Int
ern
at
i
onal
Journ
al of Ele
ctrical
an
d
Co
mput
er
En
gin
eeri
ng
(IJ
E
C
E)
Vo
l.
10
,
No.
4
,
A
ugus
t
2020
,
pp.
35
08
~
35
18
IS
S
N:
20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v
10
i
4
.
pp
35
08
-
35
1
8
3508
Journ
al h
om
e
page
:
http:
//
ij
ece.i
aesc
or
e.c
om/i
nd
ex
.ph
p/IJ
ECE
Sp
ea
ke
r specifi
c featur
e bas
ed clus
ter
in
g an
d its ap
pli
ca
ti
ons in
languag
e in
depend
ent forensic
spe
aker re
cogniti
on
Sa
t
yan
an
d
Si
ng
h
1
,
Pr
agya
Singh
2
1
School
of El
ec
t
ric
a
l
and
E
lectr
o
nic
s E
ng
ine
er
ing
,
Fij
i
Nat
ional
U
nive
rsit
y
,
Fij
i
Isl
and
2
School
of
Publ
i
c
Hea
lt
h
and
Pri
m
ar
y
C
are,
Fij
i N
at
ional
Unive
r
sit
y
,
Fi
ji
Isl
and
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Dec
5,
2018
Re
vised
Dec
18
,
2019
Accepte
d
Ja
n
11, 2
020
Forensic
spea
k
e
r
rec
ogn
it
ion
(FS
R)
is
the
pro
c
ess
of
determ
ining
whethe
r
the
source
of
a
questi
oned
voi
c
e
rec
ord
ing
(trac
e)
is
of
a
spec
ific
indi
vidu
al
(sus
pec
te
d
spe
ake
r).
Mos
t
e
xisti
ng
m
et
hods
m
ea
sure
inter
-
uttera
n
ce
sim
il
ari
ties
dir
e
ct
l
y
b
ase
d
on
s
pec
tr
um
-
base
d
c
har
acte
rist
ic
s,
th
e
resulting
cl
usters
m
a
y
not
be
wel
l
re
lated
to
spea
k
er’
s,
but
rat
he
r
to
d
iffe
r
e
nt
ac
oust
i
c
cl
asses.
Thi
s
re
sea
rch
addr
esse
s
thi
s
def
i
cienc
y
b
y
project
ing
la
ngu
age
-
inde
pend
ent
ut
teranc
es
in
to
a
ref
ere
nc
e
spac
e
eq
uippe
d
to
cov
er
the
stand
ard
voic
e
f
ea
tu
res
u
nder
l
y
ing
the
en
ti
re
u
tt
er
ance
set
.
The
n
a
c
luste
ri
ng
appr
oach
is
proposed
ba
sed
on
the
pe
ak
appr
oximat
i
on
in
orde
r
to
m
axi
m
iz
e
the
sim
il
ari
t
ie
s
bet
wee
n
l
anguage
-
inde
p
ende
nt
utt
er
anc
es
within
al
l
cl
ust
ers.
Thi
s
m
et
hod
us
es
a
K
-
m
edoi
d,
Fuzz
y
C
-
m
ea
ns,
Gus
ta
fson
and
Kess
el
and
Gath
-
Geva
al
gor
it
hm
to
eva
lu
ate
the
c
luste
r
to
w
hic
h
each
utt
er
a
nce
should
be
alloc
at
ed
,
over
coming
the
disadva
nt
age
of
tra
di
ti
ona
l
hie
rar
chi
c
a
l
cl
uster
ing
tha
t
t
he
ult
imate
out
c
om
e
ca
n
onl
y
h
it
the
opti
m
um
rec
ogni
ti
on
eff
iciency
.
The
rec
ognition
ef
fic
i
ency
of
K
-
m
edoi
d,
Fuzz
y
C
-
m
ea
ns,
Gus
ta
fson
and
Kess
el
and
Gath
-
Geva
cl
uster
i
ng
al
gorit
hm
s
are
95.
2%
,
97.
3%,
98
.
5%
a
nd
99.
7%
and
E
ER
ar
e
3
.
62%,
2.
91
%,
2
.
82%
,
and
2.
61%
respe
ctively
.
Th
e
EE
R
improve
m
ent
of
the
Gath
-
Geva
techni
que
base
d
FS
Rsy
stem
com
par
ed
wi
th
Gus
ta
fson
and
Kess
el
and
Fuzz
y
C
-
m
ea
ns
is
8.
04%
and
11
.
49
%
respe
ct
iv
ely
.
Ke
yw
or
d
s
:
Alte
rn
at
ive
du
nn in
dex (
A
DI)
Cl
assifi
cat
ion
entr
op
y
(CE)
Dun
n
in
de
x (DI)
Fu
zzy
m
axi
m
um
li
kelihood
(F
ML)
Partit
ion
c
oe
ff
i
ci
ent (
PC
)
Partit
ion
i
nd
e
x (SC)
Sepa
rati
on in
de
x(S)
Xie an
d
Be
ni i
nd
e
x (
XB)
Copyright
©
202
0
Instit
ut
e
o
f Ad
vanc
ed
Engi
n
ee
r
ing
and
S
cienc
e
.
Al
l
rights re
serv
ed
.
Corres
pond
in
g
Aut
h
or
:
Saty
anand S
i
ngh
,
School
of Elec
tric
al
an
d El
ect
ronics E
nginee
rin
g,
Fiji N
at
io
nal
U
niv
e
rsity
, F
iji
I
sla
nd.
Em
a
il
:
sat
ya
na
nd.sin
gh@
fnu.a
c.f
j
1.
INTROD
U
CTION
Sp
ea
ker
r
ec
ogniti
on
is
t
he
ge
ner
al
te
rm
us
e
d
to
i
nclu
de
al
l
the
m
any
diff
e
ren
t
ta
s
ks
o
f
di
scrim
inati
on
base
d
on
t
he
s
ound
of
t
heir
voic
es
betwee
n
on
e
pe
rs
on
an
d
an
oth
e
r
[
1].
F
or
e
ns
ic
s
m
eans
the
us
e
of
sci
e
nce
or
te
chnolo
gy in
inv
e
sti
gating
a
nd f
in
ding in
th
e court o
f
la
w f
act
s o
r
evi
den
c
e. Th
e r
ole of for
e
ns
ic
science
is to
pro
vid
e
i
nform
at
ion
(in
fac
t
or
opini
on)
t
o
as
sist
i
nv
est
igators
an
d
la
w
c
ourts
in
a
ns
we
rin
g
ques
ti
on
s
of
i
m
po
rtance
.
F
or
e
ns
ic
sp
ea
ke
r
recog
niti
on
i
s
the
m
et
ho
d
of
determ
ining
wh
et
he
r
the
ori
gin
of
a
ques
ti
on
ed
vo
ic
e
r
eco
rd
i
ng
(trace
)
is
a
par
ti
cular
person
(s
us
pe
ct
ed
sp
eake
r).
T
his
process
in
volves
c
om
par
ing
a
n
un
i
den
ti
fie
d
voic
e
rec
ordin
g
(questi
one
d
recordi
ng)
with
on
e
or
m
or
e
rec
ordin
gs
of
a
kn
own
voic
e
(the
al
le
ge
d
s
pe
aker
'
s
voic
e)
[1
]
.
Foren
sic
Au
t
om
atic
Sp
e
aker
Re
c
ogniti
on
(FASR
)
is
an
est
ablis
hed
te
rm
us
e
d
in
the
a
da
ptati
on
of
a
ut
om
atic
sp
eake
r
rec
ogniti
on
m
et
ho
ds
t
o
f
orensic
a
pp
li
cat
ion
s
.
F
or
a
utom
at
ed
sp
ea
ker
ide
ntific
at
ion
,
t
he
determ
inist
ic
or
pr
e
dicti
ve
m
od
el
s
of
th
e
vo
ic
e
of
th
e
sp
eake
r'
s
acoust
ic
char
act
e
risti
cs are c
on
tra
ste
d wit
h
the
aco
us
t
ic
ch
aracte
risti
cs of the
r
ec
ordin
gs
for q
uesti
on
[
1].
The
cl
us
te
rin
g
of
s
peaker
’s
ref
er
s
to
the
functi
on
of
gro
up
i
ng
to
ge
ther
unid
entifi
ed
sp
eec
h
expressi
on
s
ba
sed
on
t
he
vo
ic
e
c
har
act
e
risti
cs
of
a
s
peak
e
r.
T
he
c
on
ce
r
ns
an
d
needs
of
the
sp
ea
ker
recog
niti
on
c
om
m
un
it
y
hav
e
bee
n
a
m
ajo
r
m
ot
ivati
on
for
th
e
re
searc
h
on
s
peak
e
r
cl
ust
ering
f
or
m
ore
than
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N:
20
88
-
8708
Sp
e
aker
sp
eci
fi
c feature
base
d cl
us
te
rin
g and
it
s app
li
catio
ns i
n
l
anguage i
nd
e
pe
nd
e
nt …
(
Sa
ty
an
and Sin
gh
)
3509
a
deca
de
[2,
3],
in
w
hich
t
he
m
ai
n
pur
po
se
is
to
band
t
og
et
her
sp
eec
h
dat
a
ge
ner
at
e
d
by
the
sam
e
sp
ea
ker
or
sp
ea
ker’s
with
si
m
i
la
r
vo
ic
es
so
that
the
adap
ta
ti
on
of
aco
us
ti
c
m
od
el
s
can
be
m
or
e
effe
ct
ively
carri
ed
out.
