MELODIC AND METRICAL ELEMENTS OF EXPRESSIVENESS IN
HINDUSTANI VOCAL MUSIC
Yash Bhake Anki Anand P ee i Rao
Depa men o Elec ical Enginee ing, Indian Ins i u e o Technology Bombay, India
[email p o ec ed], [email p o ec ed], [email p o ec ed]
ABSTRACT
This pape p esen s an a emp o s udy he aes he ics o
No h Indian Khayal music wi h e e ence o he lexibili y
exe cised by a is s in pe o ming popula composi ions.
We s udy exp essi e iming and pi ch a ia ions o he
gi en ly ical con en wi hin and ac oss pe o mances and
p opose compu a ional ep esen a ions ha can disc imi-
na e be ween di e en pe o mances o he same song in
e ms o exp ession. We p esen he necessa y audio p o-
cessing and anno a ion p ocedu es, and discuss ou obse -
a ions and insigh s om he analysis o a da ase o wo
songs in wo agas each ende ed by en p ominen a is s.
1. INTRODUCTION
AKhayal pe o mance consis s o a composi ion ( he cho-
sen bandish o song wi h ly ics) ha also o ms he base
o imp o isa ion [1]. Tha is, while he i s occu ence o
abandish line is ypically ende ed in i s canonical o m
om memo y, he accomplished a is ende s successi e
epe i ions o he line wi h pi ch and iming a ia ions in
speci ic ways o c ea e an aes he ic expe ience app ecia ed
by lis ene s ained in he gen e. A mo i a ion o his wo k
is ha sui able compu a ional ep esen a ions can help ob-
ain insigh s o e i y hypo heses ega ding his pe o -
mance p ac ice. They can also in o m ools o gene a e
ealis ic audio o a is ic and educa ional con ex s [2].
Pe o mance exp ession, acco ding o Juslin [3], con-
ce ns " he small and la ge a ia ions in iming, dynamics,
imb e, and pi ch ha o m he mic os uc u e o a pe o -
mance and di e en ia e i om ano he pe o mance o he
same music". While he s uc u e o he songs hemsel es
ca ies speci ic emo ion – his ce ainly holds o bandish
gi en he aga ha unde lies he melody, and he seman-
ics o he ly ics, - exp essi e a ia ions con ibu e o he
emo ional impac o he pe o mance. We p esen a e iew
o some o he wo k o musicologis s who ha e s udied
Khayal music om he iewpoin o composi ion and im-
p o isa ion.
© Y. Bhake, A. Anand, and P. Rao. 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: Y. Bhake, A. Anand, and P. Rao, “Melodic and Me ical El-
emen s o Exp essi eness in Hindus ani Vocal Music ”, 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 a schola ly wo k based on se e al ocal pe o -
mances o a single bandish by di e en a is s, Mo is [4]
used his ansc ip ions o he eco ded music (well awa e
o he ambigui ies inhe en in he manual p ocess ha in-
ol es disc e e symbols) o compa e an a is ’s endi ion
wi h hei sel -decla ed no a ion o he composi ion. He
no ed ha some po ions o bandish we e ep oduced iden-
ically, o he s we e mo e lexibly p oduced, and p oceeded
o in es iga e he ex en and na u e o he di e ences. An-
o he example o such wo k is he analysis o a single li e
pe o mance by an de Mee [5], ocusing on aga- ela ed
ocal pi ch in lec ions in he aga imp o isa ion sec ion
(alap) ha igge ed emo ion as de ec ed om he audi-
ence’s spon aneous audible eac ions. He ela es hese ex-
p essi e momen s o he occu ence/execu ion o speci ic
o namen s. Finally, a close pa allel o ou p oblem is ha
o acking sanga is, which a e lineage-dependen a ia-
ions o he same ly ic lines in Ca na ic composi ions [6].
