Internati
o
nal
Journal of Ele
c
trical
and Computer
Engineering
(IJE
CE)
V
o
l.
5, N
o
. 2
,
A
p
r
il
201
5, p
p
.
31
1
~
31
7
I
S
SN
: 208
8-8
7
0
8
3
11
Jo
urn
a
l
h
o
me
pa
ge
: h
ttp
://iaesjo
u
r
na
l.com/
o
n
lin
e/ind
e
x.ph
p
/
IJECE
Classi
fication of
Emotion
a
l Speech of Children Usi
n
g
Probabilistic Neural Network
H.K.
P
a
lo,
Mi
hir Narayan
Mohanty
Depart
em
ent of
Ele
c
troni
cs and
Com
m
unication
Engine
ering,
Ins
titut
e
of
T
echni
c
a
l
Educa
tion
and
Resear
ch,
Siksha ‘O’
Anusandhan University
, B
huban
e
swar, Odisha, In
dia
Article Info
A
B
STRAC
T
Article histo
r
y:
Received Ja
n
2, 2015
Rev
i
sed
Feb
10
, 20
15
Accepted
Feb 25, 2015
Child emotions are highl
y
flex
ible and ov
erlapping. The r
eco
gnition is
a
difficu
lt t
a
sk when single
em
otion conv
e
y
s
m
u
ltiple info
rm
ations. W
e
anal
yz
e
the r
e
l
e
vanc
e and
im
portanc
e of
th
es
e fe
atures
an
d us
e tha
t
information to design classifier
arch
itectur
e. Designing of a sy
stem for
recognition of
children
emotions
with re
asonable accuracy
is
still
a ch
allenge
s
p
ecifi
cal
l
y
wit
h
reduced f
e
a
t
u
r
e s
e
t.
In this paper,
Probab
ilistic
neur
al
network (PNN)
has been designed for su
ch tas
k
of clas
s
i
ficat
io
n. P
NN ha
s
faster tr
aining
abilit
y
wi
th con
t
i
nuous class probabili
t
y
densi
t
y
functions. I
t
provides better classification ev
en with
reduced f
eatur
e set. LP_V
QC and pH
vectors
ar
e us
ed as
the fea
t
ures
for the cl
as
s
i
fier
. It has
been
att
e
m
p
ted t
o
des
i
gn the P
NN clas
s
i
fie
r
with t
h
es
e fea
t
ures
. V
a
rious
em
otions
like ang
r
y,
bore, s
a
d and ha
pp
y
have be
en c
ons
idered for thi
s
piece of work. All thes
e
emotions have b
een
coll
ected
fr
om children
in
th
ree d
i
ffer
e
nt languages as
English, Hind
i,
and Odia. Result show
s
rem
a
rkable
cl
as
s
i
fica
ti
on ac
cur
a
c
y
for these classes of emotions. It
has b
een ver
i
fied in standard d
a
tabse EMO-
DB to va
lida
t
e
t
h
e resul
t
.
Keyword:
AN
N
Hu
rst param
e
ter
LP-
VQC
pH
feat
ure
vec
t
ors
PNN
Copyright ©
201
5 Institut
e
o
f
Ad
vanced
Engin
eer
ing and S
c
i
e
nce.
All rights re
se
rve
d
.
Co
rresp
ond
i
ng
Autho
r
:
Mih
i
r
N
a
r
a
yan Mo
han
t
y
Depa
rt
em
ent
of El
ect
r
oni
cs
a
n
d
C
o
m
m
uni
cat
i
on E
n
gi
nee
r
i
n
g
,
In
stitu
te
o
f
Tech
n
i
cal Edu
catio
n and
Research
,
Si
ksh
a
‘
O
’
A
n
usa
n
d
h
a
n
Uni
v
ersi
t
y
, B
h
uba
n
e
swar
,
Odi
s
ha,
I
ndi
a
Em
a
il:
mih
i
rmo
h
a
n
t
y@so
auniv
e
rsity.ac.in
1.