Be
cause
sp
eec
h
cl
us
te
rin
g
sim
pl
y
serv
es
as
a
su
pp
le
m
entary
process
in
s
peech
recog
niti
on
,
howe
ver,
there
is
sti
ll
a
dear
th
of
stud
ie
s
ded
ic
at
ed
to
this
subj
ect
.
M
or
e
re
centl
y,
sp
eake
r
-
cl
ust
erin
g
w
ork
has
e
xperi
enc
e
d
a
ren
ai
ssa
nce
[
4],
pow
ere
d
by
re
searc
h
into
sp
oke
n
do
c
um
ent
ind
e
xing
to
ha
nd
le
bur
geonin
g
c
ollec
ti
on
s
of
acce
ssible
voic
e
data.
T
he
m
a
in
pur
pose
of
s
uch
a
n
em
erg
ing
resea
rch
t
opic
is
that
the
hu
m
an
effo
rt
ne
eded
for do
c
um
entat
ion
ca
n be
dra
m
at
ic
ally reduced
by gr
ouping s
peec
h data f
ro
m
t
he
sam
e sp
eake
r’s.
Sp
ea
ker
cl
us
te
rin
g
can
be
de
scribe
d
as
a
n
unsupe
rv
ise
d
sp
ea
ker
-
rec
og
niti
on
pro
ble
m
in
wh
ic
h
the
sp
ea
ker
rec
ogniti
on
proce
ss
[5
]
is
c
on
ce
r
ned
with
deter
m
ining
a
s
peaker
rec
ogniti
on
or
w
hether
a
s
peak
e
r
is
who
he/s
he
cl
aim
s.
Con
tra
ry
to
t
he
tra
diti
on
al
s
pea
ker
-
r
ecognit
ion
ap
proac
h,
howe
ve
r,
wh
ic
h
ass
ume
s
that
so
m
e
con
te
xt
ua
l
inform
at
ion
or
s
peec
h
detai
ls
is
acce
ssible
an
d
can
be
m
od
el
e
d
on
t
he
sp
ea
ker’s
c
on
c
ern
e
d,
sp
ea
ker
cl
us
te
r
ing
will
f
unct
ion
with
out
a
ny
awa
ren
e
ss
of
who
t
he
pote
nt
ia
l
sp
eake
r’s
a
r
e
an
d
how
m
a
ny
are
involve
d
in
the
la
ngua
ge
t
o
be
cl
us
te
re
d.
Sol
ution
s
to
t
he
s
peak
e
r
-
cl
us
te
ri
ng
issue
s
houl
d
the
refo
re
be
able
to
extract
an
d
co
m
par
e
the
sp
eech
cha
racteri
sti
cs
that
un
de
rlie
the
utterance
colle
ct
io
ns
in
an
unat
te
nde
d
m
ann
er
.
C
on
t
r
ary
to
the
tradit
ion
al
sp
ea
ke
r
-
recog
niti
on
appro
ac
h,
howev
e
r,
wh
ic
h
assum
es
that
so
m
e
con
te
xtu
al
i
nfor
m
at
ion
or
s
peech
detai
ls
is
acce
ssible
a
nd
ca
n
be
m
od
el
ed
on
t
he
sp
ea
ker’s
c
onc
ern
e
d,
sp
ea
ker
cl
us
te
r
ing
will
f
unct
ion
with
out
a
ny
awa
ren
e
ss
of
who
t
h
e
pote
nt
ia
l
sp
eake
r’s
a
r
e
an
d
how
m
a
ny
are
involve
d
in
t
he
la
ngua
ge
to
be
cl
us
te
re
d.
S
ol
ution
s
t
o
the
s
peak
e
r
-
cl
us
te
ri
ng
iss
ue
s
houl
d,
t
her
e
fore,
be
able
to
extract
a
nd
com
par
e
the
sp
eec
h
cha
ract
erist
ic
s
that
unde
rlie
the
utt
eran
ce
c
ollec
ti
on
s
in
a
n
unat
te
nd
e
d
m
ann
er
.
A
sim
il
ar
act
ivit
y
is
a
segm
entat
ion
of
sp
ea
ke
r’
s
[
6],
w
hich
see
ks
to
ide
ntify
th
e
boun
dar
ie
s
wh
e
n
a
sp
ea
ker
c
ha
ng
e
occ
ur
s
i
n
an
au
di
o
str
ea
m
con
ta
inin
g
the
s
peec
h
expressi
on
s
of
m
ulti
ple
peo
ple
.
To
gether
with
sp
eake
r
cl
us
t
ers,
the
se
gm
e
ntati
o
n
of
s
pe
aker’s
brea
ks
the
co
ntinuo
us
input
into
discrete
sta
tem
ents
that
are
easy
to
process
in
sp
e
ech/
sp
ea
ker
r
e
cogniti
on
a
nd
is,
t
her
e
fore,
an
esse
ntial
ste
p
i
n
th
e
ind
e
xing
of
spo
ken
docu
m
ents.
Sp
ea
ke
r
segm
entat
ion
can
be
acc
om
pl
ished
f
r
om
ano
t
her
a
ng
le
with
the
help
of
spe
aker
cl
us
te
rin
g.
T
her
e
m
ay
be
a
sh
ift
in
sp
eake
r’s
betw
een
two
a
dj
ac
e
nt
short
re
gions
wi
t
h
diff
e
re
nt
cl
us
te
r
in
dices.
Most
sp
ea
ker
-
cl
us
te
rin
g
m
eth
ods
c
urren
tl
y
fo
ll
ow
a
hiera
r
chical
cl
us
te
rin
g
fr
am
ework,
con
sist
in
g
of
three
m
ai
n
co
m
po
n
ents:
c
om
pu
ti
ng
inter
-
utterance
sim
i
l
arit
ie
s,
ge
ner
at
ing
a
cl
us
te
r
t
ree
a
nd
dete
r
m
ining
the
num
ber
of
cl
us
te
rs.
T
he
si
m
il
arity
equ
at
ion
was
de
sign
e
d
to
ge
ne
rate
hi
gh
e
r
va
lues
f
or
sim
i
la
riti
es
betwee
n
the
sam
e
sp
eaker'
s
utterances
a
nd
lo
wer
values
f
or
sim
i
l
arit
ie
s
between
dif
fer
e
nt
spe
aker
'
s
utterances
.
Ma
ny
si
m
il
arity
tests,
su
c
h
as
the
Ku
ll
bac
k
Lei
bl
er
(K
L
)
dista
nc
e
[6
-
8],
the
cr
os
s
pr
ob
a
bili
ty
rati
o
(CLR)
[
8],
an
d
the
ge
ner
al
i
zed
pro
ba
bili
t
y
rati
o
(G
LR)
are
analy
zed
and
co
ntrast
e
d
in
m
any
wo
r
ks
of
li
te
ra
ture.
A
cl
us
te
r
t
ree
is
c
r
eat
ed
ei
ther
in
a
bott
om
-
up
(
agg
l
om
erati
ve)
or
a
t
op
-
dow
n
(d
i
visive)
fa
sh
io
n,
accor
ding
to
s
om
e
crit
eria
der
ive
d
f
ro
m
the
m
easur
e
of
s
i
m
i
la
rity
.
The
bo
tt
om
-
up
m
eth
od
sta
rts
a
s
a
sing
le
cl
us
te
r
with
ea
ch
uttera
nce
and
the
n
m
erg
es
t
he
cl
us
te
rs
in
a
fair
m
ann
er
unti
l
al
l
the
utterances
are
f
ound
i
n
on
e
cl
us
te
r
.
Ne
ver
t
heless,
al
l
the
uttera
nces
s
ta
rt
in
a
sin
gle
cl
us
te
r
in
t
he
top
-
down
m
et
h
od,
an
d
t
he
cl
ust
er
s
are
sepa
rated
s
uccessively
un
ti
l
each
cl
us
te
r
has
ex
act
ly
one
utterance
.
T
he
resu
lt
in
g
cl
ust
er
tree
is
the
n
sp
li
t
to
m
a
intai
n
the
best
pa
rtit
ion
by
est
im
ating
the
nu
m
ber
of
cl
us
te
rs.
Re
pr
ese
ntati
ve
m
et
hods
are
bas
ed
on
the BB
N
Me
tri
c an
d
the
Bay
esi
an
I
nfo
rm
at
io
n
Crit
eri
on to
e
stim
at
e the o
pt
i
m
u
m
n
um
ber
of clusters
[8,
9].
In
a
ddit
ion
t
o
de
velo
ping
a
m
or
e
accurat
e
m
easur
e
m
ent
of
i
nter
-
utterance
sim
il
arity,
we
al
so
inv
est
igate
ho
w
the
cl
us
te
rs
can
be
opti
m
a
l
ly
pr
oduce
d
so
that
al
l
the
utt
eran
ces
within
the
cl
us
te
r
are
fr
om
the
sam
e
sp
eaker
.
Co
nventio
nal
m
et
ho
ds
ba
sed
on
ei
t
her
t
op
-
do
wn
or
bott
om
-
up
hiera
rch
ic
al
cl
us
te
ring
use
a
near
est
nei
ghbor
hood
sel
ect
ion
r
ule
to
deci
de
w
hich
pr
onouncem
ents
sh
ou
l
d
be
assig
ne
d
to
the
sam
e
cl
ass.
