MUSGO: A COMMUNITY-DRIVEN FRAMEWORK
FOR ASSESSING OPENNESS IN MUSIC-GENERATIVE AI
Rose Ba lle-Roca1Lau a Ibáñez-Ma ínez1Xa ie Se a1
Emilia Gómez1,2Ma ín Rocamo a1
1Music Technology G oup, Uni e si a Pompeu Fab a, Ba celona Spain
2Join Resea ch Cen e, Eu opean Commission, Se ille, Spain
{ ose .ba lle, lau a.ibanez, ma in. ocamo a}@up .edu
ABSTRACT
Since 2023, gene a i e AI has apidly ad anced in he
music domain. Despi e signi ican echnological ad ance-
men s, music-gene a i e models aise c i ical e hical chal-
lenges, including a lack o anspa ency and accoun abili y,
along wi h isks such as he eplica ion o a is s’ wo ks,
which highligh s he impo ance o os e ing openness.
Wi h upcoming egula ions such as he EU AI Ac encou -
aging open models, many gene a i e models a e being e-
leased labelled as ‘open’. Howe e , he de ini ion o an
open model emains widely deba ed. In his a icle, we
adap a ecen ly p oposed e idence-based amewo k o
assessing openness in LLMs o he music domain. Us-
ing eedback om a su ey o 110 pa icipan s om he
Music In o ma ion Re ie al (MIR) communi y, we e ine
he amewo k in o MusGO (Music-Gene a i e Open AI),
which comp ises 13 openness ca ego ies: 8 essen ial and 5
desi able. We e alua e 16 s a e-o - he-a gene a i e mod-
els and p o ide an openness leade boa d ha is ully open
o public sc u iny and communi y con ibu ions. Th ough
his wo k, we aim o cla i y he concep o openness in
music-gene a i e AI and p omo e i s anspa en and e-
sponsible de elopmen .
1. INTRODUCTION
Music-gene a i e AI is in oducing c i ical e hical con-
ce ns, pa icula ly ega ding i s impac on c ea i e p o-
cesses and au ho ship, po en ial legal issues om da a mis-
use, and dis up ions o exis ing business and in ellec ual
p ope y (IP) models [1–5]. Fu he mo e, hese echnolo-
gies usually exhibi a Wes e n cul u al bias, unde mining
di e si y in musical exp ession and homogenising music
gen es. Addi ionally, hey o en equi e subs an ial com-
pu a ional esou ces, con ibu ing o ene gy consump ion
amid he clima e c isis [4,6]. A u he conce n is he lack
o anspa ency in hese models, which limi s sc u iny o
© R. Ba lle-Roca, L. Ibáñez-Ma ínez, X. Se a, E. Gómez,
and M. Rocamo a. 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: R. Ba lle-Roca,
L. Ibáñez-Ma ínez, X. Se a, E. Gómez, and M. Rocamo a, “MusGO:
A Communi y-D i en F amewo k o Assessing Openness in Music-
Gene a i e AI”, 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.
hei beha iou , a ibu ion, and decision-making [5], as
well as he po en ial eplica ion o aining da a [7–9].
To add ess hese challenges, we ad oca e o he use o
open models. Open models ha e a i al ole in open sci-
ence, p omo ing anspa ency, accoun abili y, and inno a-
ion [10–12]. Alongside academic e o s [13], ech com-
panies a e inc easingly labelling AI sys ems as open. How-
e e , se e al ini ia i es p o ide only a ew componen s,
such as model weigh s and in e ence code [11], which lim-
i s euse and anspa ency, and leads o c i icism o ‘open-
washing’. Ye , de ining openness o AI models is chal-
lenging and emains a subjec o ac i e deba e. In ac ,
open model s a us is becoming highly a ac i e wi h he
legal excep ions in he EU AI Ac , making i imely o ex-
amine he equi emen s needed o a ain i [11,14].
The absence o a de ined me hodology o e alua ing
openness in music-gene a i e models c ea es a po en ial
gap in hei assessmen . Consequen ly, we pose he ques-
ion: how can openness be assessed in music-gene a i e
AI? We d aw upon and adap he wo k by Liesen eld and
Dingemanse (2024) [11], which in oduces an e idence-
based me hod o e alua ing openness in la ge language
models (LLMs). Thei p oposed amewo k is composi e
and g aded, consis ing o mul iple ca ego ies, each wi h
h ee le els o openness. Building on and ailo ing his
app oach o he music domain, we in oduce MusGO
(Music-Gene a i e Open AI): a communi y-d i en ame-
wo k o assessing openness in music-gene a i e AI. We
de elop MusGO by ac i ely in ol ing he Music In o ma-
ion Re ie al (MIR) communi y h ough a su ey and e-
ining he c i e ia based on hei eedback.
Ou main con ibu ions include (1) a ep oducible
amewo k 1 o e alua ing openness in music-gene a i e
AI, (2) a leade boa d 2showcasing openness ac oss 16
models, and (3) a ully open eposi o y 3 o enable public
sc u iny, communi y con ibu ion and u u e upda es. Wi h
his, we aim o suppo u u e wo k by p o iding a mo e
nuanced unde s anding o openness in music-gene a i e
AI, o e ing a ine-g ained e idence-based app oach o in-
o m decisions and p omo e esponsible p ac ices.
1h ps:// ose ba lle oca.gi hub.io/MusGO_ amewo k/ amewo k.
h ml
2h ps:// ose ba lle oca.gi hub.io/MusGO_ amewo k/index.h ml
3h ps://gi hub.com/ ose ba lle oca/MusGO_ amewo k
727
2. BACKGROUND AND RELATED WORK
2.1 De ining ‘open’ models
Documen a ion e o s in AI ha e suppo ed model
anspa ency by disclosing de elopmen p ocesses, da a
sou ces, and model a ibu es [15–18]. Howe e , de ining
openness in AI is challenging, as i in ol es mul iple com-
ponen s (e.g., sou ce code, documen a ion, model weigh s,
aining da a) [13] and equi es balancing openness ideals
wi h p ac ical conce ns such as da a p i acy, secu i y, and
IP igh s [19]. Recen ini ia i es ha e con ibu ed o shap-
ing such a de ini ion. One o hese is he Open Sou ce Ini-
ia i e (OSI), which has led a public consul a ion o adap
i s widely accep ed de ini ion o open-sou ce so wa e o
AI sys ems, esul ing in he elease o he Open Sou ce
AI De ini ion (OSAID) 1.0 in Oc obe 2024 [20]. I e-
qui es an AI sys em o p o ide access o i s sou ce code
and model weigh s, along wi h su icien in o ma ion abou
he aining da a. The goal is o ensu e hese models can
be eely used, modi ied, and sha ed, he eby p omo ing
anspa ency and collabo a ion wi hin he AI communi y.
