; F – Photography; 
20 “Setup and Training”, Transkribus Blog, accessed September, 2025, https://help.transkribus.org/
model-setup-and-training.
21 “How Is CER Calculated in Transkribus?”, Transkribus Blog, accessed September 5, 2025, https://
blog.transkribus.org/en/how-is-the-cer-calculated-in-transkribus.
22 Albert Pucer, Giuseppe Tartini: Inventar zbirke / Inventario della collezione: 1654–1951 (Koper: 
Pokrajinski arhiv Koper, 1993).
23 “Arhivska zbirka ‘Giuseppe Tartini’ je bila prej shranjena v Pomorskem muzeju ‘Sergej Mašera’ v 
Piranu, kjer so nekateri dokumenti služili tudi kot razstavni eksponati. Tu je bila zbirka zasilno 
in nearhivsko urejena. Dokumenti so bili oštevilčeni, vendar niso bili urejeni ne po zadevah ne 
kronološko. V rokopisu je bil izdelan tudi neke vrste inventar, delo mag. Igorja Cvetka. Leta 1986 
je zbirko dobil PAK – enota v Piranu.” Pucer, Giuseppe Tartini, 7, quotation translated by Jerneja 
Umer Kljun.
muzikološki zbornik • musicological annual lxi/2
186
G – Miscellanea. Each section is further divided into subsections marked with 
roman numerals (I – XIV) and the documents are organised chronologically.24
Within the framework of the TartiniBIS project, Pucer’s catalogue served 
as the basis for the identification of previously undiscovered or undigitised 
archival records. A comprehensive examination of the digital and physical ar-
chive uncovered various problems, such as difficult access, incomplete digiti-
sation, digitisation errors and missing or misplaced documents.25 And while 
the missing materials have now been digitised and published on the disco-
vertartini.eu website,26 the PDF documents and images available online are 
still scattered across various databases and out of context. To address the issue 
of document accessibility, further efforts could also be made by applying the 
methods presented here.
Description of the Digitally Transcribed Manuscripts
The small selection of seven digitally transcribed documents from the “scienti-
fic” section of the Tartini fonds presented in this article is built upon a manual 
transcription of Tartini’s unpublished treatise Quadratura del circolo (see entry 
#1 in Table 1 below), previously discussed by Sukljan27 and Umer Kljun28 in 
the volume In Search of Perfect Harmony: Tartini’s Music and Music Theory in 
Local and European Contexts.29 The manual transcription of the relatively short 
treatise proved to be optimal as ground truth, as it yielded the first successful 
version of the Giuseppe Tartini model, with a CER of 2.3%. Six more short 
unpublished documents that have not been transcribed before were then cho-
sen from the Tartini fonds, t. u. 3 and 4 (scientific texts), automatically transcri-
bed and manually edited to serve as ground truth for further model training 
and testing.
24 Ibid., 7.
25 “Five Studies and New Sources”, Discover Tartini, accessed September 5, 2025, https://www.
discovertartini.eu/main/pagina/19/Five-studies-and-new-sources?lang=en.
26 “Nuove fonti dagli archivi di Pirano”, Discover Tartini, accessed September 5, 2025, https://www.
discovertartini.eu/index.php/archivi/pirano.
27 Nejc Sukljan, “Tartini and the Ancients: Traces of Ancient Music Theory in the Tartini-Martini 
Correspondence”, in In Search of Perfect Harmony: Tartini’s Music and Music Theory in Local and 
European Contexts, ed. Nejc Sukljan (Berlin: P. Lang, 2022), 141–167.
28 Jerneja Umer Kljun, “Understanding Tartini and His Thought: Overcoming Translation 
Difficulties in the Correspondence between Tartini and Martini”, in In Search of Perfect Harmony: 
Tartini’s Music and Music Theory in Local and European Contexts, ed. Nejc Sukljan (Berlin: P. Lang, 
2022), 245–260.
29 Nejc Sukljan (ed.), In Search of Perfect Harmony: Tartini’s Music and Music Theory in Local and 
European Contexts (Berlin: P. Lang, 2022).
Jerneja Umer Kljun: A Digital Approach to Tartini’s Theoretical Works
187
Ta
bl
e 1
: Th
e D
ig
iti
se
d 
H
T
R
 T
ar
tin
i C
ol
lec
tio
n
E
nt
ry
 #
Id
en
tif
ie
r
# 
Pa
ge
s 
Tr
an
-
sc
rib
ed
W
or
ki
ng
 ti
tle
O
pe
ni
ng
 li
ne
s
G
en
re
C
on
te
nt
s, 
th
em
es
1
SI
 P
A
K
 P
I 
33
4 
Z
bi
rk
a 
G
iu
se
pp
e 
Ta
rti
ni
,
t. 
u.