Mo e ecen ly, he simila ques ions using sligh ly
la ge , cu a ed da a se s ha e been acili a ed by he a ail-
abili y o compu a ional me hods. P e ious compu a ional
s udies ha e examined he a iabili y o aga ph ases in
he cou se o Khayal imp o isa ion wi h espec o un-
de lying empo and me ical loca ion [7]. The con ex o
singing bandish lines, on he o he hand, en ails addi ional
cons ain s gi en he special impo ance acco ded o he
ly ics in he con ex o bandish. Recen ly, a small compa -
a i e s udy o wo pe o mances o he same bandish, wi h
e e ence o empo al de ia ions o no e e en s om he
canonical me ical posi ions [8], se ed as a p elimina y
alida ion o me hodology using au oma ically de ec ed
onse s o he audio-le el and ex -le el alignmen o man-
ual and e e ence syllable-le el ansc ip ion. Thei isual
ep esen a ion o he dis ibu ion, ac oss each pe o mance
audio, o he iming o se co esponding o a speci ic me -
ical posi ion showed he musicologically an icipa ed be-
ha iou , iz. educed de ia ion close o he sam o down-
bea o he hy hm cycle. I also helped isualise he di -
e ence be ween he wo pe o mances in e ms o he use
o iming exp essi eness. In con as , he p esen wo k (i)
uses a signi ican ly la ge da ase o a iming exp essi e-
ness s udy while also adding new a ibu es; (ii) includes
a new s udy o pi ch-based exp essi eness; (iii) p oposes
compu a ional measu es o a is -le el exp ession on each
o he s udied dimensions and alida es hei po en ial o
disc imina ing pe o mances based on exp essi eness.
In he nex sec ion we p o ide some o he music back-
68
g ound needed o app ecia e he speci ic ques ions we se
up o ou wo k and ou o e all app oach. Following his,
we p esen ou da ase , audio and ex p ocessing me hods
and he de ised compu a ional measu es. Finally, we dis-
cuss ou obse a ions and a emp s o d aw insigh s.
2. BACKGROUND
Bandish a e ly ical composi ions ha ha e been handed
down ac oss gene a ions by o al ansmission and se e as
a so o dic iona y o he associa ed aga in e ms o he
o e all pi ch mo emen and cha ac e is ic ph ases [9]. The
wo e ses o he bandish, e med s hayi and an a a, yp-
ically sung ea ly on in any conce , a e de ined by hei
ly ics and une. In he ea ly 20 h cen u y, music schola P .
V.N. Bha khande launched on a mission o collec and no-
a e adi ional composi ions om ac oss he coun y. He
de ised his own ways o no a ing he melodies ha so a
lay wi hin he amewo k o o al ansmission and la gely
wi hin he edi a y musician amilies. The no a ion, consid-
e ed a "schema ic" o m, equi es an unde s anding o he
in icacies o he aga and ala sys em in o de o in e p e
in pe o mance. Now widely espec ed, his monumen al
wo k helped p ese e he adi ional composi ions o pos-
e i y [10]. Al hough commonly ea u ed in lea ning con-
ex s, i is held ha he no a ion co esponding o a gi en
pe o med bandish is dependen o an ex en on he lineage
(gha ana) o he a is . Tex di e ences exis as well, bu
hese a e mainly due o spelling and dialec s o e b o ms,
while main aining he ecognizabili y o he ly ics.
Bha khande’s book [10] ypically p o ides a single no-
a ion sequence o e e y unique line o he song wi h a
aga swa (no e), o a sho sequence o swa , assigned o
e e y syllable o he ly ics. Figu e 1 shows he canonical
no a ion o a bandish, as a ailable
Figu e 1. The composi ion Ja Ja Re om Bha khande [10]
as a machine- eadable CSV ile o he wo e ses o 2 lines
each.
in he book and he c ea ed machine- eadable o ma ,
e aining he no e label (bo h pi ch and ly ic syllable) and
iming in o ma ion as indica ed by he labels o he 16-bea
ala cycle. In his s udy, all bandish a e se in een aal (16-
bea cycle), wi h salien ma a (bea s) like he sam (down-
bea , ’x’) and khali (9 h bea , ’o’, coinciding wi h he s a
o he second hal cycle) explici ly ma ked. The op ow
indica es bea numbe , he ow below i indica es ibhag
symbols, ma king he s a o each qua e cycle. Each line
o ly ics o he bandish spans 16 bea s, wi h each bea con-
aining ei he a syllable (o , seldom, mul iple syllables), a
es (emp y s ing), o a con inua ion o he p e ious no e
(‘s’). Each cycle is ep esen ed using wo aligned ows:
(1) No a ion ow (black), which displays he main no e
(sa gam) wi h any o namen a ion indica ed as a supe sc ip
p eceding he main no e (such as a g ace no e o a glide);
and (2) Ly ics ow ( ed), which con ains he co esponding
bandish ly ic syllables aligned o he bea posi ions and as-
socia ed no es.