While OSAID ep esen s an impo an s ep owa d se -
ing a s anda d o AI openness, i has also spa ked discus-
sion and c i icism [21–23]. Some a gue ha he s ingen
equi emen s may be challenging o mee , possibly exclud-
ing models ha , al hough no ully complian , s ill con-
ibu e meaning ully o he open-sou ce ecosys em. An-
o he conce n aised by open da a ad oca es is ha OSAID
does no equi e he aining da a i sel o be sha ed, only
de ailed in o ma ion abou i , which is some imes seen as
a o m o ‘open-washing’. This is pa icula ly ele an in
music, whe e much o he da a, such as eco dings, com-
posi ions, and ly ics, is owned by igh s holde s and o en
canno be made publicly a ailable. The eedom o use a
model o any pu pose also aises conce ns abou unin en-
ional copy igh iola ions o he delibe a e imi a ion o an
a is ’s s yle wi hou hei consen [5, 6, 24]. In con as ,
Responsible AI Licenses (RAIL) p opose a mo e nuanced
o m o openness by including es ic ions on ce ain uses.
Among o he egula o y e o s, he Eu opean Union’s
AI Ac [25] has g an ed some exemp ions o models e-
leased unde a ee and open-sou ce license, as speci ied in
A icle 53 (1)(a) and (b) [26]. The public a ailabili y e-
qui emen appea s o co e “(...) i s pa ame e s, including
he weigh s, he in o ma ion on he model a chi ec u e, and
he in o ma ion on model usage”, wi h no explici men ion
o he aining da a [27]. The Gene al-Pu pose AI Code o
P ac ice is expec ed o p o ide u he de ails on he obli-
ga ions unde A icles 53 and 55 o di e en ways o e-
leasing gene al-pu pose AI models, including open sou c-
ing, al hough he documen is s ill in d a o m [28].
Ra he han concei ing openness as a bina y s a us,
o he p oposals a e ie ed, wi h openness dis ibu ed ac oss
di e en dimensions [14, 29]. This aligns wi h p io wo k
on AI accoun abili y and anspa ency amewo ks [30–
33]. In his iew, a model’s openness is de e mined by
he ela i e openness o i s indi idual componen s, accom-
moda ing cases whe e no all c i e ia a e ully a ained.
2.2 How o assess openness?
Building on he idea ha openness is no a bina y a ibu e
bu a he a composi e o a ious elemen s, each a y-
ing ac oss di e en deg ees, se e al ini ia i es ha e aimed
o c ea e s uc u ed, ie ed, g adien -based amewo ks o
e alua e i [11, 34–36].
Fo ins ance, he Linux Founda ion’s Model Openness
F amewo k (MOF) [35] dis inguishes h ee ie s o open
AI sys ems. I p omo es comple eness and openness o e-
p oducibili y, anspa ency, and usabili y, ca ego ising AI
models based on hei le el o openness and ensu ing he
elease o key a e ac s like model a chi ec u e, aining
da a, and e alua ion esul s unde open licenses. Simila ly,
he Founda ion Model T anspa ency Index (FMTI) [34]
e alua es anspa ency in ounda ion models using 100 in-
dica o s ac oss h ee b oad domains—ups eam, model,
and downs eam— ocusing on aspec s such as da a, com-
pu e esou ces, model capabili ies, and downs eam im-
pac . While comp ehensi e, he FMTI does no allow in-
di idual da a poin s o be sc u inized o con es ed, limi ing
he anspa ency and e i iabili y o i s sco es [11].
On he o he hand, Ei as e al. [36,37] ha e con ibu ed
aluable insigh s in o he isks and oppo uni ies associa ed
wi h open-sou ce gene a i e AI models, p oposing a ax-
onomy o e alua ing hei le el o openness, pa icula ly
ocusing on he a ailabili y o code and da a. Al hough ad-
oca ing o AI democ a isa ion, hey emphasise he need
o openness o be balanced wi h esponsible AI p ac ices
o a oid isks such as misuse, bias, and unin ended conse-
quences. Thei e alua ion amewo k highligh s bo h ech-
nical openness (e.g., code and model weigh s a ailabili y)
and e hical conce ns (e.g., da a p i acy and model sa e y),
o e ing a mo e comp ehensi e pe spec i e on open-sou ce
AI compa ed o o he app oaches.
A pa icula ly ele an amewo k is he one p oposed
by Liesen eld and Dingemanse (2024) [11] which, build-
ing on p e ious wo k [29], o e s an e idence-based
me hod o e alua ing openness in LLMs, and ex ends i
o ex - o-image models. The p oposed amewo k is com-
posi e—i conside s mul iple dimensions when de e min-
ing openness—and g aded—meaning each ca ego y may
ha e di e en le els o openness: closed,pa ial and ully
open. Many o hese dimensions align wi h hose om
o he amewo ks, pa icula ly MOF. The app oach has
been es ed since July 2023 in an open leade boa d 4 ha
acks he deg ee o openness o se e al LLMs. A dis in-
guishing ea u e o his me hod is i s communi y-d i en ap-
p oach, which allows o he inspec ion and appeal o he
model’s assessmen sco es and hei suppo ing e idence.
Howe e , while he amewo k is adap able o a ious do-
mains, i is essen ial o e ise ce ain ca ego ies and in-
clude domain-speci ic elemen s o accu a ely assess open-
ness [11]. In he case o music, unique challenges, such
as he copy igh ed na u e o mos aining da a, unde sco e
he need o ailo ed c i e ia and domain-speci ic p ac ices
o ensu e a comp ehensi e openness assessmen .
4h ps://opening-up-cha gp .gi hub.io/
P oceedings o he 26 h ISMIR Con e ence, Daejeon, Ko ea, Sep embe 21-25, 2025
728
3. OPENNESS FRAMEWORK FOR MUSIC AI
3.1 F om LLMs o music
We adop ed he e idence-based amewo k in oduced by
Liesen eld and Dingemanse (2024) [11] and ailo ed i o
he music domain. Ou ini ial adap a ion in ol ed mod-
i ying e e ences o LLMs o align wi h music-gene a i e
models, o example by enaming LLM-o ien ed labels and
excluding ins uc ion uning- ela ed ca ego ies, which led
o a e ised amewo k o 11 ca ego ies om he o iginal
14. To ensu e he adap ed amewo k e lec ed he pe -
spec i es and p io i ies o he MIR communi y, we aimed
o ga he eedback on each ca ego y’s in en ion and el-
e ance, conside ing he isks o o e - eliance on single
ea u es, such as access o licensing, o de e mine open-
ness [11]. To his end, we conduc ed an anonymous online
ques ionnai e wi hin he MIR communi y.
3.2 MIR communi y su ey
3.2.1 Me hodology
The su ey included a s a emen o each ca ego y e lec -
ing he op open le el ( ully open). Pa icipan s we e asked
o assess he p oposed s a emen by (1) ele ance—how
impo an hey conside he ca ego y in de e mining he
openness o a model— using a 5-poin Like scale (1 = no
ele an , 5 = e y ele an ), and (2) ag eemen —whe he
hey ag eed, somewha ag eed, o disag eed wi h he p o-
posed s a emen . They we e encou aged o sugges modi-
ica ions and e inemen s o he s a emen s when hey dis-
ag eed o had ese a ions abou he c i e ia. The su ey
also included an open commen s sec ion o pa icipan s o
p opose new ca ego ies and aise unadd essed conce ns.