 4
, a
. u
. 2
32
52
Q
ua
dr
at
ur
a d
el 
cir
co
lo
“S
i è
 sc
op
er
to
 u
n 
fe
no
m
en
o 
ar
m
on
ico
, p
er
 d
i c
ui
 m
ez
zo
 si
 
pr
et
en
de
 o
tte
nu
ta
 la
 q
ua
dr
at
ur
a d
el 
cir
co
lo
 […
]”
sc
ie
nt
ifi
c 
te
xt
th
ird
 to
ne
;
sq
ua
rin
g 
of
 th
e c
irc
le
2
SI
 P
A
K
 P
I 
33
4 
Z
bi
rk
a 
G
iu
se
pp
e 
Ta
rti
ni
,
t. 
u.
 3
, a
. u
. 1
43
7
R
isp
os
ta
 a 
cri
tic
o 
ig
no
to
“M
on
sie
ur
, S
e l
a f
at
ica
 d
i t
an
ti 
an
ni
 
pe
r l
a r
ice
rc
a d
el 
ve
ro
 p
rin
cip
io
 
de
ll’
ar
m
on
ia
 n
on
 m
i a
ve
ss
e r
ec
at
o 
alt
ro
 va
nt
ag
gi
o 
[..
.]”
sc
ie
nt
ifi
c 
te
xt
,
let
te
r
th
ird
 to
ne
;
a d
ef
en
ce
 o
f t
he
 co
ns
eq
ue
nc
es
 
de
du
ce
d 
fro
m
 th
e t
hi
rd
 to
ne
3
SI
 P
A
K
 P
I 
33
4 
Z
bi
rk
a 
G
iu
se
pp
e 
Ta
rti
ni
,
t. 
u.
 3
, a
. u
. 1
50
8
Pr
os
eg
ui
m
en
to
 e 
Co
m
pi
m
en
to
 d
ell
a 
Ca
rta
 m
us
ica
le
“P
ro
se
gu
im
en
to
, e
 co
m
pi
m
en
to
 
de
lla
 C
ar
ta
 M
us
ica
le.
 L
a p
rim
a 
pr
op
os
iz
io
ne
 fù
 l’
ac
co
rd
o 
co
ns
on
an
te
 
sim
ul
ta
ne
o 
de
lla
 se
stu
pl
a a
rm
on
ica
 
[..
.]”
di
da
ct
ic 
m
at
er
ia
l
m
us
ic 
th
eo
ry
;
a p
ap
er
 o
n 
ra
tio
s, 
pr
og
re
ss
io
ns
, 
in
te
rv
als
 an
d 
no
ta
tio
n
4
SI
 P
A
K
 P
I 
33
4 
Z
bi
rk
a 
G
iu
se
pp
e 
Ta
rti
ni
,
t. 
u.
 3
, a
. u
. 1
51
30
D
iss
er
ta
zi
on
e 
m
us
ica
le 
su
i p
rin
cip
i 
de
ll’
ar
m
on
ia
“L
’ A
ut
or
e, 
ch
e n
el 
su
o 
Tr
at
ta
to
 d
i 
M
us
ica
 se
co
nd
o 
la 
ve
ra
 S
cie
nz
a 
de
ll'
ar
m
on
ia
 […
]”
sc
ie
nt
ifi
c 
te
xt
th
e t
ru
e f
ou
nd
at
io
n 
of
 h
ar
m
on
y
5
SI
 P
A
K
 P
I 
33
4 
Z
bi
rk
a 
G
iu
se
pp
e 
Ta
rti
ni
,
t. 
u.
 3
, a
. u
. 1
53
24
Pa
rte
 T
er
za
“I
n 
qu
es
ta
 te
rz
a p
ar
te
 s’
in
te
nd
e 
de
m
os
tra
re
, c
he
 la
 sc
ie
nz
a d
el 
pr
es
en
te
 si
ste
m
a s
ia
 id
en
tic
am
:te  
qu
ell
a s
te
ss
a d
e P
itt
ag
or
ici
 e 
di
 
Pl
at
on
e [
…
]”
sc
ie
nt
ifi
c 
te
xt
a j
us
tif
ica
tio
n 
of
 th
e p
ro
po
se
d 
ha
rm
on
ic 
sy
ste
m
 th
ro
ug
h 
an
 
ela
bo
ra
tio
n 
of
 P
lat
o’s
 T
im
ae
us
muzikološki zbornik • musicological annual lxi/2
188
E
nt
ry
 #
Id
en
tif
ie
r
# 
Pa
ge
s 
Tr
an
-
sc
rib
ed
W
or
ki
ng
 ti
tle
O
pe
ni
ng
 li
ne
s
G
en
re
C
on
te
nt
s, 
th
em
es
6
SI
 P
A
K
 P
I 
33
4 
Z
bi
rk
a 
G
iu
se
pp
e 
Ta
rti
ni
,
t. 
u.
