ENABLING EMPIRICAL ANALYSIS OF PIANO PERFORMANCE
REHEARSAL WITH THE RACH3 MIDI DATASET
Alia Mo si1∗Suhi Chi u hapudi2∗Sil an Pe e 2I an Pilko 2
Lau a Bishop3Aki a Maezawa4Xa ie Se a1Ca los Cancino-Chacón2
1Music Technology G oup, Uni e si a Pompeu Fab a, Ba celona, Spain
2Ins i u e o Compu a ional Pe cep ion, Johannes Keple Uni e si y Linz, Aus ia
3RITMO Cen e o In e disciplina y S udies in Rhy hm, Time and Mo ion, Uni e si y o Oslo, No way
4Yamaha Co po a ion, Hamama su, Japan
[email p o ec ed], [email p o ec ed]
ABSTRACT
The s udy o piano ehea sals can o e in e es ing in-
sigh s in o he s a egies adop ed by a pianis in o de
o lea n, in e p e and e en ually pe o m musical pieces.
The analysis o ehea sal p ocesses equi es compu a ional
me hods ha di e om hose used o piano pe o mance,
due o challenges like mis akes, epe i ions o musical seg-
men s, o o wa d and backwa d skips o sec ions in he
piece. The sca ci y o publicly a ailable ehea sal da a
limi s he empi ical unde s anding o hese challenges.
We elease he Rach3 MIDI Da ase , an openly a ailable
collec ion o MIDI iles con aining mo e han 750 hou s
o eco dings o piano ehea sals and co esponding Mu-
sicXML sco es by ou pianis s (3 ad anced, 1 beginne ),
collec ed o e a pe iod o mo e han 4 yea s. This da ase
eco ds he p og ession o pianis s lea ning new epe oi e,
as well as p ac icing amilia pieces, all in he Wes e n
Classical adi ion. We desc ibe he ehea sal piece iden-
i ica ion p ocess used o au oma ically labeling a po ion
o he da a in his elease. Fu he mo e, we use he Rach3
da a o highligh se e al challenges and u u e esea ch di-
ec ions pe aining o he compu a ional analysis o piano
ehea sals, speci ically symbolic ehea sal- o-sco e align-
men , ehea sal s uc u e analysis, and au oma ic mis ake
iden i ica ion.
1. INTRODUCTION
Compu a ional analysis o music pe o mance has adi-
ionally ocused on he end p oduc , ha is, he ou come
o a ehea sal p ocess, a he han ehea sal i sel . Ye
musicians spend subs an ial ime on ehea sal. Analysis
* Equal con ibu ion.
© A. Mo si, S. Chi u hapudi, S. Pe e , I. Pilko , L. Bishop,
A. Maezawa, X. Se a and C. Cancino-Chacón. Licensed unde a C e-
a i e Commons A ibu ion 4.0 In e na ional License (CC BY 4.0). A -
ibu ion: A. Mo si, S. Chi u hapudi, S. Pe e , I. Pilko , L. Bishop, A.
Maezawa, X. Se a and C. Cancino-Chacón, “Enabling Empi ical Anal-
ysis o Piano Pe o mance Rehea sal wi h he Rach3 MIDI Da ase ”, in
P oc. o he 26 h In . Socie y o Music In o ma ion Re ie al Con .,
Daejeon, Sou h Ko ea, 2025.
o ehea sal has he po en ial o imp o e unde s anding o
music lea ning and expe ise de elopmen and suppo he
de elopmen o pedagogic ools. We de ine ehea sal as
goal-o ien ed, sys ema ic p ac ice wi h he aim o lea n-
ing and becoming p o icien in playing speci ic epe oi e.
While he same amilies o music analysis app oaches a e
applicable o da a om ei he pe o mances o ehea sals,
ehea sal da a poses speci ic challenges ha a e no p esen
in polished pe o mance. The mos no able examples a e
he p esence o mis akes, jumps be ween di e en pa s
o a piece, and non-composi ional epe i ions (i.e., playing
he same passage epea edly).
To da e, esea ch on music ehea sal has been hinde ed
by a lack o da a and app op ia e compu a ional ools. Fo
ad anced musicians, ehea sal is a p ocess ha can span
mon hs o yea s, and unde s anding ha p ocess equi es
a longi udinal pe spec i e, wi h da a collec ed a di e -
en s ages. This pape in oduces he Rach3 MIDI da ase ,
which con ains mo e han 750 hou s o eco dings o piano
ehea sals by ou pianis s, mos ly in ol ing music om
he Wes e n classical adi ion. The da ase allows o a
comp ehensi e and ecologically alid compu a ional, da a-
d i en analysis o piano ehea sal o e an ex ended pe iod,
which has been limi ed in p e ious esea ch due o ech-
nical cons ain s and da a a ailabili y (c ., he scale and
scope o he s udies by Cha in and colleagues [1,2]). The
da ase will be made publicly a ailable, and, o he bes
o ou knowledge, comp ises he la ges collec ion o pi-
ano ehea sal da a. Exis ing symbolic da ase s o analysis
o piano pe o mance (e.g., (n)ASAP [3], Vienna 4x22 [4]
and Ba ik [5]), ocus on polished pe o mances.
The Rach3 MIDI da ase will con ibu e o ehea sal e-
sea ch by enabling sys ema ic s udy o ehea sal decisions.
As musicians de elop expe ise on hei ins umen , hey
also de elop mo e e ec i e ehea sal s a egies. Beginne s
a e mo e likely o epea indi idual no es, whe eas mo e
expe ienced musicians end o epea musically cohe en
sec ions o measu es [6]. Among musicians o he same
le el, some o ganize hei ehea sal sessions acco ding o
lea ning goals, o example, ocusing sepa a ely on ech-
nical challenges and musical unde s anding, while o he s
wo k in a mo e undi e en ia ed way [7]. Figu e 1 shows
484
(a) Changing ehea sal s uc u e ac oss ime o Pianis 1 ehea s-
ing Rapsodia Mexicana No. 2 by Manuel Ponce. Rehea sal num-
be indica ed on he le .