The su ey was open o 8 weeks and ecei ed 110 e-
sponses om pa icipan s linked o he MIR communi y.
Al hough we in ended b oade ep esen a ion, he sample
was biased owa ds male academics in Eu ope and No h
Ame ica. Howe e , his dis ibu ion aligns wi h he ypical
demog aphics o a endees a he In e na ional Socie y o
Music In o ma ion Re ie al (ISMIR) con e ence. 5
3.2.2 Resul s
Table 1 summa ises he su ey esul s, showing he me-
dian ele ance sco e (M) and ag eemen le els o each
ca ego y. O e all, pa icipan s mos ly ag eed wi h he
p oposed s a emen s, indica ing a gene al consensus on
he p oposed c i e ia. Ca ego ies 3: Model Weigh s,4:
Code documen a ion and 1: Open Code showed he high-
es le els o ag eemen (wi h o e 78% selec ing Yes,
I ag ee), and we e also a ed highly ele an (M1,3=5,
M4=4). Fo ca ego y 2: T aining da a, while ele ance
was high (M2=5), he e was signi ican discussion ega d-
ing he aming o he s a emen (30% selec ed Yes, I some-
wha ag ee). These esul s sugges ha elemen s ela ed o
code access and model ep oducibili y a e o g ea in e es
o he communi y.
5Complemen a y in o ma ion on he su ey, including a de ailed anal-
ysis o he pa icipan s’ backg ounds and hei esponses, is a ailable a :
h ps:// ose ba lle oca.gi hub.io/MusGO_ amewo k/su ey.h ml
Table 1: Resul s om he MIR communi y su ey, show-
ing o each ca ego y i s median ele ance and le el o
ag eemen (%) wi h he p oposed s a emen . No e ha he
c i e ia co espond o he i s adap ed e sion (see 3.1).
C i e ia Rele ance Ag eemen (%)
Median Yes, I
ag ee.
Yes, I some-
wha ag ee.
No, I do
no ag ee.
1: Open Code 5 78.18 20.00 1.82
2: T aining Da a 5 67.27 30.00 2.73
3: Model Weigh s 5 79.09 18.18 2.73
4: Code documen a ion 4 79.09 18.18 2.73
5: A chi ec u e 4 62.73 30.00 7.27
6: Pape and P ep in 4 64.55 21.82 13.64
7: Model Ca d 4 54.55 30.00 15.45
8: Da ashee 4 64.55 28.18 7.27
9: Package 3 46.36 32.73 20.91
10: API 3 39.09 30.00 30.91
11: Licensing 4 62.73 28.18 9.09
In con as , ca ego ies such as 9: Package and 10: API,
showed lowe and mo e a ied ag eemen a es (46.36%
and 39.09%, espec i ely) and lowe ele ance (M9,10=3).
This was e lec ed in he commen s, which highligh ed
conce ns abou he addi ional e o equi ed o p o ide
hese ea u es and ques ioned hei necessi y o de e min-
ing openness.
The su ey also highligh ed some ca ego ies wi h mod-
e a e ag eemen , such as 6: Pape and P ep in ,7: Model
Ca d and 8: Da ashee , all wi h a median ele ance sco e
o 4, bu wi h no able disag eemen a es. This di e si y o
esponses unde sco es he a ying p io i ies among com-
muni y membe s ega ding de ailed documen a ion ac oss
di e en a eas o he model. Commen s on hese ca ego ies
we e key o unde s anding he communi y’s pe spec i es
and e ining he ini ial c i e ia.
3.3 Re ining c i e ia based on communi y eedback
Pa icipan s’ eedback e ealed gaps and disag eemen s
ha we add essed by ca e ully e iewing all commen s,
iden i ying ecu ing issues, and e ining he c i e ia ac-
co dingly. The mos subs an ial changes in ol ed cla i-
ying he desc ip ions o each ca ego y and de ining wha
cons i u ed pa ial and ully open o each case. We in-
eg a ed music-speci ic nuances aised in he eedback,
including he challenges o sha ing IP-p o ec ed aining
da a, he need o p o ide soni ied examples, and he use
expe ience needs o musicians, who may ha e limi ed
echnical skills and domain-speci ic so wa e p e e ences,
ex ending beyond he use o APIs. In addi ion, we held
in e nal discussions wi h se e al esea che s om he Mu-
sic Technology G oup (MTG), which p o ided u he in-
sigh s ha helped e ine he amewo k.
Following he su ey esul s ega ding ca ego y ele-
ance, we in oduced wo dis inc le els wi hin ou ame-
wo k: essen ial, o ca ego ies conside ed key o assess-
ing openness, and desi able, o hose ha enhance open-
ness and p o ide addi ional alue bu a e no indispens-
able. Fu he mo e, eedback highligh ed he absence o a
ca ego y add essing model e alua ion p ocedu es. In e-
sponse, we in oduced a new ca ego y, E alua ion p o-
P oceedings o he 26 h ISMIR Con e ence, Daejeon, Ko ea, Sep embe 21-25, 2025
729
cedu e, co e ing e alua ion da a, me ics, and esul s on
model pe o mance. We also added he ca ego y Supple-
men a y ma e ial page o acknowledge he impo ance o
dedica ed websi es ha p o ide usage ins uc ions, model
demons a ions and soni ied ou pu s, which a e essen ial in
music esea ch, whe e audio canno be embedded wi hin
he pape i sel . Addi ionally, we enamed some ca e-
go ies o be e e lec hei scope: Open code →Sou ce
code,A chi ec u e →T aining p ocedu e,Pape →Re-
sea ch Pape , and API →Use -o ien ed applica ion. The
enamed ca ego y Use -o ien ed applica ion also includes
o he use -expe ience app oaches, such as eal- ime appli-
ca ions, ecognising he need o accessible ools o mu-
sicians and compose s. These music-speci ic adap a ions
we e c i ical o make he amewo k ele an o music-
gene a i e models, which di e signi ican ly om LLMs.
3.4 The MusGO amewo k
As a esul o his modi ica ion and i e a ion p ocess, we
buil he Music-Gene a i e Open AI amewo k. MusGO
comp ises 13 ca ego ies: 8 essen ial and 5 desi able. The
essen ial ca ego ies include: (1) Sou ce code, (2) T ain-
ing da a, (3) Model weigh s, (4) Code documen a ion,
(5) T aining p ocedu e, (6) E alua ion p ocedu e, (7) Re-
sea ch pape , and (8) Licensing, while desi able ca e-
go ies a e de ined by (9) Model ca d, (10) Da ashee , (11)
Package, (12) Use -o ien ed applica ion, and (13) Sup-
plemen a y ma e ial page. Essen ial ca ego ies ollow an
openness-g aded scale o 3 le els: closed (✗), pa ial (∼)
and ully (open) (✓). Ins ead, desi able ca ego ies a e bi-
na y: whe he ha elemen exis s (⭑) o no , as hey a e
conside ed complemen a y add-ons o he model, a he
han co e componen s, which allows suppo ing any e o s
made in hese a eas. A comple e desc ip ion o he c i e ia
is a ailable on ou complemen a y websi e.