 3
, a
. u
. 1
56
12
Ev
id
en
tis
sim
a 
di
m
os
tra
zi
on
e
“E
vi
de
nt
:m
a d
im
os
tra
zi
on
e 
de
ll’
ar
m
on
ica
 n
at
ur
a d
i q
ua
nt
ità
 a 
pr
io
ri.
”
sc
ie
nt
ifi
c 
te
xt
th
e i
nt
rin
sic
 h
ar
m
on
ic 
na
tu
re
 o
f 
th
e O
ne
 as
 a 
wh
ol
e
7
SI
 P
A
K
 P
I 
33
4 
Z
bi
rk
a 
G
iu
se
pp
e 
Ta
rti
ni
,
t. 
u.
 4
, a
. u
. 2
41
3
Co
nf
ro
nt
o d
ei 
pr
in
cip
j
“C
on
fro
nt
o 
de
i p
rin
cip
j n
el 
pi
an
o 
di
m
os
tra
tiv
o. 
Pr
in
cip
j i
po
te
tic
i: 
pu
nt
o, 
lin
ea
, e
 su
pe
rfi
cie
: d
ell
e 
Sc
ie
nz
e C
om
un
i [
…
]”
sc
ie
nt
ifi
c 
te
xt
a c
om
pa
ris
on
 o
f t
he
 fu
nd
am
en
ta
l 
pr
in
cip
les
 o
f t
he
 co
m
m
on
ly
 
kn
ow
n 
sc
ie
nc
es
 an
d 
th
e 
pr
in
cip
les
 o
f T
ar
tin
i’s
 h
ar
m
on
ic 
sc
ie
nc
e
Jerneja Umer Kljun: A Digital Approach to Tartini’s Theoretical Works
189
All seven autographs, written in eighteenth century cursive, present similar 
traits and while Tartini’s penmanship is quite legible and consistent, there are 
some orthographic variations that could prove useful when placing the do-
cuments on a timeline or when trying to connect various scattered fragments. 
For example, there is some oscillation in the use of double consonants, such as 
-z- and -zz- in mezo, mezi and mezzo, mezzi (mean, means).
Furthermore, “Parte terza”30 includes segments in Latin, in which Tartini 
employs the e caudata [ę] for æ (see Figure 1), and three of the seven digitised do-
cuments present an uncommon diacritic above the letter e that does not appear 
elsewhere and the (inconsistent) use of which remains uncertain (see underlined 
words in Figure 1 and excerpts from various documents in Figure 2).
Figure 1: Use of the e caudata in Latin script. 
Figure 2: Uncommon diacritics in Tartini’s manuscript.
In Tartini’s texts, the most frequently abbreviated words are Italian adverbs 
of manner ending in -mente (e.g. demostrativamente, universalmente) and 
qua litative adjectives describing roots, ratios and progressions (e.g. armonico, 
30 SI PAK PI 334 Zbirka Giuseppe Tartini, t. u. 3, a. u. 153 (“Parte Terza”).
muzikološki zbornik • musicological annual lxi/2
190
aritmetico, geometrico) or superlatives (verissimo), all categories abbreviated with 
the use of a colon and superscript (e.g. demostrativam:te, geom:che, ver:mo). Other 
methods of abbreviation include the use of a macron over an omitted letter or a 
longer stroke indicating contractions or suspensions, e.g. prople for proporzio-
nale, caplo for capitolo, as well as various forms of ordinal indicators, e.g. 2dō, 
3zā and even sesqui3zā, as shown in Figure 3 below.
Figure 3: Frequently used abbreviations.
In general, the mise-en-page is relatively straightforward, facilitating rapid au-
tomatic layout recognition. However, most of the digitised theoretical texts are 
characterized by the presence of specific graphical elements, such as musical 
and mathematical symbols that appear both in-line, which means they are 
embedded in the sentence structure and therefore must be transcribed for the 
sentence to be complete, as well as separate, floating elements,31 see Figure 4 
below. Similarly, Tartini makes frequent use of illustrative examples of both 
musical and mathematical nature, i.e. geometric drawings, sketches, calcula-
tions and musical examples, see Figure 5 below.
Since the primary goal was to obtain an effective model, a “hyper-diplo-
matic” approach to transcription proved to be the best option in the data pre-
paration stage, “so as not to confuse the computer model when it is comparing 
transcriptions to the image of the manuscript”.32 This means that the manual 
transcription follows the source material as closely as possible, including non-
standard punctuation and capitalization, superscript, un-expanded abbrevia-
tions etc. However, due to the specificities of Tartini’s text some adjustments 
and simplifications had to be made (e.g. omission of in-line graphics when a 
suitable Unicode character was not available).