A B C
(b) Va ia ions o music segmen s A, B, and C wi hin and ac oss
ehea sals. G een: MIDI sco e e e ence. Blue: Pe o mance in-
s ances ac oss ehea sals.
Figu e 1: E olu ion o p ac ice s uc u e ac oss he di e en phases o lea ning a piece.
a pianis ’s ehea sal s uc u ed in o segmen s ha e lec
changes in ocus o e ime.
This pape desc ibes da a collec ion (Sec ion 3) and
au oma ic labeling o ehea sal iles using inge p in ing
me hods (Sec ion 4). We highligh limi a ions o s a e-
o - he-a symbolic pe o mance- o-sco e alignmen o e-
hea sal da a (Sec ion 5) and p opose an al e na i e ap-
p oach o ehea sal s uc u e analysis inspi ed by pa -
e n disco e y and music s uc u e segmen a ion (Sec-
ion 6), including p elimina y a emp s a au oma ic sco e-
independen piano mis ake iden i ica ion (Sec ion 7). We
conclude wi h u u e esea ch di ec ions in he compu a-
ional analysis o piano ehea sals (Sec ion 8). 1
2. RELATED WORK
Resea ch on music ehea sal has been spa se o da e,
hough a ew s udies ha e examined ehea sal beha io s
like decision-making, goal-se ing, and p ac ice s a e-
gies [8–10]. E icsson e al. highligh ed he ole o de-
libe a e p ac ice in achie ing expe ise [11]. S udies by
Hallam [12,13], Sokolo skis [14] and Cha in [1,2,15–17]
in es iga ed ehea sal h ough obse a ions and e ospec-
i e accoun s, acking how p ac ice s a egies e ol e o e
ime. The b oade li e a u e on musical lea ning has exam-
ined how di e en p ac ice schedules a ec memo y o
pi ch and iming [18, 19]. Some esea ch has also in es i-
ga ed how isual a en ion (eye gaze) is spli be ween he
sco e and he hands du ing lea ning o piano pieces, and
how his is a ec ed by he music s uc u e [20].
Despi e his, ehea sal emains unde s udied in a da a-
d i en way, wi h much o he li e a u e based on case s ud-
ies. As no ed in Miksza’s e iew [9], no s udies ha e in-
ol ed mo e han 40 hou s o ehea sal eco dings (see Ta-
ble 1 in [9]). This is pa ly due o echnical and logis ical
limi a ions in cap u ing long- e m ehea sal da a and he
lack o e icien algo i hms o ex ac ele an in o ma ion
and pa e ns om such a la ge sou ce o da a. Win e s e
1The da ase can be downloaded om he companion websi e
h ps:// 3midi. ach3p ojec .com/ whe e u he examples
and isualiza ions a e a ailable.
al. [21] in oduced an audio-based me hod o au oma ic
p ac ice logging, o keep ack o which pieces we e pe -
o med du ing a ehea sal session. Tools ha e also been
de eloped o he au oma ic quan i a i e assessmen o pe -
o mance quali y [22,23].
3. RACH3 MIDI DATASET
The Rach3 MIDI da ase con ains o e 3,000 MIDI iles
om piano ehea sals pe o med mos ly on acous ic pi-
anos equipped wi h sys ems o enable MIDI cap u ing.
The da ase aims o be ep esen a i e o ypical ehea sal
p ac ices, ensu e ecological alidi y by e lec ing na u al
ehea sal condi ions, and emain comp ehensi e in scope
h ough di e se (i.e., mul imodal) da a sou ces o quan i-
a i e and quali a i e analysis [24]. The ull Rach3 da ase
is a mul imodal da ase ha includes synch onized audio
(cap u ed wi h mic ophones), MIDI, ideo om a came a
posi ioned o e he keyboa d, and w i en logs abou p ac-
ice s a egies and ocus. This pape ocuses solely on he
MIDI da a and o he modali ies will be add essed and e-
leased in u u e publica ions.
Da a collec ion began in Fall 2020 and now includes
o e 750 hou s o eco ded ehea sal sessions om ou pi-
anis s ( h ee ad anced, one beginne ; h ee o he pianis s
a e co-au ho s on his pape ), making his he la ges syn-
ch onized piano MIDI da ase o da e, 3.9 imes la ge han
he MAESTRO da ase (see Table 1). Figu e 2 shows a
cumula i e dis ibu ion o he pe o med no es and du a-
ion o e ime. The ad anced pianis s a e age 12.7±11.2
yea s o o mal aining a he conse a o y le el, wi h Pi-
anis s 1 and 2 holding unde g adua e o conse a o y de-
g ees in piano pe o mance. Pianis 3, a beginne , s a ed
lessons as pa o he p ojec in Summe 2024. Pianis 4
has unde g adua e-le el aining in piano pe o mance.
Rehea sals a e conduc ed on acous ic pianos equipped
wi h Silen sys ems, allowing o MIDI cap u e while p e-
se ing he na u al acous ic sound ia condense mic o-
phones. Pianis 1 uses a Yamaha GB1K Silen , Pianis 2
an Essex EUP-116E, and Pianis 4 a Yamaha Diskla ie
P oceedings o he 26 h ISMIR Con e ence, Daejeon, Ko ea, Sep embe 21-25, 2025
485
Pianis To al hou s To al no es (millions) A g. hou s pe session
All 769.6 20.9 0.94
P1 487.1 14.8 1.01
P2 142.4 3.6 0.85
P3 38.9 1.0 0.69
P4 101.2 1.5 0.93
MAESTRO 3 198.7 7.0 –
Ba ik 3.0 0.1 –
Table 1: Size compa ison o piano-cen ic da ase s wi h
synch onized MIDI and audio.
C1X. Pianis 3 eco ds on a Yamaha Cla ino a digi al pi-
ano, wi h he olume slide kep ixed on hei eache ’s
ecommenda ion.