4. ASSESSING OPENNESS
4.1 Model selec ion
We selec ed 16 s a e-o - he-a music gene a ion mod-
els: 6GANSyn h [38], Jukebox [39], RAVE [40], Musika
[41], Moûsai [42], MusicGen [43], MusicLM [44],
VampNe [45], MusicLDM [46], Music Con olNe [47],
Noise2Music [48], MeLoDy [49], DITTO-2 [50], Di -A-
Ri [51], JASCO [52], and S able Audio Open [53]. Ou
selec ion encompasses a di e se se o music-gene a i e
models wi h wa e o m ou pu s, mainly ocusing on hose
ha explici ly claim o be ‘open’ o a e associa ed wi h
openness, alongside popula models ha exhibi open-
ness beha iou s (e.g., publishing model weigh s, eleas-
ing esea ch pape s, o p o iding use -o ien ed applica-
ions). In addi ion o his, we aimed o co e a a ie y
o a chi ec u es commonly used wi hin he ield, including
Va ia ional Au oencode s (VAEs) [54], Gene a i e Ad e -
sa ial Ne wo ks (GANs) [55], T ans o me s [56], Di u-
sion models [57], and no el me hodologies such as Flow
Ma ching [58].
6In oduced om oldes o newes esea ch pape elease.
4.2 MusGO in o p ac ice
We e alua ed he selec ed models using a s uc u ed
me hodology based on a checklis o ully open le el e-
qui emen s o each ca ego y. Each model was examined
by one o he au ho s, including jus i ica ion s a emen s.
Two o he au ho s hen e iewed he e alua ions o ensu e
consis ency and minimise bias. In case o disc epancy, we
engaged in discussions and adjus ed he assessmen i e -
a i ely, ollowing consensual quali a i e esea ch p inci-
ples [59]. We ga he ed in o ma ion om o icial publish-
e s and main aine s and, when a ailable, complemen ed i
wi h hi d-pa y sou ces such as HuggingFace. 7
4.3 Openness assessmen
Figu e 1 p esen s he openness e idence-based assessmen
o he 16 selec ed models, o ming he basis o ou open-
ness leade boa d. Models a e o de ed using a weigh ed
openness sco e (O), based on essen ial ca ego ies (E) and
no malised o a 100-poin scale. Conside ing su ey ind-
ings, he h ee mos ele an ca ego ies (E1,E2and E3,
all wi h M1,2,3=5) a e weigh ed wice as much as he o h-
e s. No e ha he sco e is used o o de ing pu poses only,
and we do no in end o educe openness o a single alue.
When models achie e he same sco e, he o de is de e -
mined by he highes numbe o ul illed desi able ca e-
go ies.
The assessmen e eals signi ican a ia ion in open-
ness ac oss he conside ed models. The ca ego y T ain-
ing p ocedu e is he mos open, wi h 11 models ully open
and 5 pa ially open. Ins ead, he ca ego y T aining da a
is he mos closed, wi h only one model (S able Audio
Open [53]) achie ing ully open s a us. Rega ding sou ce
code, eigh models p o ide ull access, and wo p o ide
pa ial access. The emaining six models, besides no p o-
iding sou ce code, also do no disclose model weigh s,
code documen a ion, o apply app op ia e licensing. Mod-
els classi ied as ully open o ca ego y Model weigh s
also end o p o ide comp ehensi e documen a ion o hei
code and a e ypically licensed unde an OSI o RAIL li-
cense, e lec ing a co ela ion be ween sou ce code, model
weigh s, code documen a ion, and licensing. Rega ding
he ca ego y E alua ion p ocedu e, mos models (11/16)
a e pa ially open, o en lacking e alua ion da a o im-
plemen a ion de ails, while only i e a e ully open. Al-
hough all models analysed p o ide a esea ch pape o
equi alen echnical epo , only 11 mee he pee - e iewed
and accessibili y equi emen s. Desi able ca ego ies also
show a subs an ial a iabili y ac oss models. The ca ego y
Da ashee is he leas ul illed, wi h only one model (Mu-
sicLDM [46]) p o iding de ailed and s uc u ed da a doc-
umen a ion. Ins ead, all models include a supplemen a y
ma e ial page, which demons a es ha accompanying e-
sou ces wi h model demons a ions and sound examples
has become a communi y no m. These esul s a e made
publicly a ailable h ough a con inuously upda ed leade -
boa d, helping o iden i y incomple e openness claims and
po en ial cases o ‘open-washing’.
7h ps://hugging ace.co/
P oceedings o he 26 h ISMIR Con e ence, Daejeon, Ko ea, Sep embe 21-25, 2025
730
Figu e 1: Openness leade boa d o 16 music-gene a i e models assessed using he MusGO amewo k.
5. DISCUSSION
5.1 Music-gene a i e open AI
Ou assessmen e eals a signi ican ly di e se landscape o
music-gene a i e models in e ms o openness, highligh -
ing bo h no able e o s and signi ican oom o imp o e-
men . This si ua ion unde sco es di e en le els o com-
mi men ac oss he communi y, pa icula ly ega ding c i -
ical ca ego ies such as aining da a, sou ce code, and li-
censing. Al hough he analysed models gene ally pe o m
well in p o iding documen a ion, he le el o de ail is no
always su icien o gua an ee comple e anspa ency and
accoun abili y. Fu he mo e, desi able ca ego ies a e less
commonly add essed, especially da ashee s, poin ing o a
gap in he in o ma ion a ailable abou hese models.
Ou e alua ion alida es he MusGO amewo k as
a comp ehensi e e idence-based esou ce o assessing
openness in music-gene a i e AI. Using a composi e and
g aded amewo k allows o a mo e lexible and ine-
g ained pe spec i e on model openness. Fo ins ance,
while T aining p ocedu e is o en he mos open ca ego y,
T aining da a emains c i ical, wi h he majo i y o models
lacking su icien ly de ailed desc ip ions o he da a used.
Howe e , we acknowledge ha sha ing aining da a in he
music domain may pose signi ican challenges due o IP
igh s, which can lead o e hical and legal conce ns when
publishing his in o ma ion. We add ess his issue by con-
side ing T aining da a ully open when di ec access is e-
s ic ed due o legal conce ns, p o ided ha de ailed in o -
ma ion abou all sou ces is disclosed.
