THE FLORENCE PRICE ART SONG DATASET AND PIANO
ACCOMPANIMENT GENERATOR
Tao-Tao He
Vande bil Uni e si y
[email p o ec ed]
Ma in E. Maland o
Sam Hous on S a e Uni e si y
[email p o ec ed]
Douglas Shadle
Vande bil Uni e si y
[email p o ec ed]
ABSTRACT
Flo ence B. P ice was a compose in he ea ly 20 h cen u y
whose music e lec s he upb inging in he Ame ican
Sou h, he A ican he i age, and he Wes e n classical
aining. She is no ed as he i s A ican-Ame ican woman
o ha e a symphony pe o med by a majo o ches a.
He music has ecen ly ecei ed enewed a en ion om
bo h he public and he esea ch communi y, decades
a e he dea h. In addi ion o o he gen es, P ice was
a p oli ic compose o solo oice and piano. Music
his o ians ha e documen ed he exis ence o 134 a songs
and piano/ oice a angemen s o spi i uals and olk songs
w i en by P ice. We elease a digi al ca alog o 112 o
hese wo ks in MuseSco e, MusicXML, MIDI, and PDF
o ma . We also use his da ase o ine- une a symbolic
music gene a ion model o gene a e accompanimen s o
melodies, and we conduc a blind lis ening expe imen ha
shows ha accompanimen s gene a ed by ou model a e
pe cei ed as being e lec i e o Flo ence P ice’s s yle mo e
equen ly han accompanimen s gene a ed by a baseline
model. We elease ou model as he Flo ence P ice Piano
Accompanimen Gene a o alongside ou da ase . 1
1. INTRODUCTION
Flo ence B. P ice (Ap il 9, 1888–June 3, 1953) was a
p oli ic compose o classical music du ing he i s hal
o he wen ie h cen u y. Bo n and aised in Li le Rock,
A kansas, she s udied piano, o gan, and composi ion a
he New England Conse a o y, whe e she g adua ed wi h
hono s in 1906. Fo he nex wo decades, she p ima ily
li ed in A kansas, whe e she wo ked p incipally as a
music educa o in he egion’s seg ega ed academics. She
eloca ed o Chicago in 1927, whe e she encoun e ed a
suppo i e communi y o A ican Ame ican musicians,
a is s, dance s, and w i e s ha sus ained he c ea i e
ene gies. He ca ee as a compose eached i s i s high
poin wi h he p emie e o he awa d-winning Symphony
1h ps://gi hub.com/TT515/Flo ence_P ice_A _Song_Da ase
© T. He, M. E. Maland o, and D. Shadle. Licensed unde
a C ea i e Commons A ibu ion 4.0 In e na ional License (CC BY 4.0).
A ibu ion: T. He, M. E. Maland o, and D. Shadle, “The Flo ence P ice
A Song Da ase and Piano Accompanimen Gene a o ”, 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.
in E Mino by he Chicago Symphony O ches a in
1933— he i s ime ha a majo Ame ican o ches a
had pe o med music by an A ican Ame ican woman.
Du ing he inal wo decades o he li e, o he majo
pe o me s and ensembles, such as he con al o Ma ian
Ande son and he US Ma ine Band, championed he
music while he epu a ion g ew in e na ionally h ough
ex ensi e publica ion o he sho e piano wo ks. [1,2].
Despi e he a en ion and acclaim P ice had ecei ed
du ing he li e ime, much o he composi ional ou pu
ell in o obscu i y a e he dea h because many o he
mos impo an manusc ip s seemingly disappea ed. In
2009, howe e , hund eds o documen s om P ice’s li e
we e disco e ed in he old summe co age in Kankakee
Coun y, Illinois, including le e s, dia ies, pho og aphs,
and manusc ip sco es [3]. This disco e y spa ked in e es
om esea che s, who now had new in o ma ion o piece
oge he P ice’s li e and music, and pe o ming musicians,
who now had g ea e access o he ex ensi e ca alog. E en
so, access o he sco es has emained limi ed since mos
o he wo ks emained unpublished and mos manusc ip
sco es we e unde US copy igh p o ec ion un il 2024.
P ice exhibi ed a special in e es in he a song gen e.
A songs a e ocal composi ions ha se poe y o music,
ypically w i en o solo oice and piano, and composed
wi hin he Wes e n classical adi ion [4]. P ice’s co pus o
o e one hund ed a songs inco po a es ex s wi h di e se
subjec ma e s wi h he s ylis ically dis inc i e music.
P ice also a anged a a ie y o A ican Ame ican olk
songs o solo oice and piano, including many spi i uals–a
gen e o A ican Ame ican eligious olk song o igina ing
in he US Sou h unde ensla emen .
In his pape , we c ea e and elease a da ase o 112
o P ice’s a songs and olksong a angemen s. Making
his co pus a ailable will enhance public access o P ice’s
musical exp ession, musicological esea ch on P ice’s li e
and composi ional s yle, and echnological expe imen s
ha engage wi h he da ase in gene a i e ways.
We engage in one such expe imen by ine- uning
a gene a i e symbolic music model on ou da ase o
c ea e a model ha gene a es accompanimen s ha a e
e lec i e o P ice’s s yle. The model accep s melodies
and ou pu s accompanimen s o each inpu melody. The
ins umen a ion o a songs make hem na u ally i o
s udying ela ions be ween melody and accompanimen ,
and P ice’s ca alog o such songs o e s he oppo uni y o
s udy how a speci ic compose app oached hese ela ions.