Howe
ver,
the
cl
os
est
nei
ghbo
rho
od
sel
ect
ion
r
ule
is
ap
plied
in
a
cl
us
te
r
-
by
-
c
luster
m
ann
er
duri
ng
t
he
pr
ocedur
e
of
sp
li
tt
ing
on
e
cl
us
te
r
or
m
e
rg
i
ng
t
wo
cl
ust
ers,
r
at
her
tha
n
in
a
global
m
ann
er
that
c
onside
rs
al
l
the
cl
us
te
rs.
Con
s
eq
ue
ntly
,
hiera
rch
ic
al
cl
us
te
rin
g
ca
n
only
m
ake
each
ind
ivi
du
al
cl
ust
er
as
hom
og
e
neous
as
possi
ble,
but
the
ulti
m
a
te
goal
of
m
axim
iz
ing
overall
hom
og
eneit
y
can
no
t
be
ac
hieve
d.
To
s
olv
e
this
pro
blem
,
we
a
re
pro
po
si
ng
a
ne
w
cl
us
te
rin
g
a
ppr
oach
s
pe
ci
fical
ly
aim
ed
at
m
axi
m
izing
the
t
otal
num
ber
of
sta
tem
ents
fro
m
the
sam
e
sp
eaker
’s
within
the
cl
ust
er.
T
his
is
ac
hieve
d
by
est
im
ating
the
s
o
-
cal
le
d
cl
us
te
r
purity
in
com
bin
at
io
n
with
a
gen
et
ic
al
gorithm
-
based
optim
iz
at
ion
process
[
10]
to
f
ind
the
best
ut
te
ran
ce
par
ti
ti
on
i
ng to a
chieve m
axim
um
cluster pur
i
ty
.
2.
FUND
AM
E
N
TAL OF
SPE
AKER SPE
CI
FIC FE
ATU
R
E B
AS
ED
C
L
US
TE
RI
NG
ANALY
SIS
The
go
al
of
c
luster
a
naly
sis
is
to
cat
eg
or
i
ze
obj
ect
s
bas
ed
on
sim
il
ari
ti
es
between
them
and
to
orga
nize
data
i
nto
gro
ups.
A
m
on
g
the
uns
uper
vised
ap
pro
aches,
cl
us
te
ri
ng
te
ch
niques
do
not
us
e
pr
i
or
cl
ass
identifie
rs
.
Dif
fer
e
nt
cl
assifi
cat
ion
s
m
ay
be
associat
ed
wit
h
th
e
cl
us
te
rin
g
te
c
hn
i
qu
e
s
al
gorithm
ic
appro
ac
h.
It
is
po
ssible
to
disti
nguish
par
ti
ti
on
in
g,
hierar
c
hical
,
grap
h
-
t
heoreti
cal
m
e
tho
ds
an
d
m
et
ho
ds
bas
ed
on
the ob
j
ect
ive
f
un
ct
io
n.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
4
,
A
ugus
t
2020
:
3
5
08
-
3
5
1
8
3510
2.1.
Speaker sp
eci
fic tr
aining
d
ata
It
is
possi
ble
to
a
pp
ly
cl
ust
ering
te
ch
ni
qu
e
s
to
inf
or
m
at
ion
that
i
s
qua
ntit
at
ive
(
nu
m
erical
),
qu
al
it
at
ive
(cat
egorical
),
or
a
m
ixtur
e
of
bot
h.
T
he
cl
us
te
ri
ng
of
quantit
at
ive
data
is
bei
ng
ap
plied
in
la
ngua
ge
ind
e
pende
nt
F
SRsy
stem
.
Spec
ific
data
f
r
om
the
s
peaker
are
ty
pical
ly
ph
ysi
cal
proces
s
obse
rvat
ion
s.
Each
obser
vation
of
t
he
s
pea
ker
co
ns
ist
s
of
m
easur
ed
var
i
ables
gro
up
e
d
into
a
n
-
dim
ension
al
li
ne
vec
to
r
=
[
1
,
2
,
…
…
.
.
,
]
,
∈
.
X
denotes
a
set
of
N
obs
erv
at
io
ns
a
nd
is
rep
r
esente
d
as
a
m
at
rix
of NXn:
=
[
11
12
…
…
1
21
22
…
…
2
…
…
…
…
…
…
1
2
…
…
]
(1)
In
the
la
ngua
ge
of
patte
r
n
re
cogniti
on,
X
r
ow
s
a
re
cal
le
d
patte
rn
s
or
ob
je
ct
s,
the
columns
are
cal
le
d
the
cha
racteri
st
ic
s
or
at
tri
bu
te
s,
an
d
X
is
cal
le
d
the
patte
r
n
m
at
rix.
X
is
oft
en
re
ferred
t
o
as
the
data
m
atr
ix
in
this
stud
y.
T
he
sign
ific
a
nce
of
X'
s
colu
m
ns
and
r
ows
in
relat
ion
to
reali
ty
depends
on
the
con
te
xt.
Fo
r
exam
ple,
t
he
r
ows
of
X
m
ay
rep
rese
nt
a
sp
ea
ker
in
a
p
plica
ti
on,
a
nd
the
c
olu
m
ns
are
s
peak
e
r
s
pe
ci
fic
m
easur
em
ents.
As
cl
us
te
rin
g
is
app
li
ed
to
dy
nam
ic
syst
e
m
m
od
el
ing
a
nd
cl
assifi
cat
ion
,
the
X
rows
c
on
ta
in
tim
e
sign
al
s
m
easur
em
ents,
an
d
the
c
olum
ns
are,
f
or
exam
ple,
sp
ea
ker
sp
eci
fic
va
riables
ob
s
er
ved
i
n
the
syst
em
(se
nti
m
ent,
em
oti
on,
et
hnic
it
y,
et
c.).
T
he
pur
pose
of
cl
us
te
ri
ng
in
la
ngua
ge
i
nd
e
pe
nd
a
nt
f
or
ensic
app
li
cat
io
n
is
to
fi
nd
relat
io
ns
hi
ps
betwee
n
la
ng
uag
e
in
dep
e
ndent
syst
e
m
var
ia
bles,
cal
le
d
re
gr
ess
ors,
a
nd
sp
ea
ker
sp
eci
fi
c featu
re
de
penden
t
va
riable
s
values
, call
ed r
egr
es
san
ds
[
11]
.
2.2.
Clust
eri
n
g alg
orithm
s in F
S
R
ap
pli
ca
tio
n
Ba
sed
on
the
pur
po
se
of
cl
ust
ering,
va
rio
us
con
ce
pts
of
a
cl
us
te
r
can
be
form
ulate
d.
In
ge
ner
al
,
the
vie
w
that
a
cl
us
te
r
is
a
gr
oup
of
ob
j
ect
s
that
are
m
or
e
s
i
m
i
la
r
to
eac
h
oth
e
r
tha
n
m
em
ber
s
of
oth
e
r
cl
us
te
rs
can
be
em
br
ac
ed.
I
n
s
om
e
w
el
l
-
def
in
ed
se
nse
,
the
te
rm
si
m
il
arity
sh
ou
l
d
be
underst
ood
as
m
at
he
m
at
ic
al
si
m
il
arity.
The
si
m
il
arity
is
of
te
n
descr
i
bed
in
m
e
tric
sp
ac
es
by
m
eans
of
a
distance
sta
nd
a
r
d.
Dista
nc
e
to
so
m
e
cl
us
te
r
prot
otypica
l
obje
ct
can
be
m
ea
su
re
d
betwee
n
the
data
vecto
r
s
them
sel
ves,
or
as
a
dista
nc
e
form
a
data
vect
or.
The
prototype
s
are
us
ually
not
know
n
in
a
dv
a
nce
a
nd
ar
e
searc
hed
sim
ultaneo
us
ly
wi
th
data
par
ti
ti
on
i
ng
by
the
cl
us
te
ri
ng
al
gorithm
s.
The
m
od
el
s
m
a
y
be
vect
or
s
of
the
sam
e
size
as
the
data
obj
ect
s
,
bu
t
they
m
ay
a
lso
be
descr
i
be
d
as
ge
om
et
ric
al
“high
e
r
-
le
ve
l”
obj
ect
s,
s
uc
h
as
li
near
or
non
-
li
near
s
ub
sp
aces
or fun
ct
io
ns.
2.2.1
.
Appl
ic
ati
on o
f
K
-
me
ans a
nd K
-
me
d
oid
algorithm
s
The
m
et
ho
ds
of
ha
rd
pa
rt
it
i
on
i
ng
are
sim
ple
an
d
popul
ar,
al
th
ough
th
ei
r
res
ults
are
no
t
al
ways
reli
able
and
t
he
se
al
gorithm
s
al
so
ha
ve
num
erical
prob
le
m
s.
K
-
m
eans
an
d
K
-
m
edo
id
al
gorithm
s
al
loc
at
e
each
data
point
to
one
of
the
c
cl
ust
ers
from
a
NX
n
dim
ensional
data
set
to
m
ini
m
iz
e
the
su
m
of
sq
ua
res
within
the cluster:
∑
∑
‖
−
‖
2
∈
=
1
(2)
w
he
re
is
a
set
of
ob
j
ect
s
i
n
t
he
I
cl
us
te
r
(d
at
a
points
)
a
nd
i
s
the
m
ean
ove
r
c
l
us
te
r
I
f
or
t
hat
po
i
nt
in
(2)
in
fact
de
no
te
s
a
c
distance
sta
nd
a
rd.