In pe o mance, singe s ypically epea each line (and
some imes i s componen ph ases) se e al imes wi h a i-
a ions in oduced a e he i s u e ance is ende ed in
he canonical o m. The successi e epe i ions a e cha ac-
e ised by a ce ain luidi y in he no es and iming o whe e
a gi en syllable alls [4]. These a ia ions a e no mally
ex empo e in conce se ings bu could show simila i ies
ac oss he a is s om he same musical lineage (gha ana).
Viewing hese as exp essi e a ia ions o he bandish line,
as speci ied by i s ly ics, we in es iga e he ange and na-
u e o he exp essi e ges u es as a unc ion o he speci ic
wo d ( ia i s componen syllables) ac oss i s mul iple oc-
cu ences in he singing. Tha is, we hope o ob ain in-
sigh s abou he s uc u al momen s o he song whe e he
exp essi e ges u es a e added and p ecisely which acous ic
pa ame e s a e a ied o ealise his.
3. DATASET AND PROCESSING
Ou da ase 1is designed o se e he objec i es o ou
s udy, namely, o obse e he a ia ions ac oss epea ed u -
e ances o a gi en bandish line ac oss pe o mances o
he same and di e en a is s. We conside wo speci ic
popula composi ions, each wi h mul iple pe o mances
by se e al p ominen a is s in he gen e, ob ained om
ac oss comme cial and ee in e ne sou ces. The sung
bandish lines a e manually segmen ed om ull conce s
(which also ypically include se e al o he sec ions, includ-
ing ee imp o isa ion in he chosen aga). The de ails o
he da ase used in his s udy appea in Table 1. An im-
media e obse a ion is he skewed dis ibu ion ac oss he 4
bandish lines wi h Line 1 being sung a mo e equen ly
han any o he . This is common in he conce se ing,
gi en ha he bandish line 1 is known as he mukhda o
e ain o he song, analogous o he palla i in Ca na ic
s yle o Indian classical music.
We ca y ou ocals sepa a ion on he conce audio seg-
men s o elimina e he accompanying ins umen s. We use
a p e ained model (OpenAI/Whispe ) o speech o ex
con e sion wi h a p omp comp ising he wo ds o he song
ly ics in De anagi i sc ip , oge he wi h expec ed p onun-
cia ion a ia ions. The ob ained wo d sequence is hen
aligned a he phone le el o he audio using o ced align-
men wi h a Kaldi TDNN Hindi speech ained acous ic
model [11]. Nex , he sequence o phones is segmen ed
in o he syllables o he ly ics wo ds. Each esul ing syl-
lable’s onse ins an is hen iden i ied as he ame co e-
1Mo e de ails appea in he supplemen a y ma e ial
P oceedings o he 26 h ISMIR Con e ence, Daejeon, Ko ea, Sep embe 21-25, 2025
69
sponding o he consonan - owel (CV) ansi ion.
The esul ing alignmen s a e checked and manually co -
ec ed o he occasional e o s ha a ise mainly due o
he p esence o long owels in singing, poo enuncia ion a
imes, as well as he occu ence o signi ican pi ch in lec-
ions. We obse e he manually co ec ed onse loca ions-
i.e. he syllable onse s as ealised by he a is - ma ked
a he op o he wa e o m in he example o one ull
ala cycle o he bandish in Figu e 2. To ob ain he bea
loca ions, we anno a e he abla s oke onse s using he
sou ce-sepa a ed accompanimen ile; we manually ma k
he salien bea s (downbea ’x’ sam and 9 h bea ’o’ khali)
in 1 cycle. Assuming a consis en local empo, we di ide
each hal cycle (in e al be ween one downbea and he ad-
jacen h bea in o 8 equal pa s, he eby ob aining all he
es ima ed bea ins an s ac oss he endi ion.
Nex , we ma k he canonical loca ions o he ma ch-
ing syllables, posi ioning he syllables acco ding o he
Bha khande no a ion (as in Figu e 1). The posi ion map-
ping o he ealised syllable onse s o he co esponding
canonical syllable was implemen ed ollowing he me hod
p oposed in p e ious wo k [8]. We obse e om Figu e 2
how he ealised onse s lag he canonical loca ions mos o
he ime.
Figu e 2. Pi ch con ou (bo om) and sung syllable align-
men ( ed) wi h canonical bea posi ions o he same sylla-
ble (black) o an exce p o Ja Ja Re by ABD.