771
Ou model has applica ions in educa ion, pe o mance,
and esea ch. The model’s abili y o gene a e
accompanimen s o e s esea che s a no el lens o
analyzing P ice’s s ylis ic endencies. The model could
gene a e piano pa s o P ice’s incomple e wo ks, simila
o how an AI model helped comple e an ex ended
e sion o Schube ’s "Un inished Symphony" [5]. I can
also se e as a ool o inspi a ion o compose s and
composi ion s uden s who wan o inco po a e elemen s
o P ice’s s yle. Fo AI music esea che s, i se es as a
case s udy o model pe sonaliza ion—namely, how well
a model can cap u e a compose ’s s yle based on limi ed
da a.
Ou main con ibu ions a e: (1) We elease a no el
da ase o 112 o P ice’s a songs in MuseSco e,
MusicXML, MIDI, and PDF o ma . This is he mos
comple e digi al collec ion o Flo ence P ice’s ocal music
wo ks o da e. (2) We ain and elease he Flo ence P ice
Piano Accompanimen Gene a o , which gene a es piano
accompanimen s e lec i e o he s yle o Flo ence P ice.
(3) We conduc a blind lis ening expe imen ha shows ha
accompanimen s gene a ed by ou model a e pe cei ed as
being e lec i e o Flo ence P ice’s s yle mo e equen ly
han accompanimen s gene a ed by a baseline model.
2. RELATED WORK
Shee music has been digi ized in o compu e - eadable
o ma s in nume ous ways, including ABC no a ion,
music21 [6], MIDI, and MusicXML. Digi ized classical
music da ase s include he OpenSco e Liede Co pus [7],
which includes 1356 wo ks o ocal music (a songs) by
classical music compose s (p ima ily in he Roman ic E a),
and he OpenSco e S ing Qua e Co pus [8]. Ou da ase
adds o collec ion o compose -speci ic ca alogs, such as
he JSB Cho ales da ase [9], which encodes 382 o J.S.
Bach’s ou -pa cho ales in a ious o ma s and is widely
used in machine lea ning con ex s—see, e.g., [10,11].
The in-p og ess Flo ence B. P ice Wo ks Ca alog [12],
launched in 2023, keeps ack o in o ma ion abou all
documen ed composi ions by P ice, including whe e each
composi ion’s exis ing manusc ip s can be ound. The
mos comple e collec ion o Flo ence P ice’s ocal music
o da e is 44 A Songs and Spi i uals [13], published in
2015. Mo e o he ecen ly disco e ed songs ha e been
eco ded and eleased, including Ka en Slack’s G ammy®-
winning album Beyond he Yea s: Unpublished Songs o
Flo ence P ice [14], con aining 19 unpublished songs by
P ice and eleased in 2024. Howe e , un il now, mos o
he shee music o hese songs has emained unpublished.
P io o his wo k, he e we e no digi al eposi o ies o
P ice’s wo ks in symbolic o ma — ocal o o he wise.
T ans o me s ha e been widely used o symbolic
music gene a ion since Google’s Music T ans o me
elease in 2018 [15]. The ask o symbolic music in illing,
and speci ically accompanimen gene a ion, has been
explo ed in [15] as well as in Compose ’s Assis an [16,
17], Accomon age [18], and o he s [19–22]. In Sh edGP
[23] and P ogGP [24], i is shown ha ans o me models
Figu e 1. P ice’s manusc ip o I Remembe , p. 1. [25]
can be ine- uned on small da ase s o ain he models
o w i e in speci ic musical s yles. Howe e , ine- uning
o s yle-condi ioned accompanimen gene a ion is a mo e
sub le ques ion, and o ou knowledge is unde -explo ed,
pa icula ly in he con ex o a single compose ’s wo ks.
3. THE DATASET
3.1 Basic in o ma ion
134 di e en Flo ence P ice a songs a e documen ed o
exis be ween he Flo ence B. P ice Wo ks Ca alog [12],
he Ma ian Ande son Collec ion a he Penn Lib a ies, and
he Mullins Lib a y o he Uni e si y o A kansas. The
Mullins Lib a y a Uni e si y o A kansas holds he la ges
collec ion o P ice’s music manusc ip s, which can only be
accessed in pe son upon eques o he lib a y o h ough
pho o ep oduc ion. The Ma ian Ande son Collec ion a
he Penn Lib a ies also con ains a signi ican numbe o
P ice’s manusc ip s; o some o hese, he lib a y p o ides
pho ocopies ha a e a ailable online.
O he songs, 116 a e se o ly ics om a poe o by a
named ly icis , including se e al songs which P ice he sel
w o e he ly ics. 18 a e a angemen s o A ican-Ame ican
olk songs, including spi i uals. 132 o he songs a e sco ed
o a solo ocalis and a pianis . Fi e o he songs a e
incomple e.
Rega ding s yle, he songs a e o en highly ch oma ic,
pa icula ly in he piano accompanimen . P ice equen ly
used ou -o -key ha monies, diminished and augmen ed
cho ds, and pa allel key shi s. The whole- one scale
was a signa u e ea u e in he songs. Many o he piano
accompanimen s ea u e long appoggia u as. P ice used
P oceedings o he 26 h ISMIR Con e ence, Daejeon, Ko ea, Sep embe 21-25, 2025
772
he Juba dance hy hm in many o he composi ions,
including many a songs. Many o he a songs
include key changes and me e changes, as well as
su p ising hy hmic changes. Fo example, P ice’s song
"I Remembe " (see Figu e 1) inco po a ed augmen ed
i h ha mony, ou -o -key ch oma ic no es, syncopa ed
hy hm, ac i i y in mul iple egis e s, no es o a a ie y o
du a ions, and a me e change. The songs a e gene ally
a ound 40 ba s in leng h, wi h a ange be ween 8 and 80
ba s. Ou p ojec con ibu es o he g owing analy ical
li e a u e o P ice’s a songs [26,27].