Th
e
cl
us
te
r
prot
otypes
are
cal
le
d
in
K
-
m
ea
ns
cl
us
te
ri
ng
,
i.e
.
the cluster
ce
nters:
=
∑
=
1
,
∈
(3)
w
he
re
is
a
n
e
nt
it
y
nu
m
ber
in
.
T
he
cl
us
te
r
c
enters
are
the
near
est
ob
j
ect
s
to
the
m
ean
of
in
f
or
m
at
ion
i
n
on
e
cluster
=
{
∈
|
1
≤
≤
}
in
K
-
m
edo
id
clus
te
ring.
2.2.2
Ap
pli
ca
t
ion
of f
uz
z
y
C
-
mea
ns
algorit
hm
The
cl
ust
erin
g
al
gorithm
fo
r
Fu
zzy
C
-
m
eans
is
based
on
m
ini
m
izing
a
n
ob
j
ect
ive
functi
on
cal
le
d
functi
onal
C
-
m
eans.
Du
nn d
e
f
ines it
as:
(
;
,
)
=
∑
∑
(
)
‖
−
‖
2
=
1
=
1
(4)
w
he
re
=
[
1
,
2
,
…
.
.
,
]
,
∈
is
a
cl
us
te
r
m
od
el
vecto
r
(cen
te
r
s)
t
o
be
est
ablis
he
d
an
d
2
=
‖
−
‖
2
=
(
−
)
(
−
)
is a s
qu
a
re
d
in
ner
-
pr
oduct
distance n
orm
.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N:
20
88
-
8708
Sp
e
aker
sp
eci
fi
c feature
base
d cl
us
te
rin
g and
it
s app
li
catio
ns i
n
l
anguage i
nd
e
pe
nd
e
nt …
(
Sa
ty
an
and Sin
gh
)
3511
Stat
ist
ic
ally,
as
s
how
n
i
n
(
4)
i
s
t
he
m
axim
um
var
ia
nce
of
f
r
om
can
be
see
n
a
s
a
n
ind
ic
at
or.
Mi
ni
m
iz
ing
the
functi
on
al
c
-
m
eans
from
as
sh
ow
n
in
(
3),
and
(4)
is
a
no
nlinear
pro
blem
of
optim
iz
ation
that
can
be
s
olve
d
by
usi
ng
a
va
r
ie
ty
of
m
et
ho
ds
avail
able,
ra
ng
i
ng
f
ro
m
groupe
d
c
o
ordin
at
ion
m
ini
m
iz
a
ti
on
t
o
si
m
ulate
d
an
ne
al
ing
to
ge
ne
ti
c
al
go
rithm
s.
Ne
ve
rtheles
s,
the
m
os
t
popula
r
m
et
hod
is
a
sim
ple
Pica
rd
it
erati
on
,
known
a
s
the
fu
z
zy
c
-
m
eans
(FC
M)
al
gorith
m
,
thro
ugh
th
e
first
-
orde
r
c
onditi
ons
f
or
s
ta
ti
on
ary
po
i
nts
of
(
4)
.
Using
La
g
ra
nge
m
ulti
pliers,
the
sta
ti
on
ary
po
i
nts
of
t
he
obj
ect
ive
f
un
ct
ion
of
(
4)
c
an
be
i
den
ti
fied
b
y a
pp
ly
in
g
the
li
m
it
∑
=
1
=
1
,
1
≤
≤
to J:
̅
(
;
,
,
)
=
∑
∑
(
)
2
+
∑
(
∑
=
1
−
1
)
=
1
=
1
=
1
(5)
Figure
1
s
hows
the
resu
lt
s
of
K
-
m
edo
i
d al
go
rithm
o
f
f
our s
peak
e
r’s
of
6 s
ec v
oice da
ta
with
norm
al
iz
a
ti
on
.
Figure
1. Re
su
l
t of K
-
m
edo
id
al
gorithm
o
f
f
our
sp
ea
ke
r’
s
of
6
s
ec
vo
ic
e
da
ta
w
it
h norm
al
i
zat
ion
More
ov
e
r
by
s
et
ti
ng
the
U;
V
and
zer
o
gradi
ents o
f
(J
).
I
f
2
≥
0
,
>
1
,
then
(
,
)
∈
m
ay
m
ini
m
iz
e
(3) on
ly
if
;
=
1
∑
(
⁄
)
2
(
−
1
)
⁄
=
1
1
≤
≤
,
1
≤
≤
(6)
a
nd
=
∑
=
1
∑
=
1
,
1
≤
≤
(7)
t
he
rem
ai
nin
g
co
ns
trai
nts
∈
[
0
,
1
]
,
1
≤
≤
,
1
≤
≤
an
d
0
<
∑
<
,
1
≤
≤
=
1
are
al
so
fu
lfil
le
d
b
y
this
so
luti
on.
N
ote
that
(7)
gi
ves
vi
as
the
wei
ghte
d
m
ean
of
t
he
data
it
em
s
belonging
to
a
c
luster
wh
e
re t
he wei
ghts are
the
de
grees o
f
m
e
m
ber
sh
ip
.
The
FCM
al
gorithm
is
ba
sed
on
the
s
ta
nd
a
rd
E
uclidean
dista
nce
sta
ndar
d,
w
hich
i
nduce
s
hype
rsphe
ric
c
lusters.
The
refor
e
,
cl
us
te
rs
w
it
h
the
sam
e
form
and
ori
enta
ti
on
ca
n
on
ly
be
identifie
d,
be
cause
the
ty
pical
c
ho
ic
e
of
norm
in
du
ci
ng
m
at
rix
is:
=
or
it
ca
n
be
ch
os
e
n
as
a
diag
on
al
m
at
ri
x
w
hich
accounts
for di
ff
e
ren
t
var
ia
nc
es in t
he direct
ion o
f
t
he X c
oor
din
at
e a
xes:
=
[
(
1
1
)
2
0
…
0
0
(
1
2
)
2
…
0
…
…
…
…
…
…
…
…
…
…
…
…
0
0
…
(
1
)
2
]
(8)
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
4
,
A
ugus
t
2020
:
3
5
08
-
3
5
1
8
3512
o
r
A
can
be
de
scribe
d
as
the
inv
e
rse
of
m
at
rix
of
co
va
riance:
=
−
1
,
with
=
1
∑
(
−
̅
)
=
1
(
−
̅
)
.
Her
e
̅
de
no
te
s
the
data
sam
pl
e
m
ean.
A
in
duces
the
Ma
hal
anobis
sta
ndar
d
on
in
this
ca
se.
Fig
ur
e
2
sh
ows
the
res
ul
ts of
Fu
zzy
C
-
m
eans alg
or
it
hm
o
f
fou
r
s
pea
ker’s
of 6 sec
voic
e d
at
a
w
it
h
norm
al
iz
a
ti
on
.
Figure
2. Re
su
l
t of f
uzzy C
-
m
eans al
gorithm
of fo
ur sp
ea
ke
r’
s
of
6
sec
voi
ce data
with
no
rm
alizat
ion
2.2.3
.
Appl
ic
ati
on o
f gus
tafs
on
and kes
sel
algorithm
By
us
ing
a
n
a
dap
ti
ve
distanc
e
m
et
ho
d,
Gus
ta
fson
an
d
Ke
s
sel
exp
a
nd
e
d
t
he
sta
ndar
d
f
uz
zy
c
-
m
eans
al
gor
it
hm
to
de
te
ct
clusters
of
var
i
ou
s
g
e
om
etr
ic
sh
a
pes
in o
ne
a
rr
ay
of d
at
a
[
12
].
2
=
(
−
)
(
−
)
,
1
≤
≤
,
1
≤
≤
(9)
The
m
at
rices
are
us
ed
in
the
c
-
m
eans
f
un
ct
ion
al
as
opti
m
i
zat
ion
var
ia
ble
s,
al
lo
wing
the
cl
us
te
r
to
adap
t
the
dist
a
nce
norm
to
t
he
data'
s
local
top
ol
og
ic
al
struct
ur
e.
Let
A
denote
a
c
-
tup
le
of
m
at
rice
s
that
induce
norm
:
=
(
1
,
2
,
2
,
…
)
.
The G
K
al
gorit
hm'
s f
un
ct
io
na
l objecti
ve
is de
fine
d by:
(
;
,
,
)
=
∑
∑
(
)
2
=
1
=
1
(10)
c
onditi
ons
∈
[
0
,
1
]
,
1
≤
≤
,
1
≤
≤
,
∑
=
1
,
1
≤
=
1
≤
and
0
<
∑
<
,
1
≤
≤
=
1
m
ay
be
app
li
ed
dir
ect
ly
to
a
fixe
d
.
The
ob
je
ct
ive
functi
on
(9)
in
relat
io
n
to
howe
ver,
c
annot
be
direct
l
y
m
ini
m
iz
ed
as
it
is
li
near
in
.
This
m
eans
that
by
si
m
ply
m
a
king
le
ss
op
ti
m
ist
ic
,
can
be
rendere
d
as
l
ow
as
desi
red.
m
us
t
be
c
onstrai
ne
d
i
n
s
om
e
way
in
order
to
fi
nd
a
feasible
s
olu
ti
on.
The
usual
way
to
do
t
his
is
to
li
m
it
determ
inant.
Allo
wing
t
he
m
at
rix
to
diff
e
r
wi
th
it
s
def
i
ned
determ
inant
m
eans
opti
m
iz
ing
the
sh
a
pe
of
th
e
cl
us
te
r
w
hile
sta
yi
ng
co
ns
ta
nt
in
volum
e
‖
‖
=
,
>
0
.