Finally he ocal pi ch is ex ac ed a 10 ms in e als us-
ing an au oco ela ion based me hod o undamen al e-
quency and oicing [12]. B ie pauses and un oiced e-
gions a e linea ly in e pola ed o ob ain a con inuous pi ch
con ou o each sung syllable egion. The pi ch con ou is
con e ed o cen s by no malisa ion wi h he known pe o -
mance’s onic. E en ually, we ob ain o each pe o mance
in ou da ase , he segmen ed audio o each sung line an-
no a ed a he syllable le el wi h syllable name, bounda ies
and he pi ch (cen s) a 10 ms in e als. We use hese low-
le el ea u es o de ine quan i ies ha cap u e he singing
a ia ions ac oss epe i ions o a bandish line wi hin and
ac oss singe s. The e e ence o he compa ison is he
syllable iden i y (i.e. i s name and me ical loca ion) as
de ined in he canonical no a ion as p esen ed in Figu e 1.
Raga Bhimpalasi Yaman
Bandish Ja Ja Re Ye i Aali
Tala Teen aal Teen aal
Swa S, R, g, m, P, D, n, S S, R, G, M, P, D, N, S
# Conce s 15 13
# A is s 15 12
# Repe i ions (L1, L2, L3, L4) 167, 39, 47, 47 94, 32, 35, 23
Ma a pe min ange 138–200 111–203
Table 1. Summa y o ou da ase o conce eco dings
ac oss agas,bandish, and a is s. Swa no a ion de ails
a e in he supplemen a y.
4. MEASURING EXPRESSIVENESS
We wish o quan i y and compa e he a iabili y obse ed
in he acous ic ealisa ion o a gi en syllable, om a spe-
ci ic line o he bandish, ac oss (i) epea ed u e ances
wi hin an a is ’s pe o mance, and (ii) u e ances o he
same syllable ac oss di e en pe o mances/a is s. The
acous ic pa ame e s ha we de ec a e: (i) he onse ime
o he syllable, (ii) he syllable du a ion (as he ime in e -
al be ween he cu en syllable’s onse and ei he he on-
se o he nex syllable o he s a o he ollowing silence
segmen , whiche e occu s i s . and (iii) he pi ch con ou
shape ac oss he syllable in e al. We illus a e he p ocess
by p o iding examples o he p ocessing and analyses o
he audio ende ing o a chosen line by one a is .
4.1 Timing exp ession
The de ia ion o he de ec ed onse om i s e e ence as-
signed bea index in he canonical no a ion gi es us an
es ima e o he lag/lead o he singe o he syllable in
ques ion. We ep esen he de ia ion in e ms o ac ion
o he local bea in e al; his no maliza ion acili a es he
compa ison ac oss ins ances and conce s. We can iew
he hus measu ed iming o se s as e idence o exp essi e
iming, especially i his quan i y shows a iabili y ac oss
epe i ions o he syllable wi hin he conce .
Figu e 3. De ia ion o he sung syllable onse s om he
canonical loca ions measu ed in he uni s o bea du a ion
o Ja Ja Re Line 1 by ABD o mul iple epe i ions o he
line.
Figu e 3 cap u es he onse s o syllables in bandish 1,
Line 1 as ende ed by singe ABD. The syllable names
P oceedings o he 26 h ISMIR Con e ence, Daejeon, Ko ea, Sep embe 21-25, 2025
70
a e shown wi h hei canonical ma a loca ions a he bo -
om. We no e ha some syllables occupy 2 bea s and o h-
e s 1 bea in he canonical o m. We obse e, o exam-
ple, ha "Jaa2" and "Man" (bo h 2-bea syllables) show
a ia ions wi h a mean lag o abou one bea . "Man" in-
s ances howe e a e much mo e dispe sed. While a uni-
o m o se could po en ially indica e a s uc u al di e -
ence be ween he a is ’s e sion o he bandish and ha
o he Bha khande book, a high s anda d de ia ion (like
in "Man") poin s o he a is ’s in- he-momen exp essi e
a ia ions. On he o he hand, he 3 syllables p eceding
"Man" show nea -ze o o se s ac oss epe i ions.