The i s au ho c ea ed he da ase by manually
ansc ibing manusc ip s o P ice’s songs o MuseSco e,
and con e ing each MuseSco e ile o MusicXML, MIDI,
and PDF o ma s. Pho ocopies o he manusc ip s o 129
songs we e ob ained om he Uni e si y o A kansas and
he Penn Lib a ies websi e. A i s , we a emp ed o use
Op ical Music Recogni ion (OMR) o help digi ize he
manusc ip s. Howe e , OMR s uggled o p ocess he
manusc ip s accu a ely, so hey we e manually ansc ibed
ins ead. Each ansc ip ion was e iewed wice upon
comple ion.
3.2 The Flo ence P ice A Song Da ase
Mos o P ice’s a songs en e ed he public domain in he
Uni ed S a es o Ame ica on Janua y 1, 2024, 70 yea s
a e he compose ’s dea h in 1953. Howe e , 17 o he
129 songs we ob ained we e published be ween 1930 and
1973, and may emain unde copy igh in he USA. Ou o
legal and e hical conce ns, we exclude hese 17 songs om
ou published da ase and he aining se o ou model.
We elease he emaining 112 songs as he Flo ence P ice
A Song Da ase . 2
Ou a song da ase includes a olde o each song.
Each olde includes he MuseSco e ile, MusicXML,
MIDI, and PDF e sions o he song, as well as iles o
ly ics, me ada a, and onse s. Me ada a includes he empo
o he song, he ly icis , he i s au ho ’s judgmen o
whe he he song is happy o sad, ano he adjec i e he
i s au ho pe cei es o desc ibe he song’s mood, and he
song’s musical s yle. We p o ide wo “onse s” iles o
each song, which deno e he ba -wise sec ion bounda ies o
he song as pe cei ed by he i s and second au ho s. We
also p o ide a sc ip o gene a ing audio o e e y song in
he da ase .
Legibili y o he manusc ip s p esen ed challenges o
c ea ing he ca alog. P ice’s in en ions we e no always
clea , and in hose ins ances, ou ansc ip ion is based on
he i s au ho ’s judgemen . The au ho also co ec ed
pe cei ed e o s in he manusc ip s, mos commonly
ambigui y in acciden als. We ecognize ha despi e e o s
o p esen P ice’s composi ions as accu a ely as possible,
he e may s ill be e o s in ou ca alog. We welcome
eedback and will ac i ely manage he ca alog o p esen
P ice’s wo ks accu a ely.
2h ps://gi hub.com/TT515/Flo ence_P ice_A _Song_Da ase
4. THE FLORENCE PRICE PIANO
ACCOMPANIMENT GENERATOR
In his sec ion, we desc ibe how we c ea ed a model,
he ea e e e ed o as he FP model, o w i ing piano
accompanimen s ha e lec he s yle o Flo ence P ice.
Gi en he small size o ou da ase , i would be in easible
o ain such a model om sc a ch. The e o e, we s a ed
wi h he Compose ’s Assis an 2.1 model eleased in [17]
(which we e e o as he s a ing model) and ine- uned
i on ou da ase . The s a ing model uses a T5 [28, 29]
a chi ec u e and has 16 encode laye s, 16 decode laye s,
12 a en ion heads pe laye , and a model dimension o 576,
wi h a o al pa ame e coun o app oxima ely 192M.
The model accep s as inpu any slice o measu es
om a MIDI ile and, o each ack wi hin he slice,
a se o ba -le el masks. The model hen gene a es,
using he s uc u al in o ma ion and unmasked no es in
he inpu , no es o eplace each mask. Syn ac ically,
his is ca ied ou wi h he T5 denoising objec i e:
Simpli ying he syn ax o he model sligh ly o
explana o y pu poses, an example o a h ee-measu e inpu
o he model con aining an oboe ack and a piano ack
is [measu e] [piano] [N0,0, . . . , N0,n0] [oboe] [MASK0]
[measu e] [piano] [MASK1] [oboe] [MASK2] [measu e]
[piano] [N1,0, . . . , N1,n1] [oboe] [N2,0, . . . , N2,n2]. In
his example, [Ni,0, . . . , Ni,ni]is a sequence o okens
desc ibing he no es played by he mos ecen p e iously-
decla ed ins umen in he mos ecen p e iously-decla ed
measu e—see [16] o de ails. The oboe’s pa in he i s
measu e is masked, and bo h pa s in he second measu e
a e masked. The model’s ou pu would be [MASK0]
[M0,0,...M0,m0] [MASK1] [M1,0,...M1,m1] [MASK2]
[M2,0,...M2,m2] [eos], whe e [Mi,0, . . . , Mi,mi]is a
sequence o okens desc ibing he no es ha he model
has gene a ed o eplace MASKiin he p omp . T aining
examples a e gene a ed by aking andom measu e slices
om he aining da a, masking measu es wi hin he acks
in he slice, and aining he model o ill in he masked
g ound- u h da a. Fo alida ion and es ing, we a e only
in e es ed in accompanimen gene a ion, whe e he model
is p o ided wi h a melody and asked o w i e a piano
accompanimen . Ou accompanimen gene a ion aining
objec i e is achie ed by masking e e y measu e in he
piano ack and lea ing he melody ack (oboe in he
example) unmasked.