Wh
e
re
fo
r
e
ve
ry
cl
us
te
r
is
set
.
The
fo
ll
owin
g
expressi
on
f
or
is
ob
ta
i
ned
usi
ng
t
he
La
gr
a
nge
m
ulti
plier
m
et
ho
d
as
=
[
(
)
]
1
−
1
.
Wh
e
re
is t
he
cl
us
te
r'
s f
uzzy
cov
a
riance
m
atr
ix
def
i
ned b
y:
=
∑
(
)
(
−
)
=
1
(
−
)
∑
(
)
=
1
(11)
Rem
e
m
ber
th
at
re
placi
ng
=
[
(
)
]
1
−
1
and
(11)
with
(
9)
giv
es
a
gen
e
rali
zed
s
quare
Ma
halan
ob
is
distance
norm
betwee
n
xk
and
the
m
ean
cl
us
te
r
wh
e
re
the
co
va
rianc
e
is
weig
hted
by
the
m
e
m
ber
sh
ip
degrees.
Fig
ur
e
3
sho
ws
t
he
res
ults
of
t
he
G
us
ta
fs
on
a
nd
Kessel
cl
us
te
r
ing
al
gorithm
of
f
our
sp
ea
ker’s
of 6
sec v
oice
data
with
norm
al
iz
a
ti
on
.
2.2.4. Ap
pli
ca
tion o
f gath
-
ge
va a
l
go
ri
th
m
The
cl
us
te
ri
ng
al
go
rithm
fu
z
zy
m
axi
m
u
m
li
kelihood
(
F
ML)
us
es
a
di
sta
nce
sta
nd
a
rd
based
on
the fuzzy
m
axim
u
m
li
kelihood
Esti
m
at
es
su
gg
e
ste
d by Be
z
dek an
d D
unn [13
]:
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N:
20
88
-
8708
Sp
e
aker
sp
eci
fi
c feature
base
d cl
us
te
rin
g and
it
s app
li
catio
ns i
n
l
anguage i
nd
e
pe
nd
e
nt …
(
Sa
ty
an
and Sin
gh
)
3513
(
,
)
=
√
(
)
(
1
2
(
−
(
)
)
−
1
(
−
(
)
)
)
(12)
n
otice
that
t
hi
s
distance
no
r
m
includes
an
ex
pone
ntial
con
ce
pt,
c
ontr
ary
to
t
he
G
ust
afson
a
nd
K
essel
al
gorithm
,
and
therefor
e
dec
r
eases
faster
tha
n
the
inter
nal
pro
duct
sta
nd
a
r
d.
F
wi
de
note
s
the
fuzzy
cova
riance
m
at
rix
of it
h
cl
us
te
r, gi
ven b
y
:
=
∑
(
)
=
1
(
−
)
(
−
)
∑
(
)
=
1
,
1
≤
≤
(13)
w
he
re
w=
1
is
us
ed
in
the
or
i
gin
al
FML
al
go
rithm
,
but
the
w=2
we
igh
ti
ng
ex
pone
nt
is
us
ed
to
m
ake
the
pa
rtit
ion
m
or
e
fu
zzy
to
com
pen
sat
e
f
or
t
he
e
xposur
e
of
the
distan
c
e
sta
nd
a
r
d.
T
he
diff
e
rence
betwe
e
n
the
m
at
rix
F
i
in
G
us
ta
fson
a
nd
Kess
el
al
gorithm
and
t
he
F
wi
m
entioned
a
bove
is
t
ha
t
the
la
tt
er
does
no
t
include
t
he
we
igh
ti
ng
ex
pone
nt
m
,
bu
t
co
ns
i
sts
of
w=
1
inst
ead.
T
his
is
be
cause
the
tw
o
weig
hted
m
at
ri
ces
of
cov
a
rian
ce
de
r
ive
from
two
diff
e
re
nt
co
nc
epts
as
ge
ne
r
al
iz
at
ion
s
of
t
he
cl
assic
al
cov
a
riance
.
Th
e
α
i
is
the prio
r pro
ba
bili
ty
o
f
c
hoos
i
ng cluster
i
, t
ha
t i
s g
ive
n by:
=
1
∑
=
1
(14)
Accor
din
g
t
o
t
he
sp
ea
ke
r
dat
a
point
the
m
e
m
ber
sh
ip
de
grees
are
inte
rpr
et
ed
as
the
pos
te
ri
or
pro
bab
il
it
y
of
sel
ect
ing
the
cl
us
te
r.
Gath
a
nd
Ge
va
[
12
]
sta
te
d
that
cl
ust
ering
al
gorith
m
s
wer
e
capa
bl
e
of
detect
ing
cl
us
t
ers
of
va
ryi
ng
sh
a
pes,
siz
es
an
d
den
sit
ie
s
from
the
fuz
zy
m
axi
m
u
m
li
kelihood
e
sti
m
at
es.
The
cl
us
te
r
co
var
ia
nce
m
at
ri
x
is
us
e
d
i
n
co
m
bin
at
ion
with
an
“ex
pone
nt
ia
l”
le
ng
th,
a
nd
the
re
is
no
volum
e
const
raint
on
the
cl
us
te
r
s.
This
al
go
rith
m
,
ho
we
ver,
i
s
le
ss
r
obus
t
in
the
s
ense
that
it
nee
ds
a
good
init
ia
li
zation
be
cause
it
conv
erg
es
to
a
near
-
local
optim
um
du
e
to
the
expo
nen
ti
al
dis
ta
nce
norm
.
Figure
4
sh
ows
the
res
ul
ts of
t
he Gust
afson
-
Kessel
clusterin
g
al
gorithm
b
y t
hr
ee s
pe
aker’s
of 6 se
c voice
d
at
a
.
Figure
3. The
re
su
lt
s of t
he gust
afson
a
nd
kes
sel
cl
us
te
rin
g
al
gorithm
o
f
f
our s
peak
e
r’s
of
6 s
ec v
oice
data with
no
rm
al
iz
at
ion
Figure
4. The
re
su
lt
s of t
he gust
afson
-
kessel
cl
us
te
rin
g
al
gorithm
b
y t
hr
ee
sp
ea
ker’s
of 6
sec
vo
ic
e
data
2.2.5. Clus
ter
va
li
d
at
i
on
Fo
r
eac
h
pa
rtit
ion
,
the
validit
y
fu
nctio
n
pro
vid
es
cl
us
te
r
va
li
dity
m
easurem
ent
s.
If
the
nu
m
ber
of
cl
us
te
rs
is
unknow
n
a
pr
io
ri,
i
t
is
us
ef
ul.
The
optim
al
par
ti
tio
n
can
be
dete
r
m
ined
de
pe
nd
i
ng
on
t
he
nu
m
ber
of
cl
us
te
rs o
n
the
extrem
e p
oin
t
of
t
he vali
datio
n
in
de
xes.
C
om
pu
te
d
ind
e
xe
s inclu
de:
Pa
rtit
ion
C
oeffici
en
t (PC)
,
Cl
assifi
cat
ion
Entr
op
y
(CE
),
Partit
ion
I
nd
e
x
(S
C
),
Se
parat
ion
I
nd
e
x(S)
,
Xie
an
d
Be
ni
In
de
x
(
XB)
,
Dun
n
Inde
x
(
D
I)
a
nd
A
lt
er
nativ
e
D
unn I
nd
e
x (AD
I)
.
-
Partit
ion
C
oe
ff
ic
ie
nt
(
PC):
the
am
ou
nt
of
“
overla
p”
betwee
n
cl
us
te
rs
is
com
pu
te
d
by
(
)
=
1
∑
∑
(
)
2
=
1
=
1
.
-
Cl
assifi
cat
ion
Entr
op
y
(CE
):
it
on
ly
m
eas
ur
es
the
cl
us
t
er
par
ti
ti
on
'
s
f
uzzyness
,
w
hich
is
si
m
i
la
r
t
o
the p
a
rtit
ion
c
oe
ff
ic
ie
nt
(
)
=
−
1
∑
∑
log
(
)
=
1
=
1
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
4
,
A
ugus
t
2020
:
3
5
08
-
3
5
1
8
3514
-
Partit
ion
I
nde
x
(S
C
):
is
the
com
pactness
m
axi
m
u
m
ratio
an
d
cl
us
te
r
separ
at
io
n
rati
o.
I
t
is
a
total
of
un
i
form
ind
ivid
ual
cl
us
te
r
validit
y
te
s
ts
by
div
iding
eac
h
cl
ust
er'
s
fu
zzy
card
i
nalit
y
[1
4],
S
(
)
=
∑
∑
(
)
‖
−
‖
2
=
1
∑
‖
−
‖
2
=
1
=
1
.
Wh
e
n
c
om
par
in
g
diff
e
re
nt
par
ti
ti
on
s
wit
h
the
sam
e
nu
m
ber
of
cl
us
te
rs
,
SC is us
ef
ul.
A
low
e
r SC
v
al
ue
shows a
bette
r partit
ion
.
-
Sepa
rati
on
I
nd
ex
(S
):
unli
ke
par
ti
ti
on
in
dex
(S
C
),
t
he
se
pa
rati
on
in
dex
us
es
a
par
ti
ti
on
va
li
dity
m
ini
m
u
m
range se
pa
rati
on
[14],
(
)
=
∑
∑
(
)
2
=
1
=
1
‖
−
‖
2
,
‖
−
‖
2
.