4.2 Pi ch exp ession
Analysing he ende ed song o pi ch-based exp essi e-
ness is a a he in ol ed ask. The e a e many ways in
which he a is injec s exp essi eness in hei pe o mance
ia pi ch a ia ion. As we wan o analyse he di e en
ways o ealising he same line o a bandish, we a e in-
e es ed in he a iabili y in he pi ch con ou (PC) shape
o a syllable ac oss he mul iple epe i ions o ha line. A
g ea e di e si y o PC shapes can hen be in e p e ed as
highe exp essi eness in an a is ’s epe i ions o he line.
Figu e 4. PC quan isa ion o he nea es aga no e o one
ins ance o Ja Ja Re bandish line 1 ende ed by a is ABD
Fo each syllable, he associa ed PC spans he du a ion
o ha syllable; hence, his is no a ixed-leng h ime se ies.
We ep esen his a iable dimension PC o a gi en sylla-
ble wi h a lowe and ixed-dimensional ec o . We i s im-
plemen a piece-wise agg ega e app oxima ion (PAA) o e
he syllable PCs. PAA is a ime-se ies ep esen a ion ha
has been used widely in da a-mining asks [13]. This is
implemen ed as ollows.
The syllable PC alues a e each i s quan ised o he
nea es aga swa (no e), Figu e 4 shows his p ocess.
Nex , he quan ised PC alues o a syllable ac oss epe-
i ions, is agg ega ed by di iding he quan ised PC in o a
ixed numbe o uni o m in e als and assigning he mode
o he alues o each in e al. The numbe o ixed in-
e als is se empi ically o 10 in e als pe bea s allo -
ed o he syllable ( ea ing he syllable ex ensions indi-
ca ed by ’-’ in Figu e 1 o be a pa o he p e ious syl-
lable, he eby adding o i s allo ed bea s). The choice o
10 equal segmen s pe bea in e al is based on he empo
ange o ou da ase (110-200 BPM o 300 ms o 545 ms
Figu e 5. Th ee dis inc endi ions o he syllable "Jaa1"
by a is ABD, each ep esen ed by a ixed numbe o uni-
o m ime in e als. Each in e al is mapped based on i s
modal pi ch o he nea es aga no e, gi ing us he PAA
s ing ep esen a ion o he syllable’s pi ch shape. Fig-
u e 6 desc ibes his p ocess and he PAA s ings o he
abo e PCs.
pe bea ) and sampling pe iod (10 ms/sample) o he pi ch
con ou . This esul s in segmen s, each ep esen ed by a
sho sequence o samples o he pi ch con ou . This bal-
ance allows cap u ing dynamic pi ch luc ua ions which
a e p e alen in Hindus ani classical music (HCM), wi h-
ou o e -quan isa ion. Now, each PAA in e al wi hin a
syllable is assigned a disc e e symbol, whe e he symbols
a e d awn om a sui able alphabe which comp ises no es
(swa ) ac oss he ele an oc a e anges. The esul ing
s ing o no e alues (one pe PAA in e al) hen ep esen s
coa sely he ealised pi ch shape o he syllable. Figu e 6
shows his p ocess. The 3 s ing sequences in Figu e 6 co -
espond o he 3 PCs in Figu e 5. This ype o agg ega ion
p esen s a adeo o gene ali y s speci ici y.
Like in he case o syllable iming, we a e in e es ed
in he a ia ion, i any, in pi ch shape o a gi en syllable
ac oss epe i ions. We achie e his by compu ing he sim-
ila i y o he syllable PCs o pai s d awn om he se o
epe i ions in a single conce . The Le ensh ein edi dis-
ance [14] be ween he PAA s ings p o ides us wi h he
numbe o no e subs i u ions. We e alua e he No malised
Le ensh ein Subs i u ion Sco e (NLSS) o each pai as a
measu e o he dissimila i y. Figu e 7 shows a ma ix ep-
esen a ion (hea map) o NLSS alues o a chosen sylla-
ble as ende ed by one a is ac oss 14 epe i ions.
P oceedings o he 26 h ISMIR Con e ence, Daejeon, Ko ea, Sep embe 21-25, 2025
71
Figu e 6. P ocessing pipeline o gene a ing PAA s ing
ep esen a ions o he pi ch con ou o di e en epe i-
ions o a syllable. The s ings co espond o he PCs o
he syllable epe i ions gi en in Figu e 5 spanning 2 bea s.