To c ea e a baseline o compa ison, we ine- uned he
s a ing model on he OpenSco e Liede Co pus [7], a
collec ion o 1356 a songs by 19 h cen u y compose s. To
ain his baseline model, we used 100 songs o alida ion
and he es o aining. We used an e ec i e ba ch
size o 128 o aining and sa ed a checkpoin e e y 10
epochs. Each epoch co esponded o abou 80 op imiza ion
s eps. We ound he s a ing model unsui able as a
baseline, as i is a gene al model a he han an a song-
speci ic model, and s anda d me ics such as no e densi y,
pi ch his og am en opy (PHE), and pi ch class his og am
en opy (PCHE) [30] o gene a ed accompanimen s
di e ed widely om he ac ual accompanimen s in he
P oceedings o he 26 h ISMIR Con e ence, Daejeon, Ko ea, Sep embe 21-25, 2025
773
Figu e 2. KL-di e gence be ween me ics o gene a ed
accompanimen s and g ound u h accompanimen s in he
Liede alida ion da a, a e ine- uning he s a ing model
on he Liede Co pus. Subjec i ely-bes model highligh ed
in g een.
alida ion se —see Figu e 2. This igu e was c ea ed
by gene a ing accompanimen s o he melodies in 16-
ba snippe s selec ed a andom om he alida ion se
and compu ing he Kullback-Leible (KL) di e gence [31]
be ween he me ics o gene a ed accompanimen s a e
nepochs o ine- uning (n∈ {0,10,20,...,300}) and
he me ics o he g ound u h accompanimen s. KL-
di e gence p o ides a measu e o he di e ence be ween
he obse ed and g ound u h dis ibu ions. Figu e 2 also
shows how quickly he ine- uning p ocess ains a model
ha gene a es accompanimen s which be e app oxima e
he dis ibu ions o hese me ics in he g ound u h.
Despi e he popula i y o ine- uning gene a i e c ea i e
models on small da ase s o s yle pu poses (e.g., in a
and music), he e does no appea o be much guidance
in he li e a u e abou exac ly when o s op he ine-
uning p ocess. P e ious wo ks ha ine- uned symbolic
music models include [23, 24, 32, 33]. The mos common
guidance gi en is o sa e a aining checkpoin e e y so
Figu e 3. Pi ch his og am en opy o gene a ed
accompanimen s o melodies in he Liede aining da a,
a e ine- uning he s a ing model on he Liede Co pus.
Subjec i ely-bes model highligh ed in g een.
o en, examine some model ou pu s om he checkpoin ,
and decide whe he o keep aining. E en ually, aining is
s opped when he ou pu s no iceably deg ade as he model
o e i s. Ul ima ely, his was ou me hodology o aining
bo h he baseline model and he FP model. In one no able
excep ion, he au ho s o [24] ound success in compa ing
he PHE o gene a ed samples o he PHE o he aining
da a o decide when o s op ine- uning hei model—
hey obse ed di e gence be ween hese quan i ies as hei
model o e i . We obse ed he opposi e signal while
aining ou baseline model, bo h when compa ing o he
alida ion da a (cen e plo o Figu e 2) and o he aining
da a i sel (Figu e 3), so his did no p o ide a clea signal
as o when o s op aining. Ra he , he con e gence
in Figu e 3 is consis en wi h he model memo izing i s
aining da a. Some memo iza ion is accep able, as long
as he quali y and di e si y o gene a ed accompanimen s
o melodies ou side he aining se do no su e .
As no ed in [34], me ics commonly used o
he e alua ion o gene a ed music—bo h objec i e and
subjec i e— end o ocus on quali y a he han di e si y o
ou pu s. We wan ou ained models o be able o c ea e a
wide a ie y o high-quali y accompanimen s o any gi en
inpu melody. Fo in e ence, we use nucleus sampling [35]
wi h p= 0.95 and a empe a u e o 1.0, and we ound ha
inc easing hy hmic empe a u e o 1.5 du ing in e ence
helps main ain a di e si y o ou pu s and allows he models
o ain o mo e epochs, leading o highe -quali y ou pu s.
(The model gene a es accompanimen s one oken a a ime,
whe e a oken con ols he pi ch, posi ion, o du a ion o
a gene a ed no e. By hy hmic empe a u e, we e e o
he empe a u e pa ame e when he model is choosing he
nex posi ion o du a ion oken, bu no when i is choosing
he nex pi ch oken.)
To ain he FP model, we ine- uned he s a ing
model on ou Flo ence P ice A Song Da ase and sa ed
a checkpoin e e y 10 epochs. Due o he small size
o ou da ase , each epoch co esponded o abou 2.7
aining s eps. By lis ening o accompanimen s gene a ed
o melodies ou side he aining se s, we de e mined ha
P oceedings o he 26 h ISMIR Con e ence, Daejeon, Ko ea, Sep embe 21-25, 2025
774
aining he baseline model o 130 epochs and aining he
FP model o 100 epochs was op imal.
While aining, we examined alida ion loss, no e
densi y, PHE, PCHE, and F1, bo h on he alida ion da a
(in he case o he baseline model) and on he aining
da a, and ound no clea signal in any o hem poin ing us
owa ds he model checkpoin s we pe cei ed as op imal.