Pianis s o ganize hei ehea sal sessions eely; ypical
ehea sals include echnical wa mups (e.g., Hanon exe -
cises, scales) and epe oi e p ac ice. The epe oi e se-
lec ion ocuses on wo a eas: lea ning new pieces om
sc a ch and main aining p e iously lea ned wo ks. This
allows o analysis o di e en ehea sal s a egies: ini ial
lea ning, ongoing main enance, and elea ning. Each ad-
anced pianis ocuses p ima ily ( hough no exclusi ely)
on speci ic epe oi e: Pianis 1 on Rachmanino ’s Piano
Conce o No. 3, Op. 30, Pianis 2 on G ieg’s Piano Con-
ce o, and Pianis 4 on Bee ho en’s Piano Sona as. Fo
p ac ical easons, con ibu ing pianis s concen a e on mu-
sic om he Common P ac ice Pe iod. 2O e 100 pieces
ha e been played (coun ing indi idual mo emen s sepa-
a ely). The da ase includes ehea sal o some ou -hands
piano due s. Fo hese, each pa (p imo and secondo) is
coun ed as a sepa a e piece.
In addi ion o MIDI, he da ase includes MusicXML
sco es o he pe o med wo ks. Mos sco es we e sou ced
om MuseSco e; 3whe e una ailable, we c ea ed hem
manually using MuseSco e based on p in ed edi ions o
IMSLP 4scans (p elimina y es s wi h OMR we e un-
success ul o he complex piano wo ks included in he
da ase ). This manual sco e en y is ongoing, wi h o e
hal o he da ase cu en ly co e ed. A ull epe oi e lis
is p o ided in he Appendix. 5
The da ase also includes li e pe o mances om he
D ess Rehea sal R3ci al Se ies, whe e con ibu ing pi-
anis s pe o m o a small li e and online audience. This
se ies se es o (1) p o ide a ealis ic goal o he e-
hea sals, (2) con as ehea sal and conce se ings, and
(3) simula e eal conce condi ions using he same mul i-
modal eco ding se up. Two eci als ha e been held o da e,
ea u ing Pianis 1 pe o ming wo ks by Manuel Ponce and
Modes Musso gsky. 6
This da ase is pa o an ongoing esea ch p ojec and
will con inue o g ow h ough addi ional pe o mances, an-
no a ions, and analysis.
2This pe iod co esponds oughly o he Ba oque, Classical, Roman-
ic, and ea ly 20 h Cen u y pe iods o Wes e n Classical music.
3h ps://musesco e.com
4h ps://imslp.o g/wiki/MainPage
5See Foo no e 1 .
6h ps:// 3ci als. ach3p ojec .com.
2023-012021-01 2025-012022-01 2024-01
Mon h
0
200
400
600
800
1000
1200
Numbe o No es (in housands)
No es pe Mon h
P4
P3
P2
P1
2023-012021-01 2025-012022-01 2024-01
Mon h
0
100
200
300
400
500
600
700
800
Cumula i e Du a ion (hou s)
Cumula i e Du a ion pe Mon h
P4
P3
P2
P1
Figu e 2: Cumula i e dis ibu ion o no es and du a ion in
he Rach3 MIDI da ase .
Weigh age Accu acy P ecision Recall F1
Mac o 0.991 0.941 0.922 0.929
Weigh ed 0.991 0.989 0.991 0.989
Table 2: E alua ion o symbolic inge p in ing based piece
iden i ica ion.
4. REHEARSAL PIECE IDENTIFICATION
Du ing app oxima ely he i s wo yea s o eco ding e-
hea sals, mul iple pieces we e p ac iced and eco ded in o
a single synch onized MIDI/audio/ ideo ake o e e y e-
hea sal session (i.e., a MIDI ile and i s co esponding syn-
ch onized audio and ideo iles). Because he came as
we e some imes o e hea ing du ing long eco dings, his
p ocess was la e modi ied so ha each p ac iced piece was
eco ded in o a sepa a e MIDI/audio/ ideo ake. Mo e e-
cen ly, pianis s s a ed labeling hese iles acco ding o he
piece name. Howe e , almos 60% o ehea sal pieces e-
mained unlabeled, equi ing a semi-au oma ic app oach o
piece iden i ica ion.
Fo his pu pose, we ollowed he symbolic inge p in -
ing me hod de eloped by A z e al. [25]. We c ea ed h ee-
no e okens om MusicXML sco es, gene a ed hashes o
hese okens, and s o ed hem in a lookup able, map-
ping each hash o he co esponding sco es. Tokens we e
hen ex ac ed om ehea sal eco dings, and hei hashes
we e ma ched agains he lookup able, wi h he highes -
ma ching sco e iden i ied as he p edic ed piece.
We i s an his algo i hm on 2155 labeled MIDI iles
om Pianis 1 and Pianis 2 con aining a single piece, and
whose espec i e sco es we e digi ally and publicly a ail-
able. The lookup able consis ed o he hashes o 71 such
digi al sco es. We assessed he algo i hm’s pe o mance
wi h his labeled da a and p o ide he esul s in Table 2.
P oceedings o he 26 h ISMIR Con e ence, Daejeon, Ko ea, Sep embe 21-25, 2025
486
Figu e 3: Con usion ma ix o he piece iden i ica ion
me hod applied o 2155 labeled MIDI iles om he
da ase , yielding an accu acy o 0.99 ac oss 71 sco es.
Fou pieces we e 100% misiden i ied, all o which ap-
pea ed in only one MIDI ile: Pachulski’s P elude in C Mi-
no Op. 8 No. 1, G ieg’s ‘Sol eig’s Song’, Mendelssohn’s
Songs Wi hou Wo ds Op. 30 No. 1 and Chopin’s P elude
Op. 28 No. 7. Among he o he pieces, he e we e some
misiden i ica ions bu mos we e iden i ied wi h 100% ac-
cu acy. Common easons o misiden i ica ion included:
1) Sho ehea sal du a ions which did no p o ide enough
hashes o comp ehensi ely ep esen he piece being p ac-
iced. 2) Many epe i ions o iny agmen s whose hashes
could easily belong in o he sco es. This is especially he
case o agmen s o ch oma ic scales ha a e likely o ap-
pea in mul iple sco es. 3) Many pi ch and iming e o s,
which some imes occu ed in ea ly ehea sals.