31 Fabian C. Moss et al., “Digitizing a 19th-Century Music Theory Debate for Computational 
Analysis”, in Proceedings of the Conference on Computational Humanities Research 2021: Amsterdam, 
the Netherlands, November 17–19, 2021. https://ceur-ws.org/Vol-2989/short_paper31.pdf.
32 Bram Caers, “Teaching Handwritten Text Recognition: Can New Technologies Save Old Skills?”, 
Quaerendo 54, nos. 2-3 (2024): 207.
Jerneja Umer Kljun: A Digital Approach to Tartini’s Theoretical Works
191
Figure 4: In-line (bass clef, stave) and floating elements (the diatonic scale with  
its corresponding ratios).
Figure 5: Floating elements – corda or linea sonora AB [line segment AB representing a 
string]; a semicircle illustrating the harmonic system contained in AB.
muzikološki zbornik • musicological annual lxi/2
192
Testing Model Versions
To test the two versions of the model specifically trained on Tartini’s hand-
writing, confronting the fully automatic transcription to the one achieved with 
a publicly available model for Italian, a single page from the Tartini archive, 
which has not been previously transcribed, has been selected. To ensure the 
best possible outcome with each of the selected AI models, the test page had to 
present minimal “disturbance”, i.e. clear handwriting, little to no cancellations, 
corrections, additions, ink bleed-through or lacerations. It also had to include 
abbreviations, numbers, ratios, fractions and calculations, as they are consis-
tently present in Tartini’s theoretical work. The effectiveness of the automated 
transcription was then assessed in terms of readability, the ability of the model 
to correctly transcribe abbreviations, ratios and fractions, as well as the estima-
ted time needed for editing a single page.
The models used to transcribe the selected page from “Appunti sulla pro-
porzione armonica”33 were Transkribus Italian handwriting M134 (trained by the 
Transkribus team on 653,630 words; 6.70% CER) and my own private models 
Giuseppe Tartini I (trained on 26,402 words; 2.30% CER) and Giuseppe Tar-
tini II (trained on 51,355 words; 3.06% CER). The Italian handwriting M1 is a 
generic model trained on a diverse dataset spanning from the sixteenth to the 
nineteenth century and it is occasionally updated with community data. Giu-
seppe Tartini I is a private model trained on the manual transcription of Tarti-
ni’s unpublished manuscript “Quadratura del circolo” (diplomatic transcripti-
on; abbreviations tagged; expansions provided within the tag). Giuseppe Tartini 
II is a private model built upon the previous version of the model. It is trained 
on the Tartini Collection: one manual transcription of Tartini’s unpublished 
manuscript “Quadratura del circolo”; six documents, automatically recognized 
with Giuseppe Tartini I model, which were then edited and tagged (diplo-
matic transcription; abbreviations tagged; expansions provided within the tag; 
documents feature tables and special Unicode characters; training included 
expanding abbreviations).
The differences between the resulting transcriptions are best illustrated by 
the examples in Tables 2 and 3 below, in which errors are highlighted. 
33 Biblioteca del Conservatorio di musica Giuseppe Tartini, IT-TS0108, Collezione Tartiniana: 
scritti teorici, musiche e documenti relativi a Giuseppe Tartini in Trieste e Pirano: GT/FA/23 ms 
1, “Appunti sulla proporzione armonica”.
34 “Transkribus Italian Handwriting M1”, READ Coop, accessed September 5, 2025, https://app.
transkribus.org/models/text/38440.
Jerneja Umer Kljun: A Digital Approach to Tartini’s Theoretical Works
193
Figure 6: Excerpt from “Appunti sulla proporzione armonica” (GT/FA/23 ms1),  
top of the page.