We he e o e call o mo e esea ch in o s opping c i e ia
o s yle-based ine- uning on small da ase s. The only
clea signals we ound ha poin ed us in he igh di ec ion
we e: (1) a no iceable d op in he di e si y o gene a ed
accompanimen s o gi en melodies a e oo much aining,
and (2) ca as ophic o ge ing [36]: The s a ing model
comes equipped wi h con ol okens ha can be supplied
in p omp s o a ec he musical p ope ies (such as
no e densi y) o he gene a ed ou pu s, and e en ually
hese con ol okens become inc easingly ine ec i e (i.e.,
igno ed by he model) wi h addi ional aining. Bo h issues
began o occu a 150 epochs wi h ou FP model and a 170
epochs wi h he baseline model, and in bo h cases we e
exace ba ed by mo e aining, gi ing us an uppe bound
o ou lis ening-based sea ch.
We elease he FP model and code o in e ac ing
wi h i as he Flo ence P ice Piano Accompanimen
Gene a o . We p o ide an in e ace whe e a use can
upload any melody in MIDI o ma and download piano
accompanimen s gene a ed by he model.
5. EVALUATION
Due o he no el y o he da ase p esen ed in his pape ,
i would be di icul o ind ex e nal e alua o s who
a e amilia wi h he s yle o he music in he da ase .
The e o e, he i s and hi d au ho s, who ha e lis ened o
all o he music in he da ase , pa icipa ed as e alua o s
in a blind lis ening expe imen o assess he ex en o
which he FP model (de eloped in Sec ion 4) is pe cei ed
as w i ing accompanimen s ha a e e lec i e o Flo ence
P ice’s s yle. The second au ho , who pe o med model
checkpoin selec ion in Sec ion 4, used he FP model and
he baseline model o gene a e accompanimen s o speci ic
melodies o he e alua o s o lis en o. No con ol okens
om he s a ing model (e.g., no e densi y, pi ch ange)
we e used, in o de o gi e he models he mos lexibili y
in hei gene a ion.
We selec ed 10 melodies om Flo ence P ice’s a
songs and 10 popula melodies, and o bo h models,
gene a ed 7 accompanimen s o each melody. This
c ea ed 140 sample accompanimen s om each model. To
gene a e he accompanimen s o P ice’s melodies, we did
no use he ine- uned model om Sec ion 4. Ra he , we
employed a lea e-one-ou aining app oach, aining he
s a ing model o 100 epochs on ou da ase wi hou he
song con aining he melody, and using he esul ing model
o gene a e he accompanimen s o ha melody.
The e alua o s hen blindly and sepa a ely lis ened
o pai s o accompanimen s o each melody (one
accompanimen gene a ed by he FP model, he o he
by he baseline model) and selec ed he accompanimen
Melody ype nP ice model wins p- alue
P ice 137 82 0.026
Popula 137 99 <0.001
Table 1. Resul s om ou blind lis ening expe imen .
The lis ene s conside ed accompanimen s gene a ed by
he Flo ence P ice model o be e lec i e o P ice’s s yle
mo e o en han accompanimen s gene a ed by he baseline
model. Six ou o 280 esponses we e missing. p- alues
we e ob ained om binomial es s.
Ou Baseline
Model Model
Ha d e o s 67 47
So e o s 68 57
% accompanimen s 45.71% 57.14%
wi h no e o s
Table 2. Pe cei ed e o s in gene a ed accompanimen s
(140 om each model). Each accompanimen was 20-
60 seconds, a e aging 30 seconds, and was independen ly
examined by wo lis ene s. Bo h lis ene s ag eed ha 45%
o he clips gene a ed by he FP model had no e o s.
hey el be e e lec ed he s yle o Flo ence P ice. The
e alua o s lis ened o he same 280 samples, bu pai ed
in di e en ways. These samples we e nei he edi ed no
che y picked. These examples and many mo e (1400 in
o al) a e a ailable o he eade o lis en o. 3
We gi e he esul s o his expe imen in Table 1. The
e alua o s conside ed accompanimen s gene a ed by he
Flo ence P ice model o be e lec i e o P ice’s s yle
mo e o en han accompanimen s gene a ed by he baseline
model o bo h ypes o melodies.
As a seconda y objec i e, o help assess he quali y
o he FP model and i s po en ial o use in co-
c ea i e composi ion, he e alua o s ook no es o how
many “ha d” and “so ” e o s hey hea d in each
gene a ed accompanimen . These coun s a e p esen ed
in Table 2. Fo pu poses o his pape , a ha d e o
is a signi ican occu ence o ha monic dissonance o
melodic incohe ence, while a so e o is an occu ence
o cohe en bu coun e -in ui i e melodic o ha monic
sequences, such as an unp omp ed use o in e sions o
acciden als. Some imes he ou pu con ains no e o s as
de ined, o only so e o s ha could be ixed by a skilled
compose .
The baseline model gene a es e o - ee
accompanimen s mo e o en han he FP model, which
we ace back o he small size o ou ine- uning se .
Howe e , o co-c ea i e applica ions, he goal is no
necessa ily pe ec ion, bu a he p o iding s ylis ically
app op ia e and usable ma e ial. While an e o - ee
accompanimen is no necessa ily a good accompanimen ,
he as majo i y o ou gene a ed accompanimen s,
including hose wi h e o s, ea u e musically in e es ing
hy hmic ex u es, ha monic oicing, mo emen be ween
3h ps://gi hub.com/m-maland o/Flo ence-P ice-lis ening-examples
P oceedings o he 26 h ISMIR Con e ence, Daejeon, Ko ea, Sep embe 21-25, 2025
775
7
V.