In he nex s ep, his algo i hm was used o p edic he
pieces in he emaining unlabeled MIDI iles. In he cases
whe e he e we e mul iple pieces wi hin a single MIDI ile,
a sepa a e p e-p ocessing s ep was added o segmen he
MIDI ile a poin s whe e he e was a long silence (>4s),
assuming ha his is he poin whe e he pianis swi ches
om one piece o ano he du ing he ehea sal. The in-
ge p in ing algo i hm was hen un on hese iles/segmen s
o p edic he piece being played. The pianis ’s ehea sal
log was used o iden i y he pieces ha we e played on he
gi en day, and he inge p in ing algo i hm sea ches o
hashes co esponding only o he sco es o hose pieces.
Manual e iew o his p ocess is ongoing.
5. CHALLENGES OF PERFORMANCE-SCORE
ALIGNMENT FOR REHEARSAL DATA
Alignmen is a c ucial i s s ep owa ds quan i a i e pe -
o mance analysis. In symbolic alignmen , no e-wise
alignmen e e s o he unique ma ching o indi idual
no es, i.e., a sco e no e may be ma ched o a single pe -
o mance no e o ma ked as a dele ion, and a pe o mance
no e may be ma ched o a single sco e no e o ma ked as an
inse ion. No e alignmen algo i hms do bes wi h a one- o-
one co espondence be ween no es in he sco e and no es
Figu e 4: Compa ison o he dis ibu ion o no e ma ches,
inse ions and dele ions o pe o mances o wo pieces
in he Rach3 MIDI da ase and he (n)ASAP and Ba ik
da ase s
in he pe o mance [3, 26, 27], which is no he case o
ehea sal da a.
Figu e 4 shows he dis ibu ion o no e ma ches, in-
se ions and dele ions o mul iple pe o mances o wo
pieces in he Rach3 MIDI da ase (Rachmanino ’s Piano
Conce o, No. 3, i s mo emen and Moza ’s Twel e
Va ia ions on “Ah ous di ai-je, Maman”, K 265). We
compa e he numbe o inse ions and dele ions o he
Rachmanino and Moza pieces o hose o Lisz pieces
in he (n)ASAP da ase [3] and Moza Sona as in he
Ba ik da ase [5]. To make hese compa isons, we an
pe o mance- o-sco e alignmen s using he GlueNo e [28],
a s a e-o - he-a symbolic alignmen me hod ha uses
lea ned ep esen a ions and is claimed o be sui able o
alignmen in he p esence o la ge misma ches. The igu e
shows mo e inse ions and dele ions in he ehea sal da a
han in he polished pe o mance da a o he (n)ASAP and
Ba ik da ase s. I he alignmen me hods we e adequa e,
we would expec he p opo ions o inse ions, dele ions
and ma ches o hese wo cases o be mo e simila . We do
expec mo e e o s in he ehea sal, bu no o he ex en
shown in he plo . A po en ial ac o in he e o a e dis-
c epancy is he ex a epe i ions occu ing du ing ehea sal.
6. REHEARSAL STRUCTURE ANALYSIS
The goal o compu a ional ehea sal s uc u e analysis is
o iden i y and g oup equi alen segmen epe i ions (see
Figu e 1), yielding insigh in o how ehea sals a e o ga-
nized. We de ine equi alen segmen s as sequences o pe -
o med no es ha co espond o he same sco e passage,
e en i pe o med wi h di e en in e p e a ions o mis akes
and a ying in leng h. Gi en a sco e, pe o mance seg-
men s can be linked o co esponding sco e segmen s, wi h
all pe o mance segmen s co esponding o he same sco e
segmen s ea ed as equi alen . In he absence o a sco e,
i is necessa y o iden i y and compa e segmen s wi hin a
pe o mance.
By conside ing ea lie wo k on pa e n-disco e y and
music s uc u e segmen a ion, we conclude ha Simila -
i y Ma ix app oaches a e be e sui ed han T ansla ional
Equi alence Class (TEC) [29] app oaches o his analy-
sis. TEC me hods ea music as a spa ial a angemen o
P oceedings o he 26 h ISMIR Con e ence, Daejeon, Ko ea, Sep embe 21-25, 2025
487
Rehea sal B
Rehea sal A
Rehea sal C
C oss Simila i y Ma ix B
C oss Simila i y Ma ix A
Bin Numbe
Bin Numbe
Bin Numbe
C oss Simila i y Ma ix C
Sco e Segmen A Sco e Segmen B
En i e Sco e
Figu e 5: C oss Simila i y Ma ices wi h diagonals (blue) showing ehea sal s uc u es o h ee sessions by Pianis 1 o
Rapsodia Mexicana No. 2. Despi e he ma ked discon inui y ( ed), ehea sal C shows a comple e un- h ough pe o mance
h ough a diagonal om op-le o bo om- igh . Rehea sal A demons a es epe i ions o he same sco e segmen g owing
p og essi ely longe . Rehea sal B shows a di e en mixed p ac ice: playing a segmen once, epea edly p ac icing i s
ending, hen e u ning o an ea lie sco e loca ion. Diagonal b eaks in ehea sals B and C esul om sco e o namen a ions
which esul in misma ches be ween he ehea sal and sco e cho d bins.
e en s and loca e exac epe i ions unde ansla ion (e.g.,
ansposi ion, e e sal) in mul idimensional space. They
explici ly sea ch o se s o poin s ha can be ans o med
in o each o he ia ansla ion and a o exac ness o he
pi ch and ime ela ionships, making hem unsui able gi en
he a iabili y p esen in ehea sal.
Simila i y Ma ix app oaches accommoda e inexac
epe i ions h ough lexible p ocessing a mul iple pipeline
s ages. These me hods compu e pai wise simila i y be-
ween elemen s in ea u e sequences, ei he be ween pe -
o mance and sco e (C oss Simila i y Ma ix, CSM) o
wi hin a single pe o mance (Sel Simila i y Ma ix, SSM).