INTRODUCTION
Hum
a
n speec
h
i
s
t
h
e l
i
ngui
st
i
c
act
t
h
at
con
v
ey
s i
n
f
o
rm
at
ion a
b
out
t
h
e s
p
eake
r
.
Al
t
h
ou
gh
h
u
m
a
n
e
m
o
tio
n
s
canno
t ch
ang
e
t
h
e con
t
en
t, bu
t
it is ex
pre
ssed in
sep
a
rate
way fro
m
th
e no
rm
al sp
eech
.
Th
is
inform
ation of a speake
r
can effect hi
s/he
r
mental states.
It is challengi
ng
in ext
r
acting suitable features of
hum
an em
ot
i
onal
spee
ch t
h
at
can
best
re
p
r
e
s
ent
a
part
i
c
ul
a
r
em
ot
i
on u
n
a
m
bi
guou
sl
y
.
A
com
p
ari
s
o
n
b
e
t
w
een
di
ffe
re
nt
di
vi
si
on
of em
ot
i
ons
as har
d
-
w
i
r
e
d
vs. s
o
ci
al
l
y
l
e
arne
d a
nd
pri
m
ary
’
vs
. ‘se
c
o
n
d
ary
’
em
ot
i
ons
wi
t
h
its u
n
i
v
e
rsality in
con
t
ain
i
ng
v
a
ri
o
u
s
classes o
f
em
o
tio
n
s
is p
r
esen
ted
i
n
[1
]. A
rev
i
ew on
d
i
fferen
t
stan
d
a
rd
e
m
o
tio
n
a
l
d
a
tab
a
ses u
s
ed
b
y
d
i
fferen
t
research
ers, th
eir accessib
ility an
d
p
e
rform
a
n
ce can
b
e
foun
d
i
n
[2-3
].
Diffe
re
nt acoustic features as fundam
ental freque
n
cy (F
0), energy and duration we
re explored in m
a
ny cases
as param
e
ters of s
p
eech em
otional utteranc
es [2-3]. In
[4], statistical prope
rties of
various ac
oustic s
p
eech
e
m
otional feat
ures
like ze
ro
crossi
ng
rate (
Z
CR), Harm
on
ics-to-
N
oise-R
atio
(
HNR
), fo
rm
ant
were der
i
ved
t
o
represe
n
t e
m
otional uttera
nce
s
of Berlin dat
a
bases.
Recent
l
y researchers
have foc
u
se
d on determ
ining Hurst
param
e
ter and
on tim
e
-fre
qu
ency
pH
featu
r
es f
o
r em
oti
onal speec
h [5
-
7
]
.
EM
O-
DB database is a standa
r
d
dat
a
base as e
x
press
e
d i
n
[8]
.
It
has bee
n
u
s
e
d
i
n
m
o
st
of the cases. Vari
ouse feat
ures s
u
ch as Linea
r
predictor
coefficients
(LPC), Linea
r
predic
tor ce
pstral c
o
efficients (LPCC),
Mel fre
quency cepstral c
o
efficient
s
(MFCC), Percep
tu
al lin
ear
p
r
ed
ictio
n
(PLP), LP_VQC
and
pH vect
ors are
use
d
for em
otional
speec
h
recogn
itio
n
are fo
und
in
lite
ratu
re [9-12
]
. Si
m
ilarly, c
l
as
sifiers lik
e Mu
ltilayer p
e
rcep
tro
n
(MLP),
Rad
i
al
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Classification of
Emotional
Speech of C
h
ildren
Using
Pr
obabilistic Neural Network
(Mih
ir NM)
31
2
basi
s f
unct
i
o
n
net
w
or
k (R
B
F
N) are
use
d
f
o
r cl
assi
fi
cat
i
o
n
of di
ffe
rent
e
m
oti
ons by
a
u
t
h
o
r
s [
1
0
-
1
2
]
.
H.
K.
Palo et. al. empha
sized
on c
h
ild e
m
oti
onal s
p
eech classi
fication using differe
nt
neural ne
twork m
odels in [11-
12]
.