Table 2: Unedited automatic transcriptions with selected models
Pu
bl
ic 
m
od
el
Alla dimostrazione del trattato contenuta nel Capuoto secondo, assegnata per sopo 
il Circolo
armonico per intrinseca natura della sua costruzione et soglione, che quanto i
si dimostra per numesi linean animatici ridotto a proporzione alm.te di spese, non 
rese
al numero mune a rimetto attratto da linee, pare, e predazioni. Perche dati dice
paesi qualunque moltiglicati tra loro, il prodotto e mezzo proporzionale de due da
G
iu
se
pp
e T
ar
tin
i I
Alla dimostratione del trattato contenusa nel circolo secondo assegnata per provati 
il circolo
armonico per intrinseca natura, della sula costragione etc:, si ossone, che quanto ivi
si dimostra per numeri lineari aritmetici, ridotti a proporzione geom:ca discreta, 
non regge
al numero comune aritmetico astratto da linee, figare, e proporzioni. Perchè dati 
due
numeri qualunque moltiplicati tra loro, il prodotto è mezzo proporzionale de due 
dati
G
iu
se
pp
e T
ar
tin
i I
I
Alla dimostrazione del Trattato contenuta nel capitolo secondo assegnata per 
provare il circolo
armonico per intrinseca natura della sua costruzione etc:, si ossione, che quanto ivi
Si dimostra per numeri linea aritmetici ridotti a proporzione geom:ca discreta, non 
regge
al numero comune aritmetico astratto da linee, figure, e proporzioni. Perchè dati 
due
numeri qualunque moltiplicati tra loro, il prodotto è mezzo proporzionale de due 
dati
Similarly, if we observe a section with a more complicated layout, the efficacy 
of the models becomes even more obvious:
muzikološki zbornik • musicological annual lxi/2
194
Figure 7: Excerpt from “Appunti sulla proporzione armonica” (GT/FA/23 ms1),  
bottom of the page.
Table 3: Unedited automatic transcriptions with selected models
Pu
bl
ic 
m
od
el
ne puo esser ma ragion semplice, ma è sempre ragione, inevitabilm.te composte di 
tre ragioni espresse nell’
nell’esempio, qui sopra et poi
488
142
dalle lino
A. A da vera 3
846
243
la citta siam.te dipende la dimostrazione del cerchio armonico per intrinseca la 
natura esse essendo vero de il quadra
del seno [BD] e mezzo [prople] non tra i quadrati di £ 8, e però in tal rispetto di 
pom.a natura, e ver altrettanto e più che
G
iu
se
pp
e T
ar
tin
i I
ne può esser mai ragion semplice, ma è sempre ragione inevitabilm:te composta di 
tre ragioni espresso nelle
nell’essenqui qui sopra esposta
AB, BC.
1 a 2.
dalle linee,
Ab, Ac da umeri 1 a 3
BC, AC.
– 2 a 3
Sa ciò sottavialm:te dipende la dimostratione del cerchio armonico per intrinseca 
sia natura. Perchè essendo verte che il quadrato
del seno BD e mezzo prople tra quadrati di AB, BC, e però in tal rispetto di 
geom:ca natura, è veri altrettanto, e sia, che
G
iu
se
pp
e T
ar
tin
i I
I
nè puo esser mai ragion semplice, ma è dempre ragione inevitabilm:te composta di 
tre ragioni espresse nell’
nell’esempio qui sopra esposta
AB, BC
1 a 2
dalle linee
AB, AC, da numeri 1 a 3
BC, AC
2 a 3
Da ciò sostanzialm:te dipende la dimostrazione del cerchio armonico per 
intrinseca sua natura. Perchè essendo ved:mo che il quadrato,
del seno BD è mezzo prople tra i quadrati di AB, BC, e però in tal rispetto di 
geom:ca natura, è veri altrettanto e più, che
Jerneja Umer Kljun: A Digital Approach to Tartini’s Theoretical Works
195
Discussion
In general, it can be observed that using a public model when automatically 
transcribing Tartini’s script is not really a time-saving solution. As expected, 
the model used in this comparison proved to be unsuitable for the task, as it 
had difficulties recognizing abbreviations and even differentiating letters and 
numbers, which yielded a nearly illegible, nonsensical transcription that would 
possibly take longer to edit than transcribing the document manually. The two 
versions of the private model give far better results, even though several cor-
rections are still necessary. While the text is readable and errors mainly occur 
at character level (e.g. interchanging the letters q, g, z, p and r, f and s), a bigger 
issue stems from an inaccurate layout recognition and line segmentation (see 
Table 3 above) – an issue that can be easily prevented with proper layout edit-
ing before proceeding to text recognition.
To quantify these observations, each of the automated transcriptions was 
manually corrected and edited, the editing process was timed, and the version 
history in Transkribus and the comparison function in Word were used to 
identify the changes in the resulting transcriptions. While it only took a little 
over a minute to automatically transcribe a page with each of the models, it 
took over 47 minutes to correct the layout and the text transcribed with the 
public AI model. A comparison of the exported text files transcribed with Ita-
lian Handwriting M1 showed that 370 revisions were made to the text itself, 
while the Transkribus version history feature calculated a 20.79% CER and a 
57.52 % WER (word error rate) on the analysed page. In comparison, it took 
25 minutes to complete the same task on the page transcribed using the Giu-
seppe Tartini I model (199 text revisions detected by document comparison; 
5.72% CER and 15.22% WER calculated with Transkribus’ version history 
feature), and 19 minutes to correct the layout and text transcribed with the AI 
Giuseppe Tartini II model (98 revisions detected by document comparison, 
3.47% CER and 8.48% WER calculated with Transkribus’ version history fe-
ature). It should be noted that half of that time was spent on layout editing, 
which means the text editing will be quicker if the layout of the pages with a 
more complex structure is carefully prepared in advance.