We p esen a con inuously upda ed leade boa d show-
casing openness assessmen s using he MusGO ame-
wo k. I e lec s he cu en s a e o openness in music-
gene a i e models and ac s as a li e collabo a i e pla o m
o he communi y. Resea che s, de elope s, and s ake-
holde s a e in i ed o ac i ely con ibu e by e alua ing new
models and engaging in discussion. This app oach enables
acking model e olu ion and in eg a ing eme ging aspec s
like con ollabili y and eal- ime use. The eposi o y p o-
ides guidelines on how o con ibu e, including he de-
ailed amewo k c i e ia, ins uc ions o issues and pull
eques s, and in o ma ion on submissions e iew. 8
8h ps://gi hub.com/ ose ba lle oca/MusGO_ amewo k/ ee/main
/p ojec s/README.md
P oceedings o he 26 h ISMIR Con e ence, Daejeon, Ko ea, Sep embe 21-25, 2025
731
5.2 Beyond openness
While openness is c ucial o ensu ing anspa ency and
accoun abili y in music-gene a i e models, i alone does
no gua an ee e hical beha iou o add ess b oade impli-
ca ions. Se e al su ey pa icipan s also aised his con-
ce n. Ba ne [4] demons a es ha mos li e a u e on gen-
e a i e AI in music ends o o e look he po en ial nega-
i e e hical consequences o hese models. Howe e , ou
analysis sugges s a shi wi h inc easing a en ion ocused
on hese conce ns. Among he analysed models, 62.5%
(10/16) include e hical conside a ions, wi h a no able ise
in pape s published a e 2023 (83.3% o hese models).
Openness can help e eal whe e e hical isks a ise, es-
pecially in sensi i e a eas such as copy igh . While sha ing
code and da a suppo s anspa ency, i does no add ess
he legali y o he da a used o aining. Fo ins ance,
MusicGen [43] and JASCO [52] ob ained da a h ough
p ope legal ag eemen s, whe eas S able Audio Open [53]
is he only model using ully a ailable and accessible da a.
This highligh s he need o balance openness wi h ca e ul
copy igh conside a ions. Thus, MusGO helps iden i y and
e alua e whe he e hical claims a e subs an ia ed.
Simila ly, openness can poin ou whe he e o s o
mi iga e ha m ul o inapp op ia e uses ha e been imple-
men ed. Models like MusicLDM [46] and S able Audio
Open [53] y o mi iga e isks by pe o ming memo isa-
ion analysis o help ensu e ha aining da a is no inad-
e en ly ep oduced by hei models. In addi ion, Musi-
cLM [44] and Noise2Music [48] add ess he isks o pla-
gia ism and cul u al app op ia ion by inco po a ing sa e-
gua ds o mi iga e such issues. While MusGO does no di-
ec ly e alua e hese aspec s, u u e e sions o he ame-
wo k could help emphasise isk mi iga ion s a egies.
Ano he key conce n is he ep esen a ion and di e si y
o he aining da a. Open da ase s may s ill be biased,
wi h limi ed cul u al di e si y. Fo example, JASCO [52]
acknowledges ha i s da ase is hea ily Wes e n-cen ic,
which limi s he di e si y o he gene a ed music. Simi-
la ly, Noise2Music [48] poin s ou ha biases in aining
da a can mis ep esen musical gen es. These examples
show ha MusGO can help expose limi a ions in cul u al
di e si y and ep esen a ion by making da ase composi-
ion and documen a ion mo e accessible o sc u iny.
Finally, he economic impac o gene a i e AI on musi-
cians and c ea i e p o essionals is also a c i ical conce n.
Open models inc ease accessibili y, bu hey can po en-
ially displace human c ea o s. Many au ho s s ess ha
gene a i e models should enhance, no eplace, human c e-
a i i y. Ye , conce ns emain abou he dis up ion o adi-
ional indus ies and he economic impac on musicians.
This emphasises he need o openness o be pai ed wi h a
b oade discussion abou he e hical, social, and economic
implica ions o AI, including po en ial usage es ic ions.
Thus, e hical guidelines and sa egua ds a e essen ial o en-
su e ha AI-gene a ed music does no pe pe ua e ha m ul
s e eo ypes o in inge on he igh s o o iginal c ea o s.
MusGO o e s a s uc u ed lens o e iew models’ go e -
nance and se e as a ounda ion o e hical sc u iny.
5.3 Limi a ions
E alua ing he ca ego y T aining da a has p o en pa icu-
la ly challenging. In he music domain, da a accessibili y
is o en es ic ed by IP igh s and legal cons ain s, and
e en e ealing aining da a migh equi e a closed audi
p ocess o comply wi h hese equi emen s. Inco po a ing
such a le el o de ail in o ou amewo k would equi e a
mo e complex app oach, which may no be easible in his
con ex . Mo eo e , while ha dwa e equi emen s a e con-
side ed wi hin he T aining p ocedu e ca ego y, hei im-
pac on openness has no been ho oughly analysed. Some
models can be ained on pe sonal compu e s, while o h-
e s equi e ex ensi e compu a ional esou ces, complica -
ing ep oducibili y and limi ing accessibili y. Fo models
wi h lowe esou ce demands, ca ego ies like T aining da a
may become less ele an , shi ing he ocus owa d en-
abling a is s o ain models on hei own music.
In de eloping he MusGO amewo k, we su eyed he
MIR communi y. Howe e , we acknowledge ha ou sam-
ple was biased owa ds male academics based in Eu ope
and No h Ame ica. Expanding he sample in he u u e
o include a is s and o he s akeholde s, as well as pe -
spec i es om a b oade ange o egions, could p o ide
aluable insigh s and lead o adap a ions ha would make
he amewo k mo e inclusi e.
While MuGO p o ides a obus assessmen o open-
ness, he leade boa d does no e lec he b oade e hical
and socie al implica ions o gene a i e AI. This emains
a key di ec ion o u u e imp o emen , especially as im-
plica ions ela ed o ha dwa e equi emen s, eal- ime use,
and accessibili y a e highly ele an o a is s.
6. CONCLUSION
Wi h he ise o music-gene a i e AI, deba es a ound
he e hical implica ions o hese models ha e in ensi-
ied. Gi en he lack o anspa ency and accoun abil-
i y in hese sys ems, we ad oca e o open models. Ye ,
wha cons i u es an open model emains unde ined o
music-gene a i e AI. In his wo k, we adap an exis -
ing openness amewo k o LLMs o music-gene a i e
AI. A communi y-speci ic su ey helped us iden i y gaps
and pa icula conside a ions unique o he music do-
main. As a esul o ailo ing he c i e ia based on he
ga he ed eedback, we in oduce he MusGO amewo k.
We pu MusGO in o p ac ice by analysing 16 s a e-o -
he-a models, p o iding a public and upda able open-
ness leade boa d. Ou analysis e eals ha openness in
music-gene a i e AI is a ield in p og ess, whe e signi i-
can gaps emain in he a ailabili y o aining da a, model
weigh s, and licensing. MusGO allows o a s uc u ed
and lexible pe spec i e on model openness by conside -
ing complemen a y cha ac e is ics. I epo s on a model’s
s a us, and highligh s po en ial cases o ‘open-washing’.