-
Xie
an
d
Be
ni'
s
Inde
x
(
XB
):
a
i
m
s
at
qu
antify
ing
the
rati
o
of
total
cl
us
te
r
va
riat
ion
a
nd
cl
us
te
r
se
par
at
i
on,
(
)
∑
∑
(
)
2
=
1
=
1
‖
−
‖
2
,
‖
−
‖
2
. T
he op
ti
m
u
m
num
ber
of clu
ste
rs
s
hould m
i
nim
iz
e the in
de
x value.
-
Dun
n'
s
Inde
x
(DI)
:
it
was
ori
gin
al
ly
s
ugge
ste
d
th
at
this
inde
x
be
use
d
to
cl
assify
"
com
pact
and
well
separ
at
e
d
cl
us
t
ers.
"
It
is
there
fore
im
po
rtant
to
recalc
ulate
the
outc
om
e
of
the
cl
us
te
rin
g
as
it
was
a
hard
par
ti
ti
on alg
or
i
thm
,
(
)
=
∈
{
∈
,
≠
{
∈
,
(
,
)
∈
{
,
∈
(
,
)
}
}
}
.
-
Alte
rn
at
ive
D
unn
Inde
x
(
A
DI):
the
pu
rpose
of
m
od
ify
ing
the
ori
gina
l
Dunn
i
nd
e
x
was
t
o
sim
plify
the
cal
cul
at
ion
w
he
n
th
e
dif
fer
e
nc
e
bet
ween
two
cl
us
te
rs
w
orks
,
(
)
=
∈
{
∈
,
≠
{
∈
,
∈
|
(
,
)
−
(
−
)
|
∈
{
,
∈
(
,
)
}
}
}
.
No
te
,
the
onl
y
dif
fer
e
nce
be
tween
SC,
S
a
nd
XB
is
the
cl
us
te
r
sepa
rati
on
process
.
D
ue
t
o
re
-
pa
rtit
ion
i
ng
t
he
re
su
lt
s
w
it
h
the
hard
par
ti
ti
on
proce
ss,
the
values
of
DI
a
nd
ADI
ar
e
no
t
co
ns
ist
en
t
in
the
case
of
ov
e
rlap
ped
cl
us
te
rs
.
Fig
ur
e
5
show
s
the
re
su
lt
s
of
t
he
G
us
ta
f
s
on
-
K
essel
cl
ust
ering
al
gorith
m
by
three
s
pe
aker’s
of
6
s
ec
voic
e
data.
Table
1
il
lustr
at
e
Gu
sta
fson an
d Kessel al
go
rith
m
b
ased num
erical
v
al
ues o
f v
al
idit
y
m
easur
es of clu
ste
rin
g.
Figure
5. Val
ue
s of
dunn'
s in
dex an
d
t
he
al
t
ern
at
ive
du
nn ind
e
x o
f 6 sec
voic
e d
at
a
Table
1.
G
us
ta
fson
a
nd
kessel
al
gorithm
b
as
ed
nu
m
erical
val
ues of
validit
y
m
easur
es
of
cl
us
te
rin
g
c
1
2
3
4
5
6
7
8
9
PC
0
.87
2
8
0
.79
8
2
0
.85
5
5
0
.75
7
3
0
.76
6
6
0
.73
4
1
0
.66
9
3
0
.68
7
9
0
.68
3
7
CE
0
.22
1
9
0
.33
0
7
0
.26
4
6
0
.41
3
9
0
.40
7
8
0
.48
5
3
0
.64
4
1
0
.57
7
8
0
.63
5
5
SC
3
5
.09
0
9
0
.27
8
6
0
.32
3
7
0
.16
1
2
0
.27
7
7
0
.19
5
3
0
.14
0
5
0
.22
9
9
0
.21
2
1
S
0
.26
3
8
0
.00
3
5
0
.00
3
7
0
.00
2
1
0
.00
3
2
0
.00
2
1
0
.00
1
4
0
.00
2
3
0
.00
2
2
XB
1
0
.09
6
6
8
.49
0
1
4
.56
9
0
6
.70
5
9
6
.80
8
1
1
1
.72
2
9
5
.62
5
7
5
.75
8
4
5
.78
6
8
DI
0
.01
1
0
0
.00
5
2
0
.02
7
7
0
.00
4
3
0
.00
6
6
0
.00
5
6
0
.00
5
5
0
.00
2
1
0
.00
0
3
ADI
0
.13
9
3
0
.09
7
5
0
.01
5
4
0
.01
8
3
0
.00
0
7
0
.00
6
3
0
.00
0
0
0
.00
0
2
0
.00
0
8
3.
KERN
EL
-
BA
SED
SPE
A
KER
S
PECIFI
C
FE
ATU
RE
E
X
T
R
AC
TI
ON
AND
T
I
TS
APPLI
CA
TI
ON
Cl
assifi
cat
ion
al
gorithm
s
m
u
st
represent
th
e
obj
ect
s
to
be
cl
assifi
ed
as
points
in
a
m
ulti
di
m
ension
a
l
featur
e
sp
ace
.
Howe
ver,
on
e
can
a
pp
ly
ot
her
vecto
r
s
pa
ce
trans
form
a
ti
on
s
to
the
i
niti
al
featur
es
befo
re
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N:
20
88
-
8708
Sp
e
aker
sp
eci
fi
c feature
base
d cl
us
te
rin
g and
it
s app
li
catio
ns i
n
l
anguage i
nd
e
pe
nd
e
nt …
(
Sa
ty
an
and Sin
gh
)
3515
run
ning
the
le
a
rn
i
ng
al
gorith
m
.
Ther
e
are
t
wo
reas
ons
f
or
doin
g
this.
Fir
st,
they
can
im
pro
ve
the
pe
rfo
rm
ance
of
cl
assifi
cat
io
n
an
d
sec
ond,
they
can
re
du
c
e
the
data'
s
dim
ension
al
it
y.
The
sel
ect
ion
of
init
ia
l
featu
res
a
nd
their
tran
sf
or
m
at
ion
are
s
om
et
i
m
es
dealt
wit
h
in
the
li
te
rat
ur
e
unde
r
the
t
it
le
"feat
ur
e
ex
tract
ion
”.
To
a
vo
i
d
m
isun
de
rstan
di
ng,
this
sect
io
n
desc
ribes
only
the
la
tt
er
a
nd
descr
i
bes
t
he
first
feat
ur
e
set
.
Hope
fu
ll
y
it
w
il
l
be
m
or
e eff
ect
ive
and cla
ssific
at
ion wil
l be
faste
r.
T
he
a
ppr
oac
h
to t
he
e
xtract
ion
of f
eat
ures
m
ay
b
e eit
her
l
inear
o
r
no
nlinear
,
bu
t
there
is
a
te
chn
iq
ue
that
br
eak
s
dow
n
the
barrier
bet
ween
the
tw
o
form
s
in
so
m
e
way.
The
key
idea
beh
i
nd
the
kernel
te
chn
iq
ue
was
ori
gin
al
ly
pr
ese
nted
in
[1
5]
and
ap
plied
again
in
co
nnect
io
n
with the
g
e
ne
r
al
p
ur
po
se
S
V
M
[
16
,
17
,
18
]
fo
ll
owe
d by ot
her ke
rn
el
-
base
d
m
et
ho
ds.
3.1.
Supp
l
ying inp
ut
vari
ab
le
in
f
orm
at
io
n in
to
kernel
P
CA
A
ddit
ion
al
in
f
or
m
at
ion
to
t
he
KP
C
A
re
pr
ese
ntati
on
f
or
inte
r
pret
abi
li
ty
.
W
e
hav
e
dev
el
ope
d
a
process
to
pro
j
ect
a
gi
ven
input
var
ia
ble
into
a
subs
pa
ce
sp
an
ne
d
by
featur
e
vecto
r
s
̃
=
∑
̃
̃
(
1
)
=
1
.
We
ca
n
thi
nk
of
our
obser
vat
ion
as
a
ra
ndom
vector
=
(
1
,
2
,
…
.
.
,
)
im
ple
m
entat
ion
t
he
n
t
o
re
pr
ese
nt
the
pr
om
inence
of
the
in
pu
t
var
ia
ble
in
th
e
KP
C
A.
Co
nsi
der
in
g
a
set
of
points
of
m
at
hem
atical
fo
rm
s
=
+
∈
ℝ
w
her
e
=
(
0
,
…
.
,
1
,
…
.
,
0
)
of
kth
com
po
ne
nt
is
e
it
her
0
or
1.
N
ext,
the
pr
oj
ect
ion
points
(
)
of
these
im
ages
on
t
o
t
he
subspa
ce
s
pa
nne
d
by
t
he
feat
ure
vect
or
̃
=
∑
̃
̃
(
1
)
=
1
ca
n
be
cal
culat
ed
.
Con
si
der
i
ng (1
2) the
row vect
or g
i
ves
the
in
du
ct
io
n
c
urve
in
the
Eige
n
s
pa
ce ex
pr
es
sed
in
m
a
trix for
m
:
(
)
1
=
(
−
1
1
)
(
−
1
1
1
)
̃
(15)
f
urt
he
rm
or
e,
by
proj
ect
in
g
th
e
ta
ng
e
nt
vect
or
to
s
=
0,
we
c
an
ex
press
the
m
axi
m
u
m
chan
ge
directi
on
of
(
)
associat
ed wit
h t
he varia
ble
. Mat
rix
f
or
m
o
f
the e
xpressi
on r
e
pr
ese
nted
as
foll
ow
s:
|
=
0
=
|
=
0
(
−
1
1
1
)
̃
(
16)
w
he
re
|
=
0
=
(
1
|
=
0
,
…
…
.