U
se o
f
K-
nearest
nei
g
h
b
o
u
r
ho
o
d
Li
nea
r
di
scri
m
i
nant
cl
assi
fi
er (LD
C
) fo
r cl
assi
fi
cat
i
on of em
ot
i
ons ha
ve
b
een
also
u
s
ed b
y
m
a
n
y
research
ers [13
-
16]. A co
m
p
arativ
e stud
y on
M
u
ltilayer Percep
ton
,
Ran
d
o
m
Fo
rest,
Prob
ab
ilistic Neu
r
al Net
w
orks an
d
Su
ppo
rt
Vector Mach
ine also
h
a
s b
e
en
rep
o
rted
fo
r
th
e d
i
fferen
t
sp
eech
e
m
o
tio
n
classificatio
n
[17
-
20
]. Pro
b
a
b
ilistic n
e
ural n
e
two
r
k
h
a
s
b
e
en
u
s
ed
in
m
a
n
y
p
a
ttern
classifi
catio
n
task
s [2
1-2
2
]
.
Decision
Trees, Artificial Neu
r
al
Netw
o
r
k
s
(ANN), Pro
b
a
b
ilistic
n
e
u
r
al n
e
two
r
k
an
d
ran
d
o
m
forest techniques are a
p
plied
for cla
ssif
i
catio
n of
d
i
ff
er
en
t
sign
als [19
-
24
].
O
v
er
th
e la
s
t
d
e
c
a
d
e
s
d
i
ff
e
r
en
t
ANN classifie
r’s
perform
ance and their c
o
m
p
arison in cl
assi
fy
i
ng s
p
ee
ch an
d em
ot
i
ons
were e
x
pl
o
r
ed
by
m
a
ny researc
h
ers,
while a little am
ount of work
has
be
en re
ported
for classifi
cation of c
h
ild em
otions.
Mo
tiv
ated
b
y
th
e flex
ib
ility i
n
corp
orated
i
n
ch
ild
em
o
tio
ns, au
t
h
ors
h
a
ve tak
e
n
an
attem
p
t to
classify
four
cl
asses of em
ot
i
o
n
s
as ang
r
y
,
bor
e, sad a
nd
hap
p
y
usi
n
g LP_
V
QC
and
pH feat
ure
vect
ors
wi
t
h
PN
N
classifiers.
The pa
per i
s
o
r
gani
ze
d as fol
l
ows
.
Sect
i
o
n 2
deal
s w
ith
th
e research
m
e
th
o
d
. In
th
is sect
io
n
bo
th
th
e
m
e
t
hods f
o
r cl
assi
fi
cat
i
on and
feat
u
r
e ext
r
act
i
on a
r
e p
r
esent
e
d
.
The
r
e
sul
t
i
s
di
scus
sed i
n
Sect
i
o
n
3 an
d
fin
a
lly in
secti
o
n 4, it con
c
l
u
d
e
s t
h
is p
i
ece of work.
2.
R
E
SEARC
H M
ETHOD
2.
1. Me
th
od o
f
Cl
assific
a
ti
o
n
The choice for an appropriate
classi
fier for
e
m
otional spee
ch is a
co
m
p
lex task. Popular classifiers
for em
o
tio
n
reco
gn
itio
n su
ch as Lin
e
ar
Discrimin
a
n
t
Classifiers (LDCs) and
k
-
Nearest Neigh
bou
r
(kNN)
classifiers h
a
ve b
een
u
s
ed
in literatu
re [13-1
6
]
.
Alth
oug
h PNN is
n
o
t
necessarily th
e b
e
st classifier
b
u
t
it
provides
good statistical pr
operties. Classification accuraci
es of
Decisi
on
Trees s
u
ch as
RF, Artificial
Neural
Net
w
or
ks
(A
N
N
) a
n
d P
r
o
b
a
b
i
l
i
s
t
i
c
neural
ne
t
w
o
r
k
we
re f
o
u
n
d
t
o
be si
m
i
l
a
r.