As gene a i e models con inue o e ol e, we s i e o
mo e consis en and ho ough openness p ac ices, pa icu-
la ly when sha ing aining da a and sou ce code, o os e
anspa ency, accoun abili y, and esponsible de elopmen
in music-gene a i e AI.
P oceedings o he 26 h ISMIR Con e ence, Daejeon, Ko ea, Sep embe 21-25, 2025
732
7. ETHICS STATEMENT
This s udy in ol ed a olun a y, anonymous online su -
ey aimed a ga he ing eedback om he MIR commu-
ni y on a p elimina y adap ed openness amewo k. Pa -
icipan s we e in o med abou he scope and pu pose o
he s udy, as well as he in ended use o he collec ed
da a. Rega ding s udy design, pa icipan in o ma ion and
da a p o ec ion, we adhe ed o he guidelines and ecom-
menda ions o he Ins i u ional Commi ee o E hical Re-
iew o P ojec s (CIREP) a Uni e si a Pompeu Fab a. In
line wi h he Gene al Da a P o ec ion Regula ion (GDPR)
2016/679 (EU), all esponses we e anonymised and s o ed
secu ely. Pa icipan s we e in o med o hei da a igh s,
including access, ec i ica ion, dele ion, and wi hd awal o
consen , in acco dance wi h GDPR p o ocols.
8. ACKNOWLEDGMENTS
This wo k has been suppo ed 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 opean Union-Nex Gene -
a ion EU, and IMPA: Mul imodal AI o Audio P ocessing
(PID2023-152250OB-I00), unded by he Minis y o Sci-
ence, Inno a ion and Uni e si ies o he Spanish Go e n-
men , he Agencia Es a al de In es igación (AEI) and co-
inanced by he Eu opean Union. We hank ou colleagues
a he Music Technology G oup a Uni e si a Pompeu
Fab a o hei hough ul insigh s, cons uc i e discus-
sions and ac i e engagemen h oughou he de elopmen
o his wo k.
9. REFERENCES
[1] F. Ca no alini and A. Rodà, “Compu a ional c ea i -
i y and music gene a ion sys ems: An in oduc ion o
he s a e o he a ,” F on ie s in A i icial In elligence,
ol. 3, 2020.
[2] E. Gómez, M. Blaauw, J. Bonada, P. Chandna, and
H. Cues a, “Deep lea ning o singing p ocessing:
Achie emen s, challenges and impac on singe s and
lis ene s,” a Xi p ep in a Xi :1807.03046, 7 2018.
[Online]. A ailable: h ps://a xi .o g/abs/1807.03046
1
[3] B. L. T. S u m, M. Iglesias, O. Ben-Tal, M. Mi on, and
E. Gómez, “A i icial in elligence and music: Open
ques ions o copy igh law and enginee ing p axis,”
A s, ol. 8, p. 115, 9 2019.
[4] J. Ba ne , “The e hical implica ions o gene a i e au-
dio models: A sys ema ic li e a u e e iew,” AIES 2023
- P oceedings o he 2023 AAAI/ACM Con e ence on
AI, E hics, and Socie y, pp. 146–161, 8 2023.
[5] R. Ba lle-Roca, E. Gómez, W.-H. Liao, X. Se a,
and Y. Mi su uji, “T anspa ency in music-gene a i e
ai: A sys ema ic li e a u e e iew,” Resea ch Squa e
(p ep in ), 12 2024. [Online]. A ailable: h ps:
//www. esea chsqua e.com/a icle/ s-3708077/ 2
[6] K. Ta a , P. E icson, K. Co on, P. T. N. Del P ado,
R. Ba lle-Roca, B. Cab e o-Daniel, S. Ljungblad,
G. Diapoulis, and J. Hussain, “A shi in a is ic p ac-
ices h ough a i icial in elligence,” Leona do, ol. 57,
no. 3, pp. 293–297, 2024.
[7] N. Ca lini, J. Hayes, M. Nas , M. Jagielski, V. Sehwag,
F. T ame , B. Balle, D. Ippoli o, and E. Wallace, “Ex-
ac ing aining da a om di usion models,” in 32nd
USENIX Secu i y Symposium (USENIX Secu i y 23),
2023, pp. 5253–5270.
[8] N. Ca lini, D. Ippoli o, M. Jagielski, K. Lee,
F. T ame , and C. Zhang, “Quan i ying memo iza ion
ac oss neu al language models,” a Xi p ep in
a Xi :2202.07646, 2023. [Online]. A ailable: h ps:
//a xi .o g/abs/2202.07646
[9] D. B alios, G. Wiche n, F. G. Ge main, Z. Pan, S. Khu-
ana, C. Ho i, and J. Le Roux, “Gene a ion o epli-
ca ion: Auscul a ing audio la en di usion models,”
in ICASSP 2024-2024 IEEE In e na ional Con e ence
on Acous ics, Speech and Signal P ocessing (ICASSP).
IEEE, 2024, pp. 1156–1160.
[10] O. E. Gunde sen, Y. Gil, and D. W. Aha, “On
ep oducible ai: Towa ds ep oducible esea ch, open
science, and digi al schola ship in ai publica ions,” AI
Magazine, ol. 39, no. 3, pp. 56–68, 2018. [Online].
A ailable: h ps://doi.o g/10.1609/aimag. 39i3.2816
[11] A. Liesen eld and M. Dingemanse, “Re hinking open
sou ce gene a i e ai: open-washing and he eu ai
ac ,” in P oceedings o he 2024 ACM Con e ence
on Fai ness, Accoun abili y, and T anspa ency, se .
FAccT ’24. New Yo k, NY, USA: Associa ion o
Compu ing Machine y, 2024, p. 1774–1787. [Online].
A ailable: h ps://doi.o g/10.1145/3630106.3659005
[12] A. Spi ling, “Why open-sou ce gene a i e ai models
a e an e hical way o wa d o science,” Na u e, ol.
616, no. 7957, pp. 413–413, 2023.
[13] D. G oene eld, I. Bel agy, P. Walsh, A. Bhagia,
R. Kinney, O. Ta jo d, A. H. Jha, H. I ison,
I. Magnusson, Y. Wang e al., “Olmo: Accele a ing
he science o language models,” a Xi p ep in
a Xi :2402.00838, 2024. [Online]. A ailable: h ps:
//a xi .o g/abs/2402.00838
[14] I. Solaiman, “The g adien o gene a i e ai elease:
Me hods and conside a ions,” in P oceedings o he
2023 ACM Con e ence on Fai ness, Accoun abili y,
and T anspa ency, se . FAccT ’23. New Yo k,
NY, USA: Associa ion o Compu ing Machine y,
2023, p. 111–122. [Online]. A ailable: h ps:
//doi.o g/10.1145/3593013.3593981
P oceedings o he 26 h ISMIR Con e ence, Daejeon, Ko ea, Sep embe 21-25, 2025
733
[15] M. A nold, R. K. E. Bellamy, M. Hind, S. Houde,
S. Meh a, A. Mojsilo i´
c, R. Nai , K. N. Ramamu hy,
A. Ol eanu, D. Pio kowski, D. Reime , J. Richa ds,
J. Tsay, and K. R. Va shney, “Fac Shee s: Inc easing
us in AI se ices h ough supplie ’s decla a ions o
con o mi y,” IBM Jou nal o Resea ch and De elop-
men , ol. 63, no. 4/5, pp. 6:1–6:13, 2019.