Rela ed pa e ns eme ge as high-simila i y egions wi hin
hese ma ices, speci ically as diagonals mee ing c i e ia
o minimum leng h, simila i y h eshold, and gap ole -
ance, which a e hen conca ena ed and il e ed.
We p opose an ini ial sys em o ehea sal s uc u e
analysis based on ou de ini ion o equi alen segmen s,
using CSM and SSM o obse e di e en ins ances o e-
hea sal s uc u e in he Rach3 MIDI da ase (Figu e 5).
The pseudocode is a ailable in he Appendix on he com-
panion websi e.
A ‘cho di ica ion’ s ep handles he empo al a ia ion
ha occu s in ehea sal con ex s. No es wi h close onse s
(wi hin ∆ = 100 ms) a e g ouped in o a single ‘cho d’ bin
(a bina y 128-dimensional ec o wi h 1s a ac i e pi ch lo-
ca ions). In cho d bins, a 1 is only p esen a a no e’s onse
loca ion a he han a all loca ions whe e i is held, emo -
ing he e ec o no e o se di e ences. Fu he mo e, since
cho d bins a e only c ea ed when he e is a leas one no e
onse , silences a e emo ed om he sequence, elimina -
ing empo di e ences due o exp essi e choices o ins a-
bili y a i ac s. The MIDI sequences o be compa ed (pe -
o mance o sco e) a e ep esen ed as 128 ×nma ices.
Ins ead o measu ing simila i y be ween cho d bin pai s
h ough Euclidean dis ances, we p opose a mo e lexible
app oach. One o he 128×ncho d bin sequences (pe o -
mance o CSM, sco e o SSM) is con e ed o a 128 ×n
pi ch p o ile sequence, whe e he i- h pi ch p o ile (pi) is a
p obabili y dis ibu ion cap u ing pi ch ela ionships in he
local neighbo hood o each ac i e no e in he i- h cho d
bin (bi). Each piis ob ained by applying a local smoo h-
ing 1D con olu ion window (w) o he pi ch axis o each
bin, allowing us o ea each en y j∈ {0,...,127}in
pias he p obabili y o obse ing pi ch jin he con ex
o bi. Unde he assump ion ha pi ches in any bin bia e
bina y independen e en s, he simila i y be ween biand
any pi ch p o ile pkis exp essed as he likelihood o pk
ep esen ing he obse ed pi ches in bi, o mula ed as he
ollowing Be noulli likelihood:
L(pk|bi) =
127
Y
j=0
pbi,j
k,j ·(1 −pk,j)1−bi,j (1)
whe e pk,j ep esen s he p obabili y o pi ch jbeing
ac i e in p o ile k, and bi,j ∈ {0,1}indica es whe he
pi ch jis obse ed in cho d bin i. Applying his com-
pu a ion o each cho d bin iagains all pi ch p o iles pk
(whe e k= 0, . . . , n −1) cons uc s he simila i y ma ix
by conca ena ing he likelihood esul s o each cho d bin.
Fo CSM, we compa e pe o mance cho d bins wi h pi ch
p o iles om he sco e ( esul ing in an nsco e ×npe ma-
ix), whe eas o SSM, bo h pi ch p o iles and cho d bins
come om he pe o mance ( esul ing in an n×nma ix).
To ind ele an egions in he simila i y ma ix, we a-
P oceedings o he 26 h ISMIR Con e ence, Daejeon, Ko ea, Sep embe 21-25, 2025
488
e se diagonals o iden i y hose mee ing p especi ied min-
imum leng h, simila i y h eshold, and gap ole ance pa-
ame e s. The diagonals a e hen pos -p ocessed by g oup-
ing hem acco ding o ho izon al and e ical o e lap a ios
and me ging g oups based on diagonal in e sec ions. The
esul is g oups o diagonals, each e lec ing a unique e-
pea ed segmen .
Figu e 5 shows an applica ion o he ehea sal s uc u e
analysis desc ibed abo e o h ee s ages o ehea sal. In
he i s ( ehea sal A), he pianis p ac ices om a speci ic
s a ing poin , and g adually ex ends he p ac iced segmen
o include he nex sec ion in he piece. La e ( ehea sal B),
du ing a di e en ehea sal session, he pianis epea s a
segmen mul iple imes be o e mo ing o a second segmen
elsewhe e in he piece, which is also epea ed. Finally, e-
hea sal C ocuses on ull un- h ough ehea sals whe e he
goal is o play he piece om s a o inish. These end o
happen la e in he lea ning p ocess.
Quan i a i e e alua ion using common pa e n disco -
e y me ics [30] is no easible due o incompa ibili y wi h
ou de ini ion o equi alen segmen s; anno a ing a Rach3
MIDI e alua ion se is planned o u u e wo k. Though
simple, his app oach is ha d o une, as op imal hype pa-
ame e s depend on pe o mance de ails. In Figu e 5, pe -
o mance B illus a es how an unsui able ∆ o cho d bins
led o misma ched sco e and pe o mance segmen s. Ad-
di ional simila i y ma ices (see supplemen a y ma e ials)
show ha no e inse ions cause diagonal o se s, c ea ing
ex a sub-segmen g oups. Fu u e wo k should ocus on
p edic ing hype pa ame e s om MIDI da a, explo ing al-
e na i e cho d p o iles, and imp o ing diagonal g ouping
o handle inse ion-induced o se s.
7. SCORE-INDEPENDENT AUTOMATIC PIANO
MISTAKE IDENTIFICATION
E ec i e mis ake iden i ica ion sys ems can imp o e he
p ocessing o music ehea sal da a. In o ma ion abou p e-
dic ed mis ake loca ions and ypes can be inco po a ed in o
s uc u al analysis o alignmen pipelines and enable spe-
cialized analysis in hese a eas. Piano pe o mance mis-
akes a e ypically ca ego ized as pi ch o hy hm de ia-
ions om he sco e, wi h pe o mance- o-sco e alignmen
se ing as he p ima y iden i ica ion me hod. Gi en he
challenges highligh ed in Sec ions 5 and 6, we in es iga e
whe he he app oach p oposed in [31], which ains mod-
els o sco e-independen au oma ic iden i ica ion o con-
spicuous piano pe o mance mis akes, can label egions
ha migh be pa icula ly di icul o p ocess due o mis-
akes, such as he no e addi ions leading o he o se diag-
onals in Figu e 5.