AN
N
req
u
i
r
e
d
by
fa
r t
h
e
hi
ghe
st
cal
cul
a
t
i
on t
i
m
e
s, w
h
ereas t
h
e t
r
ai
ni
ng a
n
d
t
e
st
i
ng of R
F
t
ook
us
ual
l
y
lon
g
e
r
t
h
an P
N
N. A
r
t
i
f
i
c
i
a
l
Neura
l
Net
w
or
k has
m
a
ny
di
sadva
nt
ages
, suc
h
a
s
com
p
l
e
x opt
im
i
zati
on, l
o
w
ro
bust
n
ess a
n
d m
u
ch t
r
ai
ni
n
g
t
i
m
e
.
R
a
nd
om
Fores
t
i
n
co
nt
rast
i
s
easy
t
o
u
s
e, s
i
nce o
n
l
y
o
n
e
vari
a
b
l
e
nee
d
s
t
o
be
set
by
t
h
e
user
. H
o
we
ver
,
i
t
s
cl
assi
fi
cat
i
on a
ccuraci
es ca
n
not
sat
i
s
fy
t
h
e
m
achi
n
e-l
ear
ni
ng m
e
t
hods
w
h
ereas i
t
s
r
o
bu
st
ness wa
s am
on
g t
h
e
best
[
2
3-
2
4
]
.
PN
N i
s
a speci
al
cl
ass of R
a
d
i
al
basi
s fu
nct
i
on
(R
B
F
)
net
w
or
k u
s
ed
fo
r cl
assi
fi
cat
i
on [
1
7
-
1
8
,
20
-
24]
.
It is
bene
ficial to m
a
ny appilicati
ons i
n
cludi
n
g s
p
eech bec
a
use
of the
s
p
e
e
d
of learning. It is
one
of the non-
p
a
ram
e
tric
m
e
th
od
s. Th
is typ
e
of n
e
t
w
ork ad
op
ts
pro
b
a
b
ilistic
m
e
th
od
to
classify
d
a
ta. Th
e PNNs are
effective
discri
minative classifiers with s
e
ve
ral outsta
ndi
ng characteristics
,
nam
e
ly
: they are ha
ving a
orde
r
of m
a
gnitude
m
u
ch faster a
n
d accurate t
h
an m
u
ltilaye
r perce
p
tron
networks.
The
ne
tworks are
rel
a
tively
in
sen
s
itiv
e t
o
o
u
tliers
h
a
v
i
ng
an
inh
e
ren
tly p
a
rallel structu
r
e an
d gu
aran
teed
to
conv
erg
e
to an op
ti
m
a
l
classifier as t
h
e size of the
training set inc
r
eases.
P
N
N i
s
base
d
on B
a
y
e
s o
p
t
i
m
al
cl
assi
fi
cat
i
on.
PN
N
s
can
gene
rate accurate pre
d
icted target
proba
bility scores wi
t
h
no l
o
cal minima issues. No
extensi
v
e retra
i
ning i
s
necessa
ry
i
f
t
r
a
i
ni
ng
sam
p
l
e
s are a
dde
d
o
r
re
m
oved. T
h
ese
charact
e
r
i
s
t
i
c
s have
m
a
de PN
Ns
very
p
o
pul
ar a
n
d
success
f
ul.
In
PNN, th
e
o
p
e
ration
s
are o
r
g
a
n
i
zed in
t
o
m
u
ltila
yered
feed
forward n
e
tw
ork with four layers:
in
pu
tlayer,
p
a
ttern
layer, summa
tio
n
and
o
u
t
p
u
t
layer.
Fun
d
a
m
e
n
t
ally
it is b
a
sed on
Bayesian
cl
assifier
tech
n
i
qu
e.Th
e
first layer si
m
p
ly d
i
strib
u
t
es th
e in
pu
t to
th
e n
e
uro
n
s
in
th
e
p
a
ttern
layers.