Voice
P iceModelPiano
P iceModel
BaselineModelPiano
Baseline
oend he
sun he
he
he he wings o wings o se
hang
ime sun ime changeo change
o
o
o
gi gi song,song, Fo Fo
edge mo
longlongdayday changeandand andand
ow.
amam gladgladII
and he mo wo ld on ow.wo ld hang
Fo
Which
owFo he ow
amI hings hings
so so
on
3
3
3
3
3
3
33
=114
Figu e 4. Fi s 9 measu es o gene a ed accompanimen s o Rhapsody.
egis e s, and an app op ia e amoun o embellishmen s.
These a e all s andou ea u es o he a song gen e.
Ou pe cei ed e o - ee accompanimen gene a ion
a e o 45% he e o e sugges s ha he FP model can
meaning ully suppo compose s seeking inspi a ion in
P ice’s s yle.
Finally, as a case s udy, we compa e wo
accompanimen s gene a ed o he melody o Rhapsody,
composed by Flo ence P ice. One was gene a ed by he
baseline model, and he o he was gene a ed by ou model
a e ine- uning on he ca alog o P ice’s songs excluding
Rhapsody. The i s nine (ou o 22) ba s o he wo
accompanimen s a e gi en in Figu e 4. The wo 22-ba
samples a e a ailable o he eade o lis en o. 4
We analyze he hy hm, ange, and ha mony o he
wo accompanimen s. The baseline model’s hy hm is
based on eigh h no es, and de ia es o a mo e ac i e
hy hm ea u ing 32nd no es a ba eigh . The FP model’s
hy hm is based on iple s and de ia es o a qua e -
no e hy hm in ba six, be o e e u ning o iple s and
eaching a climax a ba nine wi h ascending iple cho ds.
Flo ence P ice o en uses as ascending cho ds in he
accompanimen o each climaxes in he songs, so his is
an example o he FP model inco po a ing P ice’s musical
language in o i s gene a ions. The FP model sample is
ac i e in a wide ange, simila o many o P ice’s a
songs accompanimen s. Bo h accompanimen s ea u e
ha monies cohe en wi h he con ex o he melody and he
accompanimen ’s own p og ession, al hough hey di e in
ba nine, whe e he baseline model’s ha mony is F# majo
4h ps://gi hub.com/m-maland o/Flo ence-P ice-lis ening-
examples/ eleases/download/ 1.0.0/Rhapsody.zip
and he FP model’s ha mony is B♭majo ( ea ing he
A# in he melody as a B♭). B♭majo is mo e su p ising
and exp essi e in he con ex o he home key D majo ,
and al hough P ice’s o iginal composi ion had F# majo
ha mony in ba nine, he B♭majo ha mony is encou aging
because i e lec s P ice’s endency o use su p ising
and exp essi e ha mony in he accompanimen s. By
compa ing he ha mony, hy hm, and ange o he sample
accompanimen s, we showcase he FP model’s capaci y o
gene a e accompanimen s ha align wi h P ice’s musical
ocabula y and exp essi e endencies.
6. CONCLUSION
We ha e c ea ed and eleased a digi al ca alog o Flo ence
P ice’s wo ks o solo oice and piano in MuseSco e,
MusicXML, MIDI, and PDF o ma . Using he ca alog,
we ha e ine- uned and eleased a symbolic music model
o gene a e piano accompanimen s e lec i e o P ice’s
composi ional s yle. We ha e conduc ed a lis ening
expe imen and concluded ha he accompanimen s
gene a ed by he ine- uned model success ully cap u ed
elemen s o P ice’s s yle. We hope ha ou digi al ca alog
can se e as a pilla o a is s, lis ene s, and esea che s
in e es ed in lea ning mo e abou P ice’s songs, and we
hope ha ou model can se e as a ool bo h o compose s
and hose in e es ed in analyzing P ice’s music.
7. ETHICS STATEMENT
All o he music we ha e eleased in he Flo ence P ice A
Song Da ase and used o ain he Flo ence P ice Piano
Accompanimen Gene a o en e ed he public domain in
P oceedings o he 26 h ISMIR Con e ence, Daejeon, Ko ea, Sep embe 21-25, 2025
776
he Uni ed S a es o Ame ica on Jan. 1, 2024. We ob ained
access o P ice’s manusc ip s wi h pe mission om he
Uni e si y o A kansas Da id W. Mullins Lib a y and
he Uni e si y o Pennsyl ania Kislak Cen e o Special
Collec ions, Ra e Books, and Manusc ip s.
8. REFERENCES
[1] R. L. B own, The Hea o a Woman: The Li e and
Music o Flo ence B. P ice. U bana, IL, USA:
Uni e si y o Illinois P ess, 2020.
[2] S. Ege, “Flo ence P ice and he Poli ics o He
Exis ence,” Ame ican Music, ol. 39, no. 1, pp. 1–23,
2021.
[3] K. T. S ewa d, “A e Los Sco es A e Found
in Abandoned House, Musicians Gi e Li e o
Flo ence P ice’s Music,” 2018, kUAR, Uni e si y
o A kansas a Li le Rock Public Radio. Accessed
Ma ch 26, 2025. [Online]. A ailable: h ps:
//www.ual public adio.o g/local- egional-news/2018-
05-04/a e -los -sco es-a e- ound-in-abandoned-
house-musicians-gi e-li e- o- lo ence-p ices-music
[4] A. Ha ison, “Music app ecia ion,” 2021. [Online].