To summarise, the performance of the second version of the private model 
is already quite satisfactory, but it could be improved further with more data 
and “fine-tuning”, as there are still some issues to be addressed. For example, 
most floating elements had to be excluded from the training process because 
the software cannot yet recognise musical notation or other graphics. However, 
efforts have been made towards handwritten music recognition (OMR) wit-
hin Transkribus, which entails textual encoding of the written music.35 Whi-
35 “Jorge Calvo-Zaragoza – Handwritten Music and Text Recognition in Transkribus”, YouTube 
video, 08:02, accessed September 5, 2025, https://www.youtube.com/watch?v=TvDyF9L7sYE.
muzikološki zbornik • musicological annual lxi/2
196
le the method proposed by Calvo-Zaragoza36 did not seem efficient enough 
at this time or even compatible with training an HTR model, the automatic 
recognition of Tartini’s music may be well within reach in the near future.
In Lieu of a Conclusion: Possible Applications
As the digitisation of the Tartini collection is an ongoing project, there is no 
concrete conclusion to be made yet. It is easy to see, however, how this approach 
could be useful in compiling a fully digitised, machine-readable database of Tar-
tini’s scholarly work that could be adapted for research (e.g. corpus-based studi-
es) or critical editions. The process of model training is a time-consuming pro-
cess, nevertheless, the efficiency of the tool is truly expressed when working with 
a large collection of documents,37 creating a large data set. And as Nicholas Cook 
observed, “working with larger data sets will open up new areas of musicology”.38 
In addition, the use of HTR in university courses has already proven to be an 
engaging method for teaching palaeography and digital editing, revitalizing the 
discussion about text editions.39 With HTR (and AI in general), we are standing 
at a crucial moment in digital humanities, a moment that requires a sensible and 
well-considered approach, but nevertheless “a moment of opportunity”40.
Bibliography
Primary sources
Biblioteca del Conservatorio di musica Giuseppe Tartini, IT-TS0108
Collezione Tartiniana: scritti teorici, musiche e documenti relativi a Giuseppe Tartini 
in Trieste e Pirano
GT/FA/23 ms1 (“Appunti sulla proporzione”)
Pokrajinski arhiv Koper/Koper Regional Archives 
SI PAK PI 334 Zbirka Giuseppe Tartini 
t. u. 3
a. u. 143 (“Risposta a critico ignoto”)
a. u. 150 (“Proseguimento e Compimento della Carta musicale”)
a. u. 151 (“Dissertazione musicale sui principi dell’armonia”)
a. u. 153 (“Parte Terza”)
a. u. 156 (“Evidentissima dimostrazione”)
t. u. 4 
a. u. 232 (“Quadratura del circolo”)
a. u. 241 (“Confronto dei principj”)
36 Ibid.
37 Glavič, “Primer uporabe programa transkribus […]”, 88.
38 Nicholas Cook, “Towards the Complete Musicologist?”, ISMIR 2005, 4, https://ismir2005.ismir.
net/documents/Cook-CompleatMusicologist.pdf.
39 Caers, “Teaching Handwritten Text Recognition”.
40 Cook, “Towards the Complete Musicologist?”, 1.
Jerneja Umer Kljun: A Digital Approach to Tartini’s Theoretical Works
197
References
“Archives and Catalogue.” Discover Tartini. Accessed September 5, 2025. https://www.dis-
covertartini.eu/archivi?type=1&lang=en.
Caers, Bram. “Teaching Handwritten Text Recognition: Can New Technologies Save 
Old Skills?.” Quaerendo 54, nos. 2-3 (2024): 198–209. https://brill.com/view/journals/
qua/54/2-3/article-p198_6.pdf; DOI: 10.1163/15700690-BJA10024. 
Cook, Nicholas. “Towards the Complete Musicologist?.” ISMIR 2005. https://ismir2005.
ismir.net/documents/Cook-CompleatMusicologist.pdf.
“Five Studies and New Sources of Documentation Relating to Giuseppe Tartini, His Con-
temporaries, Pupils and Followers.” Discover Tartini. Accessed September 5, 2025. htt-
ps://www.discovertartini.eu/main/pagina/19/Five-studies-and-new-sources?lang=en.