Th
is layer u
s
ing
th
e
gi
ve
n set
o
f
d
a
t
a
poi
nt
s as t
h
e cent
e
rs f
o
r
m
s
t
h
e Gaus
si
an f
u
nct
i
ons
. I
t
or
gani
zes t
h
e
t
r
ai
ni
n
g
set
s
u
ch t
h
at
each input vec
t
or is represent
e
d by
an individual proces
sing layer called
the summ
a
tion layer. There a
r
e as
many processi
ng elem
ents as the
num
ber
of classes to be
recognized in
the s
u
mm
a
tion
layer. T
h
e
distances
f
r
o
m
th
e in
p
u
t testin
g
v
ect
or
s to
t
h
e input tr
ain
i
n
g
v
ect
o
r
s is co
m
p
u
t
ed
an
d
a v
ect
o
r
t
h
at ind
i
cates th
e
closenes
s of the training a
nd
testing
inputs is produce
d
. For each class of
testing inputs the summ
ation layer
su
m
s
th
e
con
t
rib
u
tion
s
o
f
p
r
ev
iou
s
layer
ou
t
p
u
t
b
y
g
i
v
i
n
g
a
n
e
t ou
tpu
t
vecto
r
o
f
p
r
o
b
a
b
ilities.
Basically
an
avera
g
ing operation of the
outputs from
the
pattern layer for each cl
ass is perform
e
d by the summation layer
A PNN uses Parzen
window
probabilistic
density function esti
m
a
tor for
each
of the classes with a mi
xture of
Gau
s
sian
b
a
sis fu
n
c
tion
s
[2
2
]
. Figure
1
sh
ows the
b
a
sic
stru
cture o
f
prob
ab
ilistic
n
e
ural n
e
two
r
k
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
IJEC
E V
o
l
.
5, No
. 2, A
p
ri
l
20
15
:
31
1 – 3
1
7
31
3
Fig
u
re
1
.
Basic Pro
b
ab
ilistic Neural
Netwo
r
k
St
ru
ct
u
r
e
Fo
r an
inpu
t p
a
ttern
, pat
t
ern
vect
or
di
m
e
nsi
on
d,
s
m
oot
hi
ng para
m
e
t
e
r
, with
each
inp
u
t
training vector
,
co
nsi
d
e
r
e
d
t
o
be a
cent
e
r
o
f
a ke
rnel
f
unct
i
on t
h
e
out
put
of t
h
e
pat
t
e
rn
l
a
y
e
r i
s
gi
ven
b
y
[2
1]
∅
,
1
2
,
,
2
(1
)
For
classes of em
o
tio
n
s
t
h
e prob
ab
ility d
e
n
s
ity fun
c
tion
of eac
h clas
s
i
s
gi
ve
n by
t
h
e
equat
i
o
n
1
2
1
,
,
2
(2
)
Whe
r
e i
=
1,
2,
…,
m
and
is to
tal nu
m
b
er of train
i
ng
p
a
ttern
s
for each
class
;
is th
e
p
th
i
nput
vector. Here
m
t
a
kes t
h
e
val
u
e of
fo
u
r
co
rre
spo
n
d
i
n
g t
o
a
n
gry
,
b
o
re,
sad
and
ha
ppy
s
p
e
ech em
ot
i
ons.
The
varia
n
ce
should
be c
h
osen judiciously as it
acts as a sm
oot
hi
n
g
fact
o
r
t
o
s
o
ft
e
n
th
e su
rface. Fo
r sm
all
th
e
classifier m
a
y not
ge
neraliz
e well
a
nd ca
n c
r
eate
distinct m
odes.
Larger
allo
ws i
n
terpo
l
atio
n b
e
t
w
een
poi
nt
s. Very
l
a
rge
ap
prox
im
a
t
e th
e pdf to b
e
Gau
ssian resu
l
tin
g
loss
o
f
d
e
t
a
ils.
To d
i
stingu
ish class
to
wh
ich
i
n
pu
t
v
ect
or
b
e
lon
g
s t
h
e Bayesian d
eci
sio
n
ru
le is app
lied
.