Figu e 7. Hea map showing NLSS o each pai o "Jaa1"
syllable PCs d awn om he se o epe i ions o line 1 o
bandish Ja Ja Re ende ed by a is ABD
5. OBSERVATIONS AND DISCUSSION
We a e in e es ed in he wi hin-a is a ia ion o a gi en
bandish line and i s syllables. This can acili a e po en-
ially aluable insigh s abou (i) he p e e ed loca ions (in
e ms o chosen syllables) o exp essi e ges u es by indi-
idual a is s and ac oss a is s and (ii) he ex en and na-
u e o exp essi e ges u es o a gi en a is . Due o space
limi a ions, we p esen he analysis esul s o line 1 o one
bandish, wi h he o he bandish p esen ed in he supple-
men a y.
Figu e 8 p esen s o each a is and syllable, he s an-
da d de ia ion (s.d.) o he iming de ia ion (as cap u ed
in he example o a is ABD in Figu e 3). The s.d. helps
us ocus on he a iabili y o o se s a he han on ac-
ual o se alues (which migh be a ibu ed o s uc u al
di e ences be ween he a is ’s e sion and Bha khande’s
e sion o he bandish, a he han exp ession- ela ed).
In Figu e 8, we no e he dominance o he i s 3 syl-
lables o mos a is s. The ull ange o beha iou s, how-
e e , includes IN a one end wi h minimal a ia ions o DG
and RK, who in oduce new a ia ions on p ac ically all
syllables. Tha IN does no exe cise any lexibili y is no
Figu e 8. S anda d de ia ion o he dis ibu ion o he ac-
ional iming de ia ion o e e y syllable o e mul iple ep-
e i ions in one endi ion, ac oss a is s.
Figu e 9. Mean NLSS o e all pai s o epe i ions o each
syllable in line 1 o bandish Ja Ja Re, compu ed pe a is .
se ing as a dissimila i y measu e ac oss di e en epe i-
ions o a syllable by an a is .
su p ising gi en ha his pe o mances we e explici ly c e-
a ed o closely ollow he p esc ibed Bha khande no a ion,
as discussed he e [15]. RK, on he o he hand, is consid-
e ed a i uoso musician.
Agg ega ing ac oss he ows, we ob ain he pe sylla-
ble beha iou ac oss a is s in Figu e 10. The mean alues
show ha he i s 3 syllables ca y he mos exp essi e im-
ing, wi h "Jaa2" also showing he mos sp ead ac oss a is s
(consis en wi h Figu e 8). We see, o example, ha PT
shows a la ge ange in pe syllable SD, again ag eeing wi h
Figu e 8. In he case o he a is ABD, we can obse e ha
he empo al de ia ion is sp ead ela i ely e enly, while
peaking o a pa icula syllable "Man", which is he down-
bea . This can be easily app ecia ed in lis ening o he au-
dio, which can be accessed in he supplemen a y ma e ial.
To assess pi ch a iabili y, we calcula e he a e age
numbe o pi ch subs i u ions by aking he mean o he
NLSS o all pai s o epe i ions pe a is and pe syllable,
(no malised by he s ing leng h), calcula ed ac oss all pos-
sible pai s o he gi en syllable u e ances wi hin a conce .
Figu e 9 shows he alues pe a is and pe syllable o Line
1. We can obse e in Figu e 8 and Figu e 9 ha bo h he
pi ch and empo al a ia ion ac oss epe i ions and ac oss
di e en a is s is mo e p ominen a he beginning o he
line. Nea he end o he line, he pi ch a ia ion inc eases,
which can be a ibu ed o he emphasis on a seman ically
P oceedings o he 26 h ISMIR Con e ence, Daejeon, Ko ea, Sep embe 21-25, 2025
72
Figu e 10. Box-plo o s.d. o iming de ia ion pa am-
e e pe syllable o Ja Ja Re Line 1 as agg ega ed ac oss
conce s and a is s. The ala-cycle ends a he syllable
"Ne", and a new cycle s a s om downbea a he syllable
"Man".
Figu e 11. Box-plo o he mean o he a e aged NLSS pe
syllable o Ja Ja Re Line 1 as agg ega ed ac oss conce s
and a is s.
impo an wo d- "Mandi a a".
The syllables "Pa" and "Ne" a e nea he ala-cycle
bounda y, and we obse e a minimum pi ch and empo al
a ia ion; his is consis en wi h he obse a ions o [8]. As
he pe o me heads owa ds he ala-cycle bounda y, hei
o e all endency o a ia ion and imp o isa ion dec eases,
aiming a he owa ds esol ing he melodic and empo al
exp ession hey ha e come up wi h in he pa icula ala-
cycle.