Glavič, Jože. “Primer uporabe programa transkribus in izdelava modela za avtomat-
sko optično prepoznavanje znakov za poenostavljeno transkribiranje ročno pisane 
gotice.” Moderna arhivistika 3, no. 1 (2020): 86–97. DOI: https://doi.org/10.54356/
MA/2020/1/ECAD5603.
“How Is CER Calculated in Transkribus.” Transkribus Blog. Accessed September 5, 2025. 
https://blog.transkribus.org/en/how-is-the-cer-calculated-in-transkribus.
“Jorge Calvo-Zaragoza – Handwritten Music and Text Recognition in Transkribus.” You-
Tube video. 08:02. Posted by Transkribus on 12 october 2022. Accessed September 5, 
2025. https://www.youtube.com/watch?v=TvDyF9L7sYE.
Malagò, Giorgia, ed. Giuseppe Tartini / Pisma in Dokumenti / Letters and Documents. Vol. 1. 
Translated by Jerneja Umer Kljun and Roberto Baldo. Trieste: EUT Edizioni Univer-
sità di Trieste, 2020.
Moss, Fabian C., Maik Köster, Melinda Femminis, Coline Métrailler, and François Ba-
vaud. “Digitizing a 19th-Century Music Theory Debate for Computational Analy-
sis.” In Proceedings of the Conference on Computational Humanities Research 2021: 
Amsterdam, the Netherlands, November 17–19, 2021. https://ceur-ws.org/Vol-2989/
short_paper31.pdf.
Muehlberger, Guenter, Louise Seaward, Melissa Terras, Sofia Ares Oliveira, Vicente Bosch, 
Maximilian Bryan, Sebastian Colutto, et al. “Transforming Scholarship in the Archives 
through Handwritten Text Recognition: Transkribus as a Case Study.” Journal of Docu-
mentation 75, no. 5 (2019): 954–976. https://doi.org/10.1108/JD-07-2018-0114.
Nockels, Joe, Paul Gooding, Sarah Ames, and Melissa Terras. “Understanding the Applica-
tion of Handwritten Text Recognition Technology in Heritage Contexts: A System-
atic Review of Transkribus in Published Research.” Archival Science 22, no. 3 (2022): 
367–392. https://doi.org/10.1007/s10502-022-09397-0.
“Nuove fonti dagli archivi di Pirano.” Discover Tartini. Accessed September 5, 2025. https://
www.discovertartini.eu/index.php/archivi/pirano.
“OCR vs. HTR.” Transkribus Blog. Accessed September 5, 2025. https://blog.transkribus.
org/en/insights/ocr-vs-htr.
Pucer, Albert. Giuseppe Tartini: Inventar Zbirke / Inventario Della Collezione: 1654–1951.
Koper: Pokrajinski arhiv Koper, 1993.
“Setup and Training.” Transkribus Blog. Accessed September 5, 2025. https://help.transkri-
bus.org/model-setup-and-training.
Sukljan, Nejc, ed. In Search of Perfect Harmony: Tartini’s Music and Music Theory in Local and 
European Contexts. Berlin: P. Lang, 2022. DOI: 10.3726/b20325. 
Sukljan, Nejc. “Tartini and the Ancients: Traces of Ancient Music Theory in the Tartini-
Martini Correspondence.” In In Search of Perfect Harmony: Tartini’s Music and Music 
muzikološki zbornik • musicological annual lxi/2
198
Theory in Local and European Contexts, edited by Nejc Sukljan, 141–167. Berlin: P. Lang, 
2022.
“TartiniBIS.” Interreg Italia – Slovenija. Accessed September 5, 2025. https://www.ita-slo.
eu/it/tartini-bis.
“TARTini.” Interreg Italia – Slovenija. Accessed September 5, 2025. https://2014-2020.ita-
slo.eu/it/tartini.
“Text-Image Matching.” Transkribus. Accessed September 5, 2025. https://help.transkri-
bus.org/import-existing-transcriptions-with-text-image-matching.
“Thematic Catalogue.” Discover Tartini. Accessed September 5, 2025. https://www.discov-
ertartini.eu/archivi?type=5&lang=en.
“Training Text Recognition Model.” Transkribus. Accessed September 5, 2025. https://help.
transkribus.org/training-text-recognition-models.
Transkribus Blog. Accessed September 5, 2025. https://blog.transkribus.org/en .
“Transkribus Italian Handwriting M1.” READ Coop. Accessed September 5, 2025. https://
app.transkribus.org/models/text/38440 .
“Transkribus.” Transkribus. Accessed August 20, 2025. https://www.transkribus.org/.