A ailable: h ps://lib a y.achie ing hed eam.o g/
alamomusicapp ecia ion/chap e /a -song/
[5] E. Da is, “Schube ’s ’Un inished’ Symphony
comple ed by a i icial in elligence,”
Classic FM, 2019. [Online]. A ailable:
h ps://www.classic m.com/compose s/schube /
un inished-symphony-comple ed-by-ai/
[6] M. Cu hbe , C. A iza, B. Hogue, and J. W.
Obe hol ze , “music21: A Toolki Fo Compu e -
Aided Musicology.” [Online]. A ailable: h p://
web.mi .edu/music21/
[7] M. R. H. Go ham and P. Jonas, “The OpenSco e Liede
Co pus,” in Music Encoding Con e ence P oceedings
2021, S. Münnich and D. Rizo, Eds. Humani ies
Commons, 2022, pp. 131–136.
[8] M. R. H. Go ham, M. Redbond, B. Bowe , and
P. Jonas, “The “OpenSco e S ing Qua e ” Co pus,”
in P oceedings o he 10 h In e na ional Con e ence
on Digi al Lib a ies o Musicology. Milan I aly:
ACM, No . 2023, pp. 49–57. [Online]. A ailable:
h ps://dl.acm.o g/doi/10.1145/3625135.3625155
[9] C. Z. Huang and T. Cooijmans. (2012) JSB-
Cho ales-Da ase . [Online]. A ailable: h ps:
//gi hub.com/czhuang/JSB-Cho ales-da ase
[10] G. Hadje es, F. Pache , and F. Nielsen, “DeepBach:
a S ee able Model o Bach Cho ales Gene a ion,”
in P oc. 34 h In . Con . Machine Lea ning, se .
P oceedings o Machine Lea ning Resea ch, D. P ecup
and Y. W. Teh, Eds., ol. 70. PMLR, 06–11 Aug 2017,
pp. 1362–1371.
[11] Y. Wu, J. Ga dne , E. Manilow, I. Simon,
C. Haw ho ne, and J. Engel, “The Chambe
Ensemble Gene a o : Limi less High-Quali y MIR
Da a ia Gene a i e Modeling,” a Xi p ep in
a Xi :2209.14458, 2022.
[12] D. Shadle. (2023) Flo ence B. P ice Wo ks Ca alog.
[Online]. A ailable: h ps:// bp iceca alog.ne /
[13] F. B. P ice, 44 A Songs and Spi i ual A angemen s,
R. Hea d, Ed. Faye e ille, AR, USA: Cla Nan
Edi ions, 2015.
[14] K. Slack, “Beyond he yea s,” 2024, album.
[Online]. A ailable: h ps://www.a ie- eco ds.com/
eleases/beyond- he-yea s-ka en-slack/
[15] C.-Z. A. Huang, A. Vaswani, J. Uszko ei , I. Simon,
C. Haw ho ne, N. Shazee , A. M. Dai, M. D. Ho man,
M. Dinculescu, and D. Eck, “Music T ans o me ,” in
In . Con . Lea ning Rep esen a ions, 2019.
[16] M. Maland o, “Compose ’s Assis an : An In e ac i e
T ans o me o Mul i-T ack MIDI In illing,” in P oc.
24 h In . Socie y o Music In o ma ion Re ie al Con .,
Milan, I aly, 2023, pp. 327–334.
[17] M. Maland o, “Compose ’s Assis an 2: In e ac i e
Mul i-T ack MIDI In illing wi h Fine-G ained Use
Con ol,” in P oc. 25 h In . Socie y o Music
In o ma ion Re ie al Con ., San F ancisco, CA, USA,
2024, pp. 438–445.
[18] J. Zhao and G. Xia, “Accomon age: Accompanimen
A angemen Via Ph ase Selec ion and S yle T ans e ,”
in P oc. 22nd In . Socie y o Music In o ma ion
Re ie al Con ., Online, 2021, pp. 833–840.
[19] J. Ens and P. Pasquie , “MMM : Explo ing Condi ional
Mul i-T ack Music Gene a ion wi h he T ans o me ,”
a Xi p ep in a Xi : 2008.06048, 2020.
[20] J. Ens and P. Pasquie , “Flexible Gene a ion wi h he
Mul i-T ack Music Machine,” in Ex ended Abs ac s
o he La e-B eaking Demo Session o he 21s
In . Socie y o Music In o ma ion Re ie al Con .,
Mon éal, Canada, 2020.
[21] P. Pasquie , J. Ens, N. F ade , P. T iana, D. Rizzo i, J.-
B. Rolland, and M. Sa i, “MIDI-GPT: A Con ollable
Gene a i e Model o Compu e -Assis ed Mul i ack
Music Composi ion,” 2025. [Online]. A ailable: h ps:
//a xi .o g/abs/2501.17011
[22] R. Guo, I. Simpson, C. Kie e , T. Magnusson, and
D. He emans, “MusIAC: An Ex ensible Gene a i e
F amewo k o Music In illing Applica ions wi h
Mul i-le el Con ol,” in A i icial In elligence in Music,
Sound, A and Design. E oMUSART 2022. Lec u e
No es in Compu e Science, T. Ma ins, N. Rod íguez-
Fe nández, and S. M. Rebelo, Eds., ol. 13221.
Sp inge , Cham, 2022, pp. 341–356.
P oceedings o he 26 h ISMIR Con e ence, Daejeon, Ko ea, Sep embe 21-25, 2025
777
[23] P. Sa men o, A. Kuma , D. Xie, C. Ca , Z. Zukowski,
and M. Ba he , “Sh edGP: Gui a is s yle-
condi ioned abla u e gene a ion wi h ans o me s,” in
P oceedings o he 16 h In e na ional Symposium
on Compu e Music Mul idisciplina y Resea ch.