[16] T. Geb u, J. Mo gens e n, B. Vecchione, J. W.
Vaughan, H. Wallach, H. D. III, and K. C aw o d,
“Da ashee s o da ase s,” Commun. ACM, ol. 64,
no. 12, p. 86–92, no 2021.
[17] S. Holland, A. Hosny, S. Newman, J. Joseph, and
K. Chmielinski, “The da ase nu i ion label,” Da a
P o ec ion and P i acy, ol. 12, no. 12, p. 1, 2020.
[18] M. Mi chell, S. Wu, A. Zaldi a , P. Ba nes, L. Vasse -
man, B. Hu chinson, E. Spi ze , I. D. Raji, and T. Ge-
b u, “Model ca ds o model epo ing,” in P oceed-
ings o he Con e ence on Fai ness, Accoun abili y,
and T anspa ency, se . FAT* ’19. New Yo k, NY,
USA: Associa ion o Compu ing Machine y, 2019, p.
220–229.
[19] D. R. Desai and M. Riedl, “Be ween copy igh and
compu e science: The law and e hics o gene a i e
ai,” a Xi p ep in a Xi :2403.14653, 2024. [Online].
A ailable: h ps://a xi .o g/abs/2403.14653
[20] Open Sou ce Ini ia i e (OSI), “The Open Sou ce AI
De ini ion – e sion 1.0,” 10 2024, las accessed:
Ma ch 26, 2025. [Online]. A ailable: h ps://opensou
ce.o g/ai/open-sou ce-ai-de ini ion
[21] K. Robison (The Ve ge), “Open-sou ce AI mus e eal
i s aining da a, pe new OSI de ini ion,” 10 2024,
las accessed: Ma ch 23, 2025. [Online]. A ailable:
h ps://www. he e ge.com/2024/10/28/24281820/ope
n-sou ce-ini ia i e-de ini ion-a i icial-in elligence-m
e a-llama
[22] R. Gall (The News S ack), “The Open Sou ce AI
De ini ion: Wha he C i ics Say,” 11 2024, las
accessed: Ma ch 23, 2025. [Online]. A ailable:
h ps:// henews ack.io/ he-open-sou ce-ai-de ini io
n-wha - he-c i ics-say/
[23] A. Ta kowski and P. Kelle , “The Open Sou ce AI
De ini ion is a s ep o wa d in de ining openness in
AI,” 11 2024, las accessed: Ma ch 26, 2025. [Online].
A ailable: h ps://open u u e.eu/blog/ he-open-sou c
e-ai-de ini ion-is-a-s ep- o wa d-in-de ining-opennes
s-in-ai/
[24] H. H. Jiang, L. B own, J. Cheng, M. Khan, A. Gup a,
D. Wo kman, A. Hanna, J. Flowe s, and T. Geb u, “Ai
a and i s impac on a is s,” in P oceedings o he
2023 AAAI/ACM Con e ence on AI, E hics, and Soci-
e y, 2023, pp. 363–374.
[25] Eu opean Pa liamen and Council o he Eu opean
Union, “Regula ion (EU) 2024/1689 o he Eu opean
Pa liamen and o he Council o 13 June 2024
laying down ha monised ules on a i icial in elligence
(A i icial In elligence Ac ),” 2024, las accessed:
Ma ch 23, 2025. [Online]. A ailable: h p://da a.eu op
a.eu/eli/ eg/2024/1689/oj
[26] ——, “A icle 53: Obliga ions o P o ide s o
Gene al-Pu pose AI Models (A i icial In elligence
Ac ),” 2024, las accessed: Ma ch 23, 2025. [Online].
A ailable: h ps://a i icialin elligenceac .eu/a icle/53/
[27] E. A. O ice, “Gene al-Pu pose AI Models in he
AI Ac - Ques ions & Answe s,” 03 2025, las
accessed: Ma ch 26, 2025. [Online]. A ailable:
h ps://digi al-s a egy.ec.eu opa.eu/en/ aqs/gene al-p
u pose-ai-models-ai-ac -ques ions-answe s
[28] EU AI O ice, “Second D a o he Gene al-
Pu pose AI Code o P ac ice,” 12 2024, las
accessed: Ma ch 26, 2025. [Online]. A ailable:
h ps://digi al-s a egy.ec.eu opa.eu/en/lib a y/second
-d a -gene al-pu pose-ai-code-p ac ice-published-w
i en-independen -expe s
[29] A. Liesen eld, A. Lopez, and M. Dingemanse,
“Opening up cha gp : T acking openness, ans-
pa ency, and accoun abili y in ins uc ion- uned ex
gene a o s,” in P oceedings o he 5 h In e na ional
Con e ence on Con e sa ional Use In e aces, se .
CUI ’23. New Yo k, NY, USA: Associa ion o
Compu ing Machine y, 2023. [Online]. A ailable:
h ps://doi.o g/10.1145/3571884.3604316
[30] F. Königs o e and S. Thalmann, “Ai documen a ion:
A pa h o accoun abili y,” Jou nal o Responsible Tech-
nology, ol. 11, p. 100043, 2022.
[31] J. Ma hews, “Pa e ns and an i-pa e ns, p inciples and
pi alls: accoun abili y and anspa ency in ai,” AI
Magazine, ol. 41, no. 1, pp. 82–89, 2020.
[32] J. Walmsley, “A i icial in elligence and he alue o
anspa ency,” AI & socie y, ol. 36, no. 2, pp. 585–
595, 2021.
[33] C. No elli, M. Taddeo, and L. Flo idi, “Accoun abili y
in a i icial in elligence: wha i is and how i wo ks,”
Ai & Socie y, ol. 39, no. 4, pp. 1871–1882, 2024.