Umer Kljun, Jerneja. “Understanding Tartini and His Thought: Overcoming Translation 
Difficulties in the Correspondence between Tartini and Martini.” In In Search of Perfect 
Harmony: Tartini’s Music and Music Theory in Local and European Contexts, edited by 
Nejc Sukljan, 245–260. Berlin: P. Lang, 2022.
“What Is Ground Truth?.” Transkribus. Accessed September 5, 2025. https://help.transkri-
bus.org/import-existing-transcriptions-with-text-image-matching .
“What Is HTR and How Does It Work.” Transkribus Blog. Accessed September 5, 2025. htt-
ps://blog.transkribus.org/en/what-is-handwriting-recognition-and-how-does-it-work.
SUMMARY
The article outlines the potential next steps in digitising Giuseppe Tartini’s scientific texts 
kept at the Piran branch of the Regional Archives Koper, through the integration of Hand-
written Text Recognition (HTR) technology for automatic or assisted transcription. It first 
gives a brief overview of the general concepts of HTR, describing the main features of the 
Transkribus app, including the training process for a customized text recognition model. 
The focus then shifts towards the primary sources selected for a test run with the HTR app, 
describing the main features of Tartini’s theoretical manuscripts. In addition, an experiment 
was conducted to test the performance of the trained model by comparing the results of two 
private custom AI models against the results obtained with a publicly available AI model 
for handwritten Italian. The resulting transcriptions were compared in terms of readability, 
the ability of the model to accurately recognize numbers, abbreviations, ratios and fractions, 
as well as the estimated time needed for editing a single automatically recognized page.
Jerneja Umer Kljun: A Digital Approach to Tartini’s Theoretical Works
199
POVZETEK
Digitalni pristop pri obravnavi Tartinijevih teoretičnih del: 
učenje in preizkušanje prilagojenega umetnointeligenčnega 
modela za prepoznavanje rokopisnih besedil
Članek zarisuje eno izmed možnih nadaljnjih poti pri digitalizaciji znanstvenih besedil Giu-
seppeja Tartinija, ki jih hrani piranska enota Pokrajinskega arhiva Koper, in sicer z uporabo 
tehnologije HTR (handwritten text recognition), namenjene samodejnemu prepoznavanju in 
transkribiranju rokopisnega gradiva. Uvod prinaša kratek pregled osnovnih načel tehnologije 
HTR in opisuje glavne značilnosti programske opreme Transkribus, ki omogoča učenje pri-
lagojenih umetnointeligenčnih modelov za prepoznavanje rokopisnih besedil. Po podrobni 
razčlenitvi temeljnih značilnosti izbranih Tartinijevih rokopisov, tj. primarnih virov, izbranih 
za preizkušanje programske opreme, sledi opis preizkusa učinkovitosti izbranega pristopa za 
obdelavo gradiva. Nazadnje je predstavljena primerjava rezultatov dveh zasebnih prilagoje-
nih umetnointeligenčnih modelov z rezultati, pridobljenimi z javno dostopnim modelom za 
italijanščino, in sicer z vidika natančnosti pri prepoznavanju in prepisovanju števil, krajšav, 
razmerij in ulomkov, berljivost besedila in splošni učinkovitosti modelov, pri čemer je upo-
števan tudi čas urejanja in revizije samodejno generirane transkripcije.
ABOUT THE AUTHOR
JERNEJA UMER KLJUN ( Jerneja.umerkljun@ff.uni-lj.si) is a translator, translation teach-
er and researcher exploring the fields of Translation Studies, Sociolinguistics and Digital 
Humanities. She teaches Italian-Slovene translation and several language courses at the 
Department of Translation Studies (Faculty of Arts, University of Ljubljana) as well as at 
the Academy of Music of the University of Ljubljana. She has translated Giuseppe Tarti-
ni’s letters into Slovene and authored several scientific articles and a monograph on code- 
switching and language attitudes.
O AVTORICI
JERNEJA UMER KLJUN ( Jerneja.umerkljun@ff.uni-lj.si) je prevajalka, visokošolska uči-
teljica in raziskovalka na področju prevodoslovja, sociolingvistike in digitalne humanistike. 
Je avtorica številnih strokovnih in znanstvenih člankov ter nedavno objavljene monografije o 
kodnem preklapljanju in odnosu do jezika. Na Oddelku za prevajalstvo Filozofske fakultete 
Univerze v Ljubljani poučuje italijanske jezikovne predmete ter prevajanje med italijanščino 
in slovenščino. Redno sodeluje z Akademijo za glasbo Univerze v Ljubljani, kjer poučuje 
italijanščino za pevce. Posveča se prevajanju avdiovizualnih vsebin ter leposlovnih, huma-
nističnih in zgodovinskih besedil, med katerimi velja zlasti omeniti prevod zbirke pisem 
Giuseppeja Tartinija.