In [31] a Tempo al Con olu ional Ne wo k (TCN) was
ained on a p i a e da ase o mis ake-anno a ed piano pe -
o mances, including bo h sigh - ead and p ac iced pe o -
mances. To compensa e o limi ed aining da a, hey
p e ained a TCN au oencode on a di e en p i a e se
o unanno a ed p o essional MIDI eco dings be o e ine-
uning wi h he anno a ed se . Howe e , his app oach
yielded only modes imp o emen s, likely due o domain
misma ch be ween p o essional-g ade da a and he es se .
Acco dingly, we in es iga e whe he we can eplica e
hei expe imen and ain a sco e-independen au oma ic
piano mis ake de ec ion model. We p e- ain hei same
TCN au oencode a chi ec u e wi h unlabeled iles om
he Rach3 MIDI da ase , ollowed by ine- uning a inal
classi ica ion laye wi h labeled syn he ic piano mis ake
da a gene a ed wi h he app oach and oolki in [32]. The
syn he ic mis akes oolki applies al e a ions o mis ake-
ee MIDI pe o mances based on a p oposed axonomy
o pe o mance mis akes. I e u ns a modi ied e sion o
an inpu pe o mance wi h he applied mis akes, and he
co esponding anno a ion ile wi h mis ake ypes and loca-
ions. We use he ecommended inpu pe o mance iles
indica ed on he oolki ’s webpage. 7This collec ion in-
cludes ac ual pe o mances (Vienna 4x22 [4], SMD [33],
and 32 iles om ASAP [34]) and music sco es wi hin
he beginne o in e media e p o iciency le els. The ou -
pu mis ake labels we e summa ized in o a bina y label
ma king he p esence o absence o a mis ake a disc e e
poin s in he esul ing piano oll. We use hese da a o c e-
a e ain, alida ion and es spli s and used he ansc ip-
ion p ecision/ ecall/F1 me ics o e alua ion since he es-
ima ed and g ound- u h anno a ions can be ea ed as no e
e en s a p ede ined pi ches. Ou bes aining con igu-
a ion achie ed 0.445 a e age F1-Measu e, 0.400 a e age
p ecision, and 0.528 a e age ecall on he es se (all o
syn he ic da a).
Al hough ini ial quali a i e obse a ions sugges ha
model p edic ions end o clus e a ound mis ake anno a-
ions, he loca ions a e no exac . This imp ecision would
comp omise ou abili y o ely on such mis ake p edic ions
o imp o e ou ehea sal da a p ocessing pipelines. Fu -
he in es iga ion is needed o de e mine which syn he ic
mis ake ypes can be e ec i ely lea ned, and o ex end
he oolki o c ea e mis akes ha ep esen obse a ions
om he Rach3 MIDI Da ase , as cu en pa ame e s a e
se mo e heu is ically. Fu he mo e, i is possible o c ea e
a small collec ion o human-anno a ed mis ake da a om
he Rach3 MIDI da ase o be used o es ing.
8. CONCLUSION
This pape in oduced he Rach3 MIDI da ase , he la ges
publicly a ailable collec ion o piano ehea sal da a,
eco ded o e ou yea s wi h ou pianis s. I o ms pa o
an ongoing p ojec wi h u u e eleases planned, including
audio and ideo eco dings. Using he da ase , we explo ed
c i ical compu a ional challenges associa ed wi h piano e-
hea sal analysis, applying s a e-o - he-a me hods in h ee
a eas: symbolic ehea sal- o-sco e alignmen , ehea sal
s uc u e analysis, and au oma ic mis ake iden i ica ion.
Ou indings demons a e ha exis ing me hods equi e
subs an ial adap a ion o ehea sal analysis. The Rach3
da ase p o ides bo h he ounda ion o compu a ional e-
hea sal analysis and empi ical e idence o me hodological
gaps ha mus be add essed.
7h ps://gi hub.com/Alia-mo si/piano-synmis
P oceedings o he 26 h ISMIR Con e ence, Daejeon, Ko ea, Sep embe 21-25, 2025
489
9. ACKNOWLEDGMENTS
This wo k has been suppo ed by he Aus ian Science
Fund (FWF), g an ag eemen PAT 8820923 (“Rach3: A
Compu a ional App oach o S udy Piano Rehea sals”), by
he Eu opean Resea ch Council (ERC) unde he EU’s
Ho izon 2020 esea ch & inno a ion p og amme, g an
ag eemen No. 101019375 (“Whi he Music?”), by IA
y Música: Cá ed a en In eligencia A i icial y Música"
(TSI-100929-2023-1), unded by he Sec e a ía de Es ado
de Digi alización e In eligencia A i icial, and he Eu o-
pean Union-Nex Gene a ion EU, unde he p og am Cá e-
d as ENIA 2022 pa a la c eación de cá ed as uni e sidad-
emp esa en IA, and he Resea ch Council o No way
h ough i s Cen es o Excellence scheme, p ojec numbe
262762.
10. REFERENCES
[1] R. Cha in and G. Im eh, “P ac icing Pe ec ion:
Piano Pe o mance as Expe Memo y,” Psychological
Science, ol. 13, no. 4, pp. 342–349, Ap . 2005.
[Online]. A ailable: h ps://www. aylo ancis.com/
books/9781135685461
[2] R. Cha in, T. Lisboa, T. Logan, and K. T. Begosh,
“P epa ing o memo ized cello pe o mance: he
ole o pe o mance cues,” Psychology o Music,
ol. 38, no. 1, pp. 3–30, Jan. 2010. [Online].
A ailable: h p://jou nals.sagepub.com/doi/10.1177/
0305735608100377
[3] S. D. Pe e , C. E. Cancino-Chacón, F. Fos-
ca in, A. P. McLeod, F. Henkel, E. Ka ys inaios,
and G. Widme , “Au oma ic No e-Le el Sco e- o-
Pe o mance Alignmen s in he ASAP Da ase ,”
T ansac ions o he In e na ional Socie y o Mu-
sic In o ma ion Re ie al, ol. 6, no. 1, pp.