.
,
|
=
0
)
a
nd
|
=
0
=
(
,
)
|
=
0
=
(
∑
(
,
)
=
1
)
|
=
0
=
∑
(
,
)
=
1
|
=
=
(
,
)
|
=
W
he
re
delta
of
K
ronec
ker
is
represe
nted
a
s
an
d
ra
dial
bas
is
ke
rn
el
a
s
(
,
)
=
(
−
‖
−
‖
2
)
=
(
−
∑
(
−
)
2
=
1
)
. After c
onside
rin
g
=
+
∈
ℝ
:
|
=
0
=
(
,
)
|
=
=
−
2
(
,
)
(
−
)
=
−
2
(
,
)
(
−
)
W
he
re
the
t
rainin
g
point
=
.
T
hu
s
,
by
ap
plyi
ng
(
13),
it
is
possible
to
l
oca
ll
y
rep
rese
nt
any
giv
e
n
i
nput
var
ia
ble
plo
t
in
KP
C
A
.
Furt
her
m
or
e,
by
usi
ng
(
14),
it
is
po
ssi
ble
to
rep
rese
nt
the
ta
ngent
vect
or
ass
ociat
ed
with
any
gi
ve
n
input
va
riabl
e
at
each
sa
m
ple
po
i
nt
[1
9]
.
Ther
e
fore,
a
vecto
r
fiel
d
ca
n
be
dr
a
wn
on
KP
CA
ind
ic
at
in
g
the
grow
t
h
directi
on
of
a
gi
ven
var
ia
ble.
T
here
are
s
om
e
existi
ng
te
ch
niqu
es
to
co
m
pu
te
z
for
sp
eci
fic
kernels
[20
]
. For a
Gau
ssia
n ker
nel
(
,
)
)
=
(
−
‖
−
‖
2
/
2
2
)
,
m
us
t sat
isfy t
he
f
ollo
wing c
onditi
on.
=
∑
(
‖
−
‖
2
/
2
2
)
=
1
∑
=
1
(
−
‖
−
‖
2
)
/
2
2
(17)
4.
E
X
PERI
MEN
TAL SET
UP
To
e
valuate
th
e
eff
ic
ie
ncy
of
kernel
-
base
d
s
peak
e
r
-
sp
eci
fic
featu
re
e
xtract
ion
te
c
hniq
ues,
a
la
ng
uage
ind
e
pende
nt
ut
te
ran
ces
recog
niti
on
e
xperim
ent
was
pe
rform
ed.
The
ex
pe
rim
ent
include
s
520
Japa
nese
wor
ds
from
the
ATR
Japa
nese
C
la
ngua
ge
set
V
oice
data
base,
80
s
pea
ker’s
(
40
m
en
and
40
Fem
al
e).
Audi
o
sam
ples
of
10
iTa
uk
ei
sp
ea
ke
r’
s
we
re
colle
c
te
d
at
ran
dom
and
unde
r
unf
avou
rab
le
co
ndit
ion
s
.
The
av
erage
durati
on
of
t
he
trai
ning
sam
ples
was
six
sec
onds
pe
r
s
pea
ke
r
f
or
al
l
10
s
pe
aker’s
a
nd
out
of
t
wen
ty
utter
ances
of
e
ach
sp
ea
ke
r
just
one
was
us
e
d
f
or
trai
ni
ng
pur
pose
[2
1]
.
F
or
m
at
chi
ng
pur
poses
t
he
rem
ai
nin
g
19
vo
ic
e
sa
m
ples
wer
e
us
e
d
f
ro
m
the
corp
us
.
We
ha
ve
rec
orde
d
utterances
f
or
t
his
inv
est
igati
on
wer
e
at
on
e
sit
ti
ng
for
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
4
,
A
ugus
t
2020
:
3
5
08
-
3
5
1
8
3516
each
s
pea
ker
.
The
te
xt
f
or
t
he
utterances
was
rand
om
l
y
sel
ect
ed
by
s
peak
e
r.
The
m
ai
n
voic
e
r
eco
rd
i
ng
s
consi
st
of
bo
t
h
m
al
e
and
fem
al
e
sp
eake
r’
s
of
t
wen
ty
uttera
nce
of
each
usi
ng
sam
pling
r
at
e
of
16
kHz
with
16 b
it
s/sam
ple
[22
]
.
Sp
eec
h
featu
re
s,
each
c
on
sist
ing
of
24
Me
l
-
scal
e
fr
e
quen
cy
cepstral
co
eff
ic
ie
nts
(MF
CC
s),
we
re
extracte
d
from
these
uttera
nce
s
for
eve
ry
20
-
m
s
Ha
m
m
ing
-
windowe
d
f
ra
m
e
with
10
-
m
s
fr
am
e
sh
ifts.
P
rior
t
o
MFC
C
com
pu
ta
ti
on
,
voic
e
act
ive
detect
ion
was
ap
plied
to
rem
ov
e
sal
ie
nt
non
-
sp
ee
ch
reg
i
on
s
th
at
m
ay
be
include
d
i
n
an
utterance
[
23,
24
]
.
O
ur
basic
strat
egy
is
to
create
an
utterance
-
in
depen
den
t
GMM
usi
ng
al
l
the
utterance
s
to
be
cl
us
te
re
d,
f
ollo
wed
by
an
adap
ta
ti
on
of
the
uttera
nc
e
-
in
dep
e
ndent
GMM
fo
r
ea
ch
of
the
utteran
ces
us
in
g
m
axi
m
u
m
a
po
ste
rior
i
(M
AP)
est
i
m
ati
on
[
25
]
.
This
te
ch
nique
com
es
fr
om
the
GMM
-
UB
M
strat
egy
[
26
]
fo
r
FSR
in
wh
ic
h
the
nec
essary
s
peak
e
r
-
ex
plici
t
m
od
e
ls
ar
e
m
ade
by
tun
i
ng
the
par
am
et
ers
of
a
wides
pr
ea
d
s
peak
e
r
m
od
el
pr
e
-
prepare
d
by
util
iz
ing
disc
our
se
inf
or
m
at
ion
f
ro
m
nu
m
erous
s
pe
aker’s.
W
e
ar
e
us
i
ng
Lan
gu
age
in
de
pende
nt
Ga
us
sia
n
m
ixtur
e
m
od
el
in
g
f
ollo
we
d
by
MA
P
adap
ta
ti
on i
n
l
angua
ge
in
de
pe
nd
e
nt
forensi
c
sp
ea
ker rec
og
niti
on
.
Thro
ughout
th
e
exp
e
rim
ent,
10400
uttera
nc
es
wer
e
us
e
d
a
s
trai
ning
data
and
t
he
rem
ain
in
g
31,
200
utterances
were
us
e
d
a
s
te
st
data.
The
sa
m
pl
ing
rate
of
the
au
dio
si
gn
al
is
16
kHz.
13
Me
l
-
C
epstral
coeffic
ie
nts
ex
tract
ed
us
i
ng
25.
6
m
s
H
am
m
i
ng
wind
ow
s
w
it
h
10
m
s
sh
ifts.
Fig
ur
e
6
sho
w
the
eq
ual
er
r
or
rate
(EER)
of
K
-
m
edo
i
d,
Fu
zzy
C
-
m
eans,
G
us
ta
fson
a
nd
Kess
el
and
Gat
h
-
G
eva
cl
ust
eri
ng
al
gorithm
s
Gev
a
f
or
6
sec
of
voic
e
da
ta
of
ATR
Jap
anese
C
la
ngua
ge.
T
he
fore
nsi
c
sp
eake
r
rec
ogniti
on
e
ff
ic
i
ency
of
of
K
-
m
edo
id,
Fu
zzy
C
-
m
eans,
Gu
sta
fson
a
nd
Kessel
and
G
at
h
-
Gev
a
cl
us
te
rin
g
al
gori
thm
s
are
95
.
2%
,
97.
3%
,
98.
5%
an
d
99.7%
an
d
EE
R
are
3.
62%,
2.91%
,
2
.
82
%
and
2.6
1%
re
sp
ect
ively
.
T
he
EER
im
pr
ove
m
ent
of
Gath
-
Ge
va
te
chn
iq
ue
base
d
FSRsy
ste
m
c
om
par
ed
with
Gu
sta
fson
a
nd
Kessel
an
d
F
uz
zy
C
-
m
eans
is
8.0
4%
an
d
11.
49%
resp
ect
ively
.
Table
2
il
lustra
te
e
ff
ic
ie
ncy
a
nd
EERo
f
t
he
FSRsy
ste
m
fo
r
of
K
-
m
edo
id,
Fu
zzy
C
-
m
eans,
G
us
ta
f
s
on
and
Kessel
an
d
Gath
-
Gev
a
c
lusterin
g
al
gor
it
h
m
s
resp
ect
ively
fo
r
ATR
Japa
nese
C
la
ng
ua
ge
.
Fi
gure
7
sho
w
the
eq
ual
e
rro
r
rate
(EER)
K
-
m
edo
id
,
F
uz
zy
C
-
m
eans,
Gu
sta
fson
a
nd
Kessel
a
nd
Gath
-
Ge
va
cl
ust
ering
al
gorithm
s
Ge
va
f
or
6
sec
of
vo
ic
e
data
of
10
iTa
uk
ei
s
peak
e
r’s
cross
la
ng
ua
ge.The
FSReff
ic
ie
nc
y
of
of
K
-
m
edo
id
,
Fuzz
y
C
-
m
eans,
Gu
sta
fson
an
d
Kessel
and
G
at
h
-
Gev
a
cl
us
t
erin
g
al
gorith
m
s
are
93
.2%,
96
.