Agg ega ing ac oss ows o Figu e 9, we ob ain he pe
syllable beha iou o e all a is s in Figu e 11. The do ed
line joining he means indica es ha he pi ch a ia ion de-
c eases as one app oaches he ala-cycle bounda y a he
syllable "Ne", while he he pi ch exp essi eness is highe
in he s a and end o he line, which alls on he 7 h and
5 h bea s o he ala-cycle espec i ely, which is a om
he cycle bounda y, hence p o iding mo e scope o iming
and pi ch exp essi eness. We see, o example, ha AC
and AK show high pi ch a ia ion ac oss he epe i ions o
he same line.
Hie a chical clus e ing o all he pai s om he se o
epe i ions o a syllable by an a is p o ides us wi h in o -
ma ion abou he a ia ion clus e s. A h eshold can be de-
ined ha decides i a a ia ion belongs o a clus e o no .
Mo e di e se a ia ions would indica e mo e numbe o
clus e s, indica ing highe exp essi eness. Dend og ams
a e excellen o isualising such clus e s. The ealisa ion
o he p e ious syllable has an in luence on which clus e
he ollowing syllable a ia ion would belong o.
Compa ing Figu e 8 and Figu e 9, we no e ha exp es-
si e ges u es ha u ilise pi ch a e no necessa ily a he
same loca ions ha exhibi iming de ia ion in e ms o
he p e e ed syllable. I is a he in e es ing o look a he
leas amoun o exp essi eness in bo h pi ch and iming lie
wi h he syllables ha a e nea he ala-cycle bounda y. An
analysis a he indi idual audio le el would p o ide a mo e
accu a e pic u e o he co ela ions, i any, and is le o
u u e wo k.
Ou obse a ions, epo ed he e on he Line 1 o one
bandish, la gely hold wi h he second bandish. The unde -
lying easons o he choice o speci ic syllables exhibi ing
la ge a iabili y a e simila o hose discussed by Mo -
is [4]. These include la ge a ia ions a line o ph ase
ending syllables due o he e ec o p e ious and nex con-
ex s, and he choice o syllables belonging o mo e emo-
ionally loaded wo ds in he ly ics.
6. CONCLUSION
In his pape , we a icula ed he p oblem o modeling ex-
p essi e a ia ions in he con ex o pe o mance o Hin-
dus ani adi ional composi ions by es ablished a is s o
he gen e. A well-known bandish in he chosen aga is
always sung a he beginning o a conce wi h mul iple
epe i ions o he lines, ma ked by a ia ions in he loca-
ion and ype o he exp essi e ges u es. Based on ou
p oposed me hodology, we showed ha i is possible o
a i e a sys ema ic pa e ns ac oss a is s by ea ing he
syllables o he ly ics as e e ence poin s o a s udy o he
ange o a ia ion. This also helped us discuss in e es ing
co ela ions be ween he oles o melody and hy hm in ex-
p essi eness.
We p esen ed a da ase ha was anno a ed wi h a combi-
na ion o manual and au oma ic ools o ob ain a ich epos-
i o y o dis inc ealiza ions o he lines o wo popula
adi ional composi ions. While much u he explo a ion
emains possible, his wo k demons a es he po en ial o
compu a ional models o imp o isa ion in he con ex o
composi ions in he Khayal gen e. This wo k lays he
ounda ion o gene a i e applica ions by cap u ing high-
le el pe o mance ea u es ha e lec an a is ’s dis inc i e
s yle. A p elimina y expe imen was pe o med o gene a e
he empo al de ia ions discussed abo e, whe e he dis i-
bu ions o med by all he de ia ions o a syllable om i s
canonical loca ion o an a is we e used o sample ou
new poin s o each syllable. A sine- one based audio was
syn hesized om he gene a ed pi ch con ou . The e e -
ence (Bha khande canonical o m) and gene a ed acks a e
a ailable in he supplemen a y. This le s us c ea e in ini e
possibili ies o ende ing he same line, while a he same
ime cap u ing some hin o he a is ’s s yle. Ex ending
his app oach o o he acous ic dimensions o exp ession
such as imb e and dynamics can enable he gene a ion o
classical music ha embodies he unique iden i y o indi-
idual a is s.