Zenodo, No . 2023, p. 112–121. [Online]. A ailable:
h ps://doi.o g/10.5281/zenodo.10110154
[24] J. Lo h, P. Sa men o, C. Ca , Z. Zukowski, and
M. Ba he , “P ogGP: F om gui a p o abla u e
neu al gene a ion o p og essi e me al p oduc ion,” in
P oceedings o he 16 h In e na ional Symposium on
Compu e Music Mul idisciplina y Resea ch. Zenodo,
No . 2023, p. 122–133. [Online]. A ailable: h ps:
//doi.o g/10.5281/zenodo.10110236
[25] F. P ice, “I Remembe ,” A chi al objec MC 988a 2
(objec ID 501224), Special Collec ions Depa men ,
Uni e si y o A kansas Lib a ies, Faye e ille, AR, box
12, Folde 27, Reposi o y ID: 501224; accessed June
23, 2025. [Online]. A ailable: h ps://ua k.as.a las-
sys.com/ eposi o ies/2/a chi al_objec s/501224
[26] D. J. H. J , “Flo ence B. P ice’s Composi ional
S yle in Song Se ings o Paul Lau ence
Dunba ,” Mas e ’s Thesis, Uni e si y o
Ne ada, Las Vegas, 2025. [Online].
A ailable: h ps://digi alschola ship.unl .edu/
cgi/ iewcon en .cgi?a icle=4621&con ex =
hesesdisse a ions
[27] C. Jobson, “Flo ence P ice: An Analysis o
Selec A Songs wi h Tex by Female Poe s,”
Doc o al Disse a ion, Uni e si y o Miami, 2019.
[Online]. A ailable: h ps://schola ship.miami.edu/
esplo o/ou pu s/doc o al/Flo ence-P ice-An-Analysis-
o -Selec /991031447224302976
[28] C. Ra el, N. Shazee , A. Robe s, K. Lee, S. Na ang,
M. Ma ena, Y. Zhou, W. Li, and P. J. Liu, “Explo ing
he Limi s o T ans e Lea ning wi h a Uni ied Tex -
o-Tex T ans o me ,” Jou nal o Machine Lea ning
Resea ch, ol. 21, no. 140, pp. 1–67, 2020.
[29] T. Wol , L. Debu , V. Sanh, J. Chaumond, C. Delangue,
A. Moi, P. Cis ac, T. Raul , R. Lou , M. Fun owicz,
J. Da ison, S. Shlei e , P. on Pla en, C. Ma, Y. Je ni e,
J. Plu, C. Xu, T. Le Scao, S. Gugge , M. D ame,
Q. Lhoes , and A. Rush, “T ans o me s: S a e-
o - he-A Na u al Language P ocessing,” in P oc.
2020 Con . Empi ical Me hods in Na u al Language
P ocessing: Sys em Demons a ions. Associa ion o
Compu a ional Linguis ics, Oc . 2020, pp. 38–45.
[30] S.-L. Wu and Y.-H. Yang, “The Jazz T ans o me on
he F on Line: Explo ing he Sho comings o AI-
Composed Music Th ough Quan i a i e Measu es,” in
P oc. 21s In . Socie y o Music In o ma ion Re ie al
Con ., Mon éal, Canada, 2020, pp. 142–149.
[31] S. Kullback and R. A. Leible , “On In o ma ion and
Su iciency,” The Annals o Ma hema ical S a is ics,
ol. 22, no. 1, pp. 79 – 86, 1951. [Online]. A ailable:
h ps://doi.o g/10.1214/aoms/1177729694
[32] L. Casini and B. L. T. S u m, “T ad o me :
A T ans o me Model o T adi ional Music
T ansc ip ions,” in P oceedings o he Thi y-
Fi s In e na ional Join Con e ence on A i icial
In elligence, IJCAI-22, L. D. Raed , Ed. In e na ional
Join Con e ences on A i icial In elligence
O ganiza ion, 2022, pp. 4915–4920, AI and
A s. [Online]. A ailable: h ps://doi.o g/10.24963/
ijcai.2022/681
[33] S. Mossmy , E. Halls öm, B. L. S u m, V. H.
Vegebo n, and J. Wedin, “F om Jigs and Reels o
Scho isa och Polsko : Gene a ing Scandina ian-like
Folk Music wi h Deep Recu en Ne wo ks,” in 16 h
Sound and Music Compu ing Con e ence (SMC2019),
2019.
[34] M. Plasse , S. Pe e , and G. Widme , “Disc e e
Di usion P obabilis ic Models o Symbolic
Music Gene a ion,” in P oceedings o he Thi y-
Second In e na ional Join Con e ence on A i icial
In elligence, IJCAI-23, 8 2023, pp. 5842–5850.
[35] A. Hol zman, J. Buys, L. Du, M. Fo bes, and Y. Choi,
“The Cu ious Case o Neu al Tex Degene a ion,” in
In . Con . Lea ning Rep esen a ions, 2020.
[36] M. McCloskey and N. J. Cohen, “Ca as ophic
in e e ence in connec ionis ne wo ks: The sequen ial
lea ning p oblem,” se . Psychology o Lea ning
and Mo i a ion, G. H. Bowe , Ed. Academic
P ess, 1989, ol. 24, pp. 109–165. [Online].
A ailable: h ps://www.sciencedi ec .com/science/
a icle/pii/S0079742108605368
P oceedings o he 26 h ISMIR Con e ence, Daejeon, Ko ea, Sep embe 21-25, 2025
778