6%,
97.7%
a
nd
98.
8%
a
nd
EER
are
4.23%,
3.4
2%,
3.3
3%
an
d
3.11%
resp
e
ct
i
vely
.
The
E
ER
im
pr
ov
em
ent
of
Gath
-
Ge
va
te
c
hn
i
qu
e
based
FSRsy
ste
m
com
par
ed
with
Gu
sta
fson
an
d
Kessel
an
d
F
uzzy
C
-
m
eans
is
7.0
7%
and
9.9
6%
res
pecti
vely
.
Ta
bl
e
3
il
lustrate
Eff
ic
ie
ncy
a
nd
EERo
f
the
F
SRsy
stem
fo
r
of
K
-
m
edo
i
d,
Fu
zz
y
C
-
m
eans,
Gusta
fson
a
nd
Ke
s
sel
and
Gat
h
-
Gev
a
cl
us
te
ri
ng
al
gorithm
s
r
especti
vely
f
or
10
iTa
ukei
s
pe
aker’s
cro
ss
lan
guage
.
In
ad
diti
on,
t
o
deci
de
how
m
any
cl
u
ste
rs
sho
uld
be
pr
oduce
d,
the
cl
us
te
rin
g
m
et
h
od
ha
s
bee
n
integrate
d
wit
h
the
Ba
ye
sia
n
in
form
at
ion
crit
erion.
E
xp
e
rim
ental
resu
lts
show
t
hat
th
e
nu
m
ber
of
c
lusters
autom
at
ic
ally
determ
ined
ca
n
a
ppr
ox
im
at
e
the
act
ual
popula
ti
on
siz
e
of
t
he
s
pea
ke
r.
A
n
e
xperi
m
ental
evaluati
on
of
t
he
pe
rfo
rm
ance
of
t
he
f
or
e
nsi
c
r
eco
gn
it
io
n
syst
e
m
eff
ic
ie
ncy
of
K
-
m
edo
id,
Fu
zzy
C
-
m
eans,
Gu
sta
fson
a
nd
Kessel
an
d
G
at
h
-
Ge
va
cl
ust
erin
g
al
gorith
m
s
are
95
.
2%,
97
.
3%
,
98.
5%
and
99.7
%
a
nd
EER
are
3.62
%
,
2.9
1%,
2.8
2%
a
nd
2.61
%
res
pe
ct
ively
.
The
E
ER
i
m
pr
ovem
ent
of
Gath
-
G
eva
te
ch
nique
base
d
FSRsy
ste
m
co
m
par
ed
with
Gustafs
on a
nd Kesse
l an
d
F
uzz
y C
-
m
eans is 8.04% a
nd
11.49%
res
pecti
vely
.
Figure
6.
Eq
ua
l err
or r
at
e
of
K
-
m
edo
id
, fuz
zy
C
-
m
eans,
gus
ta
fs
on a
nd kessel
and
gath
-
gev
a
f
or 6 sec
of voice
data of
ATR
j
a
pan
e
se C la
ng
ua
ge
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N:
20
88
-
8708
Sp
e
aker
sp
eci
fi
c feature
base
d cl
us
te
rin
g and
it
s app
li
catio
ns i
n
l
anguage i
nd
e
pe
nd
e
nt …
(
Sa
ty
an
and Sin
gh
)
3517
Table
2.
E
ff
ic
i
ency an
d EER
of the
FSRsy
stem
f
or of
K
-
m
edo
i
d,
f
uzzy C
-
m
eans,
gusta
f
so
n a
nd k
es
sel
and
gath
-
ge
va
cl
us
t
erin
g
al
gorit
hm
s r
especti
vely
f
or
ATR
j
apa
nese C la
ngua
ge
Ef
f
icien
cy
in
%
EER
in
%
K
-
m
ed
o
id
9
5
.2
3
.6
2
Fu
zzy
C
-
m
e
an
s
9
7
.3
2
.91
Gu
staf
so
n
and
Kes
sel
9
8
.5
2
.82
Gath
-
Gev
a
9
9
.7
2
.61
Figure
7.
Eq
ua
l err
or r
at
e
of
K
-
m
edo
id
, fuz
zy
C
-
m
eans,
gusta
fs
on a
nd kessel
and
gath
-
gev
a
for 6
sec
of voice
data of it
au
kei
sp
ea
ker’s c
ro
s
s
langua
ge
Table
3.
E
ff
ic
i
ency an
d EER
of the
FSRsy
stem
f
or of
K
-
m
edo
i
d,
f
uzzy C
-
m
eans,
gusta
f
so
n a
nd k
es
sel
and
gath
-
ge
va
cl
us
t
erin
g
al
gorit
hm
s r
especti
vely
f
or
10 it
auk
ei
sp
ea
ker’s c
r
oss l
an
gu
a
ge
Ef
f
icien
cy
in
%
EER
in
%
K
-
m
ed
o
id
9
3
.2
4
.23
Fu
zzy
C
-
m
e
an
s
9
6
.6
3
.42
Gu
staf
so
n
and
Kes
sel
9
7
.7
3
.33
Gath
-
Gev
a
9
8
.8
3
.11
5.
CONCL
US
I
O
N
This
stud
y
ex
a
m
ined
m
et
ho
ds
f
or
im
pr
oving
the
m
easur
em
ent
of
inter
-
utterance
si
m
il
arity
fo
r
sp
ea
ker
cl
us
te
r
ing
.
The
relat
ion
s
hi
ps
of
sim
il
arit
y
between
the
uttera
nces
to
be
cl
us
te
re
d
can
be
e
xploit
ed
m
or
e
easi
ly
and
e
ff
ic
ie
ntly
by
us
in
g
a
vo
ic
e
c
har
ac
te
ris
ti
c
ref
ere
nce
s
pace.
W
e
pr
ese
nte
d
s
ever
al
i
m
ple
m
entat
io
ns
for
the
de
velo
pm
ent
of
ref
e
re
nce
s
pa
ces.
I
n
par
ti
c
ular,
i
n
order
to
ca
ptu
r
e
th
e
m
os
t
represe
ntati
ve
char
act
e
risti
cs
of
the
voic
es
of
sp
ea
ke
r’
s
,
the
ref
ere
nce
sp
ace
was
re
pr
ese
nted
as
a
set
of
ei
genv
ect
ors
obta
ined
by
ap
plyi
ng
the
te
c
hn
i
qu
e
of
sel
f
-
voic
e
to
the
set
of
ex
pr
es
sion
s
t
o
be
cl
ust
ered.
This
has
res
ulted
in
a
sig
nific
ant
i
m
pr
ovem
e
nt
in
sp
ea
ke
r
-
c
lusterin
g
pe
rfo
rm
ance
com
pa
red
to
t
he
tradi
ti
on
al
inter
-
s
peec
h
sim
il
arity
m
eas
ur
e
base
d
on
the
ge
n
erali
ze
d
li
kelihoo
d
r
at
io.
H
ow
e
ve
r,
we
ha
ve
rese
arch
e
d
the
m
et
ho
d
f
or
creati
ng
cl
us
te
rs
outsi
de
tradi
ti
on
al
hierar
c
hi
cal
cl
us
te
ring
s
o
that
al
l
withi
n
-
cl
us
te
r
uttera
nces
can
be
from
a
sing
le
s
pea
ker
as
fa
r
as
possi
ble.
T
his
c
rite
r
ion
was
co
ncei
ved
as
a
pr
o
ble
m
of
cal
culat
in
g
a
nd
op
ti
m
iz
ing
the
overall
purity
of
the
cl
us
te
r.
By
re
pr
ese
nting
cl
us
te
r
purity
as
a
f
unct
io
n
of
inter
-
utter
ance
si
m
il
arity
and
app
ly
ing
t
he
gen
et
ic
al
gorithm
to
find
a
so
luti
on
to
thi
s
prob
le
m
,
we
hav
e
dem
on
strat
e
d
a
f
ur
th
er
i
ncrea
se
in
sp
ea
ke
r
-
cl
ust
erin
g
e
ff
ic
ie
ncy
c
ompare
d
t
o
c
onve
ntion
al
ag
glom
erati
on
hiera
rch
ic
a
l
cl
us
te
rin
g.
REFERE
NCE
S
[1]
S.
Singh,
“
Fore
nsic
and
Autom
at
i
c
Speak
er
R
e
cogni
ti
on
S
y
st
e
m
,
”
in
In
te
rnati
onal
Journal
of
Elec
tri
cal
and
Computer
Engi
n
ee
ring (
IJE
C
E)
,
vol.
8
,
no
.
5
,
pp
.
2804
-
2811,
Oc
to
ber
2018.
[2]
S.
Singh,
“
Supp
ort
Vec
tor
Mac
hine
Based
Approac
hes
For
Real
Ti
m
e
Autom
at
ic
Speake
r
Re
co
gnit
ion
S
y
st
em,”
Inte
rnational
Jo
urnal
of Appl
i
ed Engi
ne
ering
R
ese
arch
,
vo
l. 13, n
o.
10
,
pp
.
8561
-
8567,
2018
.
Evaluation Warning : The document was created with Spire.PDF for Python.