P oceedings o he 26 h ISMIR Con e ence, Daejeon, Ko ea, Sep embe 21-25, 2025
73
7. ACKNOWLEDGMENTS
We ake his oppo uni y o acknowledge and exp ess ou
g a i ude o all hose who suppo ed and guided us du ing
his esea ch wo k. We a e hank ul o Madhumi ha S.,
whose hesis wo k was a c ucial base o ou p ojec . We
also hank M . Himanshu Sa i, and M s. Hemala Ranade
o hei schola ly musicological insigh s, which ga e di-
ec ion o ou wo k.
8. REFERENCES
[1] B. Wade, “Music in India: The classical adi ions,”
Manoha P ess, 2001.
[2] C. E. Cancino-Chacón, M. G ach en, W. Goebl, and
G. Widme , “Compu a ional models o exp essi e mu-
sic pe o mance: A comp ehensi e and c i ical e-
iew,” F on ie s in Digi al Humani ies, ol. 5, p. 25,
2018.
[3] P. N. Juslin, “Cue u iliza ion in communica ion o emo-
ion in music pe o mance: Rela ing pe o mance o
pe cep ion.” Jou nal o Expe imen al Psychology: Hu-
man pe cep ion and pe o mance, ol. 26, no. 6, p.
1797, 2000.
[4] A. Mo is, T ansmission and pe o mance o Khayal
composi ions in he Gwalio gha ana o Indian ocal
music. PhD hesis, Uni . O London, S.O.A.S., 2004.
[5] W. Van de Mee , “Audience esponse and exp essi e
pi ch in lec ions in a li e eco ding o legenda y singe
kesa bai ke ka ,” in Exp essi eness in music pe o -
mance: Empi ical app oaches ac oss s yles and cul-
u es, D. Fabian, R. Timme s, and E. Schube , Eds.
Ox o d Uni e si y P ess (UK), 2014, pp. 170–184.
[6] S. Sanka an, P. V. K. Sekha , and A. M. Hema, “Au-
oma ic segmen a ion o composi ion in ca na ic music
using ime- equency c cc empla es,” in P oceedings
o 11 h In e na ional Symposium on Compu e Music
Mul idisciplina y Resea ch (CMMR), 2015.
[7] K. K. Ganguli and P. Rao, “A s udy o a iabili y in
aga mo i s in pe o mance con ex s,” Jou nal o New
Music Resea ch, ol. 50, no. 1, pp. 102–116, 2021.
[8] Y. Bhake and P. Rao, “Exp essi e iming in Hindus-
ani ocal music,” in P oc. o ICASSP 2025 Wo kshop
on Indian Music Analysis and Gene a i e Applica ions
(WIMAGA), Hyde abad, India, 2025, accessed a :link.
[9] S. Rao and P. Rao, “An o e iew o Hindus ani music
in he con ex o compu a ional musicology.” Jou nal
o New Music Resea ch, ol. 43, no. 1, 2014.
[10] V. Bha khande, K amik Pus aka Malika. Sangee
Ka yalaya Ha h as, India, 2013.
[11] D. Po ey, A. Ghoshal, G. Boulianne, L. Bu ge ,
O. Glembek, N. Goel, M. Hannemann, P. Mo licek,
Y. Qian, P. Schwa z e al., “The kaldi speech ecog-
ni ion oolki ,” in IEEE 2011 wo kshop on au oma ic
speech ecogni ion and unde s anding. IEEE Signal
P ocessing Socie y, 2011.
[12] Y. Jadoul, B. Thompson, and B. De Boe , “In oducing
pa selmou h: A py hon in e ace o p aa ,” Jou nal o
Phone ics, ol. 71, pp. 1–15, 2018.
[13] J. Lin, E. Keogh, L. Wei, and S. Lona di, “Expe ienc-
ing SAX: a no el symbolic ep esen a ion o ime se-
ies.” Da a Mining and knowledge disco e y, ol. 31,
no. B, pp. 107–144, Ap il 2007.
[14] W. J. Hee inga, “Measu ing dialec p onuncia ion di -
e ences using Le ensh ein dis ance,” 2004.
[15] “2000 classic composi ions om Bha khande on cd,”
h ps://sc oll.in/a icle/726180/, accessed: 2024-04-10.
P oceedings o he 26 h ISMIR Con e ence, Daejeon, Ko ea, Sep embe 21-25, 2025
74