[34] R. Bommasani, K. Klyman, S. Longp e, S. Kapoo ,
N. Maslej, B. Xiong, D. Zhang, and P. Liang,
“The ounda ion model anspa ency index,” a Xi
p ep in a Xi :2310.12941, 2023. [Online]. A ailable:
h ps://doi.o g/10.48550/a Xi .2310.12941
[35] M. Whi e, I. Haddad, C. Osbo ne, X.-Y. Liu, A. Ab-
delmonse , and S. Va ghese, “The model openness
amewo k: P omo ing comple eness and openness o
ep oducibili y, anspa ency and usabili y in ai,” a Xi
p ep in a Xi :2403.13784, 2024. [Online]. A ailable:
h ps://doi.o g/10.48550/a Xi .2403.13784
P oceedings o he 26 h ISMIR Con e ence, Daejeon, Ko ea, Sep embe 21-25, 2025
734
[36] F. Ei as, A. Pe o , B. Vidgen, C. Sch oede , F. Pizza i,
K. Elkins, S. Mukhopadhyay, A. Bibi, A. Pu ewal,
C. Bo os, F. S eibel, F. Kesh ka , F. Ba ez, G. Smi h,
G. Guadagni, J. Chun, J. Cabo , J. Impe ial, J. A.
Nolazco, L. Landay, M. Jackson, P. H. S. To ,
T. Da ell, Y. Lee, and J. Foe s e , “Risks and
oppo uni ies o open-sou ce gene a i e ai,” a Xi
p ep in a Xi :2405.08597, 2024. [Online]. A ailable:
h ps://doi.o g/10.48550/a Xi .2405.08597
[37] F. Ei as, A. Pe o , B. Vidgen, C. S. de Wi ,
F. Pizza i, K. Elkins, S. Mukhopadhyay, A. Bibi,
B. Csaba, F. S eibel, F. Ba ez, G. Smi h, G. Guadagni,
J. Chun, J. Cabo , J. M. Impe ial, J. A. Nolazco-Flo es,
L. Landay, M. Jackson, P. Rö ge , P. H. S. To ,
T. Da ell, Y. S. Lee, and J. Foe s e , “Nea o mid-
e m isks and oppo uni ies o open sou ce gene a i e
ai,” a Xi p ep in a Xi :2404.17047, 2024. [Online].
A ailable: h ps://doi.o g/10.48550/a Xi .2404.17047
[38] J. Engel, K. K. Ag awal, S. Chen, I. Gul ajani,
C. Donahue, and A. Robe s, “GANSyn h: Ad e sa ial
neu al audio syn hesis,” in In e na ional Con e ence
on Lea ning Rep esen a ions (ICLR), 2019. [Online].
A ailable: h ps://open e iew.ne / o um?id=H1xQVn
09FX
[39] P. Dha iwal, H. Jun, C. Payne, J. W. Kim, A. Rad o d,
and I. Su ske e , “Jukebox: A gene a i e model
o music,” a Xi p ep in a Xi :2005.00341, 2020.
[Online]. A ailable: h ps://a xi .o g/abs/2005.00341
[40] A. Caillon and P. Esling, “Ra e: A a ia ional
au oencode o as and high-quali y neu al audio
syn hesis,” a Xi p ep in a Xi :2111.05011, 2021.
[Online]. A ailable: h ps://a xi .o g/abs/2111.05011
[41] M. Pasini and J. Schlü e , “Musika! as in ini e wa e-
o m music gene a ion,” in P oceedings o he 23 d
In e na ional Socie y o Music In o ma ion Re ie al
Con e ence (ISMIR), 2022.
[42] F. Schneide , O. Kamal, Z. Jin, and B. Schölkop ,
“Moûsai: Tex - o-music gene a ion wi h long-con ex
la en di usion,” a Xi p ep in a Xi :2301.11757,
2023. [Online]. A ailable: h ps://a xi .o g/abs/2301.1
1757
[43] J. Cope , F. K euk, I. Ga , T. Remez, D. Kan , G. Syn-
nae e, Y. Adi, and A. Dé ossez, “Simple and con ol-
lable music gene a ion,” Ad ances in Neu al In o ma-
ion P ocessing Sys ems (Neu IPS), ol. 36, 2024.
[44] A. Agos inelli, T. I. Denk, Z. Bo sos, J. En-
gel, M. Ve ze i, A. Caillon, Q. Huang,
A. Jansen, A. Robe s, M. Tagliasacchi e al.,
“Musiclm: Gene a ing music om ex ,” a Xi
p ep in a Xi :2301.11325, 2023. [Online]. A ailable:
h ps://a xi .o g/abs/2301.11325
[45] H. F. Ga cia, P. See ha aman, R. Kuma , and B. Pa do,
“Vampne : Music gene a ion ia masked acous ic o-
ken modeling,” 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.
[46] K. Chen, Y. Wu, H. Liu, M. Nezhu ina, T. Be g-
Ki kpa ick, and S. Dubno , “Musicldm: Enhanc-
ing no el y in ex - o-music gene a ion using bea -
synch onous mixup s a egies,” in ICASSP 2024 - 2024
IEEE In e na ional Con e ence on Acous ics, Speech
and Signal P ocessing (ICASSP), 2024, pp. 1206–
1210.
[47] S.-L. Wu, C. Donahue, S. Wa anabe, and N. J. B yan,
“Music con olne : Mul iple ime- a ying con ols o
music gene a ion,” IEEE/ACM T ans. Audio, Speech
and Lang. P oc., ol. 32, p. 2692–2703, May 2024.
[Online]. A ailable: h ps://doi.o g/10.1109/TASLP.20
24.3399026
[48] Q. Huang, D. S. Pa k, T. Wang, T. I. Denk,
A. Ly, N. Chen, Z. Zhang, Z. Zhang, J. Yu,
C. F ank e al., “Noise2music: Tex -condi ioned
music gene a ion wi h di usion models,” a Xi
p ep in a Xi :2302.03917, 2023. [Online]. A ailable:
h ps://a xi .o g/abs/2302.03917
[49] M. W. Y. Lam, Q. Tian, T. Li, Z. Yin, S. Feng, M. Tu,
Y. Ji, R. Xia, M. Ma, X. Song, J. Chen, Y. Wang,
and Y. Wang, “E icien neu al music gene a ion,” in
P oceedings o he 37 h In e na ional Con e ence on
Neu al In o ma ion P ocessing Sys ems, se . NIPS ’23.
Red Hook, NY, USA: Cu an Associa es Inc., 2024.
[50] Z. No ack, J. McAuley, T. Be g-Ki kpa ick, and N. J.
B yan, “DITTO-2: Dis illed di usion in e ence- ime -
op imiza ion o music gene a ion,” 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.
[51] J. Nis al, M. Pasini, C. Aouameu , M. G ach en, and
S. La ne , “Di -a- i : Musical accompanimen co-
c ea ion ia la en di usion models,” 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.
[52] O. Tal, A. Zi , I. Ga , F. K euk, and Y. Adi, “Join
audio and symbolic condi ioning o empo ally con-
olled ex - o-music gene a ion,” 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.
[53] Z. E ans, J. D. Pa ke , C. Ca , Z. Zukowski, J. Tay-
lo , and J. Pons, “S able audio open,” a Xi p ep in
a Xi :2407.14358, 2024.
[54] D. P. Kingma and M. Welling, “Au o-encoding
a ia ional bayes,” in 2nd In e na ional Con e ence on
Lea ning Rep esen a ions, ICLR 2014 - Con e ence
T ack P oceedings. In e na ional Con e ence on
Lea ning Rep esen a ions (ICLR), Dec 2014. [Online].
A ailable: h ps://a xi .o g/abs/1312.6114 10
P oceedings o he 26 h ISMIR Con e ence, Daejeon, Ko ea, Sep embe 21-25, 2025
735