27–42, Jun. 2023. [Online]. A ailable: h p:
// ansac ions.ismi .ne /a icles/10.5334/ ismi .149/
[4] W. Goebl, “The Vienna 4x22 Piano Co pus,” 1999.
[Online]. A ailable: h ps://doi.o g/10.21939/4X22
[5] P. Hu and G. Widme , “The ba ik-plays-moza
co pus: Linking pe o mance o sco e o musicological
anno a ions,” in P oceedings o he 24 h In e na ional
Socie y o Music In o ma ion Re ie al Con e ence,
ISMIR 2023, Milan, I aly, No embe 5-9, 2023,
A. Sa i, F. An onacci, M. Sandle , P. Bes agini,
S. Dixon, B. Liang, G. Richa d, and J. Pauwels,
Eds., 2023, pp. 297–303. [Online]. A ailable: h ps:
//doi.o g/10.5281/zenodo.10265283
[6] L. M. G uson, “Rehea sal skill and musical com-
pe ence: does p ac ice make pe ec ?” in
Gene a i e P ocesses in Music: The Psychol-
ogy o Pe o mance, Imp o isa ion, and Com-
posi ion. Ox o d Uni e si y P ess, 01 2001.
[Online]. A ailable: h ps://doi.o g/10.1093/acp o :
oso/9780198508465.003.0005
[7] K. Miklaszewski, “A case s udy o a pianis p epa ing
a musical pe o mance,” Psychology o Music, ol. 17,
no. 2, pp. 95–109, 1989. [Online]. A ailable:
h ps://doi.o g/10.1177/0305735689172001
[8] S. Reid, “P epa ing o pe o mance,” in Musical
Pe o mance, 1s ed., J. Rink, Ed. Camb idge
Uni e si y P ess, Dec. 2002, pp. 102–112. [Online].
A ailable: h ps://www.camb idge.o g/co e/p oduc /
iden i ie /CBO9780511811739A015/ ype/book_pa
[9] P. Miksza, “A Re iew o Resea ch on P ac-
icing: Summa y and Syn hesis o he Ex an
Resea ch wi h Implica ions o a New Theo-
e ical O ien a ion,” Bulle in o he Council o
Resea ch in Music Educa ion, no. 190, pp.
51–92, Oc . 2011. [Online]. A ailable: h ps:
//schola lypublishingcollec i e.o g/bc me/a icle/
doi/10.5406/bulcou esmusedu.190.0051/255279/
A-Re iew-o -Resea ch-on-P ac icing-Summa y-and
[10] E. R. How, L. Tan, and P. Miksza, “A PRISMA e-
iew o esea ch on music p ac ice,” Musicae Scien-
iae, ol. 26, no. 3, pp. 455–697, Sep. 2022.
[11] K. A. E icsson, R. T. K ampe, and C. Tesch-Rome ,
“The Role o Delibe a e P ac ice in he Acquisi ion o
Expe Pe o mance,” Psychological Re iew, ol. 100,
no. 3, pp. 364–403, 1993.
[12] S. Hallam, “P o essional Musicians’ App oaches o
he Lea ning and In e p e a ion o Music,” Psychology
o Music, ol. 23, no. 2, pp. 111–128, Oc . 1995.
[Online]. A ailable: h p://jou nals.sagepub.com/doi/
10.1177/0305735695232001
[13] S. Hallam, I. Papageo gi, M. Va a igou, and
A. C eech, “Rela ionships be ween p ac ice, mo i-
a ion, and examina ion ou comes,” Psychology o
Music, ol. 49, no. 1, pp. 3–20, Jan. 2021. [On-
line]. A ailable: h p://jou nals.sagepub.com/doi/10.
1177/0305735618816168
[14] J. Sokolo skis, D. He emans, and E. Chew, “A
no el in e ace o he g aphical analysis o mu-
sic p ac ice beha io s,” F on ie s in Psychology,
ol. Volume 9 - 2018, 2018. [Online]. A ail-
able: h ps://www. on ie sin.o g/jou nals/psychology/
a icles/10.3389/ psyg.2018.02292
[15] R. Cha in and G. Im eh, “"Pulling Tee h and To u e"
: Musical Memo y and P oblem Sol ing,” Thinking
& Reasoning, ol. 3, no. 4, pp. 315–336, No . 1997.
[Online]. A ailable: h ps://www. and online.com/doi/
ull/10.1080/135467897394310
[16] ——, “A Compa ison o P ac ice and Sel -Repo as
Sou ces o In o ma ion Abou he Goals o Expe
P ac ice,” Psychology o Music, ol. 29, no. 1, pp.
39–69, Ap . 2001. [Online]. A ailable: h p://jou nals.
sagepub.com/doi/10.1177/0305735601291004
P oceedings o he 26 h ISMIR Con e ence, Daejeon, Ko ea, Sep embe 21-25, 2025
490
[17] R. Cha in and T. Lisboa, “P ac icing pe ec ion: How
conce solois s p epa e o pe o mance,” Ad ances
in Cogni i e Psychology, ol. 2, no. 2, pp. 113–130,
Jan. 2006. [Online]. A ailable: h p://www.ac-psych.
o g/en/download-pd / olume/2/issue/2/id/13
[18] M. Wisehea , A. A. D’Souza, and J. Chae, “Lack o
spacing e ec s du ing piano lea ning,” PLOS ONE,
ol. 12, no. 8, pp. 1–13, 08 2017. [Online]. A ailable:
h ps://doi.o g/10.1371/jou nal.pone.0182986
[19] C. E. Ca e and J. A. G ahn, “Op imizing music lea n-
ing: Explo ing how blocked and in e lea ed p ac ice
schedules a ec ad anced pe o mance,” F on ie s in
psychology, ol. 7, p. 1251, 2016.
[20] M. A. Ca a, “The e ec o p ac ice and musical s uc-
u e on pianis s’ eye-hand span and isual moni o ing,”
Jou nal o eye mo emen esea ch, ol. 16, no. 2, pp.
10–16 910, 2023.
[21] R. M. Win e s, S. Gu u ani, and A. Le ch, “Au oma ic
p ac ice logging: In oduc ion, da ase & p elimina y
s udy.” in P oceedings o he 17 h In e na ional
Socie y o Music In o ma ion Re ie al Con e ence.
ISMIR, Aug. 2016, pp. 598–604. [Online]. A ailable:
h ps://doi.o g/10.5281/zenodo.1416224
[22] A. Le ch, C. A hu , A. Pa i, and S. Gu u ani,
“An In e disciplina y Re iew o Music Pe o mance
Analysis,” T ansac ions o he In e na ional Socie y
o Music In o ma ion Re ie al, ol. 3, no. 1,
pp. 221–245, No . 2020. [Online]. A ailable: h p:
// ansac ions.ismi .ne /a icles/10.5334/ ismi .53/
[23] S. Gu u ani, K. A. Pa i, C.-W. Wu, and A. Le ch,
“Analysis o Objec i e Desc ip o s o Music Pe o -
mance Assessmen ,” in P oceedings o he In e na-
ional Con e ence on Music Pe cep ion and Cogni ion
(ICMPC15/ESCOM10), G az, Aus ia, 2018.
[24] C. E. Cancino-Chacón and I. Pilko , “The Rach3
Da ase : Towa ds Da a-D i en Analysis o Piano
Pe o mance Rehea sal,” in Mul iMedia Modeling,
S. Rudinac, A. Hanjalic, C. Liem, M. Wo ing,
B. Jónsson, B. Liu, and Y. Yamaka a, Eds.
Cham: Sp inge Na u e Swi ze land, 2024, pp.
28–41. [Online]. A ailable: h ps://doi.o g/10.1007/
978-3-031-56435-2_3
[25] A. A z , S. Böck, and G. Widme , “Fas iden i ica ion
o piece and sco e posi ion ia symbolic inge -
p in ing,” in P oceedings o he 13 h In e na ional
Socie y o Music In o ma ion Re ie al Con e ence,
ISMIR 2012, Mos ei o S.Ben o Da Vi ó ia, Po o,
Po ugal, Oc obe 8-12, 2012, F. Gouyon, P. He -
e a, L. G. Ma ins, and M. Mülle , Eds. FEUP
Edições, 2012, pp. 433–438. [Online]. A ailable: h p:
//ismi 2012.ismi .ne /e en /pape s/433-ismi -2012.pd
[26] S. D. Pe e , “Online Symbolic Music Alignmen Wi h
O line Rein o cemen Lea ning,” in P oceedings o
he 24 h In e na ional Socie y o Music In o ma ion
Re ie al Con e ence, ISMIR 2023, Milan, I aly,
No embe 5-9, 2023, A. Sa i, F. An onacci, M. San-
dle , P. Bes agini, S. Dixon, B. Liang, G. Richa d,
and J. Pauwels, Eds., 2023, pp. 634–641. [Online].
A ailable: h ps://doi.o g/10.5281/zenodo.10265367
[27] E. Nakamu a, K. Yoshii, and H. Ka ayose, “Pe o -
mance E o De ec ion and Pos -P ocessing o Fas
and Accu a e Symbolic Music Alignmen ,” in P o-
ceedings o he 18 h In e na ional Socie y o Mu-
sic In o ma ion Re ie al Con e ence (ISMIR 2017),
Suzhou, China, 2017.
[28] S. Pe e and G. Widme , “The GlueNo e: Lea ned Rep-
esen a ions o Robus and Flexible No e Alignmen ,”
in P oceedings o he 25 h In e na ional Socie y o
Music In o ma ion Re ie al Con e ence, ISMIR 2024,
San F ancisco, Cali o nia, USA and Online, No embe
10-14, 2024, B. Kaneshi o, G. J. Myso e, O. Nie o,
C. Donahue, C. A. Huang, J. H. Lee, B. McFee, and
M. C. McCallum, Eds., 2024, pp. 603–610. [Online].
A ailable: h ps://doi.o g/10.5281/zenodo.14877409
[29] D. Me edi h, K. Lems öm, and G. Wiggins, “Algo-
i hms o disco e ing epea ed pa e ns in mul idimen-
sional ep esen a ions o polyphonic music,” Jou nal o
New Music Resea ch, ol. 31, 04 2003.
[30] T. Collins, “Disco e y o epea ed hemes and
sec ions,” Re ie ed 4 h May, h p://www. musici .
o g/mi ex/wiki/2013: Disco e y o Repea ed Themes &
Sec ions, 2013.
[31] A. Mo si, K. Ta sumi, A. Maezawa, T. Fujishima, and
X. Se a, “Sounds ou o Pläce? Sco e-Independen
De ec ion o Conspicuous Mis akes in Piano Pe o -
mances,” in P oceeding o he 24 h In e na ional Soci-
e y on Music In o ma ion Re ie al (ISMIR), No embe
5-9, 2023.
[32] A. Mo si, H. Zhang, A. Maezawa, S. Dixon, and
X. Se a, “Simula ing piano pe o mance mis akes o
music lea ning,” P oceedings o he 21s Sound and
Music Compu ing Con e ence (SMC 2024), July 2024.
[33] M. Mülle , V. Konz, W. Bogle , and V. A i i-Mülle ,
“Saa land music da a (SMD),” 2011.
[34] F. Fosca in, A. McLeod, P. Rigaux, F. Jacquema d,
and M. Sakai, “ASAP: a da ase o aligned sco es and
pe o mances o piano ansc ip ion.” in P oceedings
o he 21 h In e na ional Socie y o Music In o ma ion
Re ie al Con e ence, ISMIR 2020, Mon eal, Canada,
Oc obe 11-16, 2020, 2020, pp. 534–541. [Online].
A ailable: h p://a chi es.ismi .ne /ismi 2020/pape /
000127.pd
P oceedings o he 26 h ISMIR Con e ence, Daejeon, Ko ea, Sep embe 21-25, 2025
491