Here it
is
assum
e
d that
a p
r
io
ri
proba
bi
lities and losse
s associated
with incorr
ect decision is same for each class. By
th
is ru
le th
e est
i
m
a
ted
class
of th
e p
a
ttern
, base
d o
n
t
h
e
o
u
t
p
ut
of al
l
t
h
e
sum
m
a
t
i
on l
a
y
e
r neu
r
ons i
s
gi
ve
n by
a
r
g
m
a
x
,
1
,2,
…
,
(3
)
2.
2. Fea
t
ure
E
x
tr
acti
on
Selecting a sui
t
able feature that can best represen
t
a pa
rt
i
c
ul
ar em
ot
i
on i
s
one o
f
t
h
e f
o
r
e
m
o
st
wor
k
s
in
th
e field
o
f
e
m
o
tio
n
s
recog
n
ition
[1
1
]
.
Sin
ce ch
ild
em
o
tio
n
s
are
v
e
ry
flex
ib
le th
e sel
ectio
n
b
e
co
m
e
s
m
o
re
tediou
s.
In
o
r
der t
o
ca
ptu
r
e
diffe
re
nt aspec
t
s of
v
o
cal tra
c
t
m
echanism
bot
h tim
e-freq
u
ency
param
e
ter an
d
sp
ectral p
a
ram
e
ters
as features
h
a
v
e
b
e
en
atte
m
p
ted
.
Furth
e
r t
h
e effect
o
f
feat
u
r
e red
u
c
tion
cap
a
b
i
lity o
f
v
ector q
u
a
n
tizatio
n
with
LPC
sp
ectral
feature we
re
c
o
mpare
d
with pH
feat
ure vectors
for PNN
cla
ssifiers.
The t
w
o
di
f
f
er
ent
m
e
t
hods
fo
r feat
ure
ext
r
ac
t
i
on i
s
e
x
pl
ai
ne
d as
f
o
l
l
o
w
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Classification of
Emotional
Speech of C
h
ildren
Using
Pr
obabilistic Neural Network
(Mih
ir NM)
31
4
pH time
-fre
q
uency fe
ature v
ectors
pH
is a tim
e
-freq
u
e
n
cy
voca
l
source
featu
r
e [7-
8
]
.
It ex
p
l
ores the tim
e-
fre
que
ncy
m
u
lti-resol
u
tion
capability of di
screte wa
velet
transform
(DWT) to ca
pt
ur
e the
higher order correlations of the
speech sam
p
le
s
t
hus a
b
l
e
t
o
cl
assi
fy
t
h
e em
ot
i
ons i
n
a
bet
t
e
r way
.
It
c
o
nsi
s
t
s
of
a
vect
o
r
of
H
u
rst
c
o
m
pone
nt
0
1
obtaine
d from
each fram
e of the signal c
onc
erne
d. Si
nce
Hurst
param
e
ter closely related to the decaying rate
o
f
au
to
correlatio
n
co
efficien
t
fun
c
tion
(ACF),
1
1
. The
rel
a
t
i
o
nshi
p
of t
h
e H
u
rst
param
e
t
e
r
as
→
∞
i
s
gi
ve
n as
~
2
1
(4
)
A value
o
f
H
valu
es o
f
0
0
.
5
i
ndi
c
a
t
e
s em
ot
i
ons
wi
t
h
hi
g
h
e
n
e
r
gy
, si
nce t
h
e
d
o
m
i
nant
hi
g
h
fre
que
ncy com
p
one
n
ts ha
ve a -9dB/octave roll-off a
nd th
e
ACF ra
pidly decays to zero.
Em
otions with lower
ener
gy
t
e
n
d
s t
o
have
H
val
u
es
o
f
0
.5
1
with
-15
d
B
/o
ctav
e
PSD
ro
ll-o
f
f and
lead
s to
a slo
w
ly
vani
s
h
i
n
g
AC
F
.
R
e
Scal
ed a
d
j
u
st
ed
ra
nge
(R
/
S
) st
at
i
s
t
i
c
, Hi
guc
hi
m
e
t
hod
and
A
b
ry
-
V
ei
t
c
h (
A
V) est
i
m
at
or ca
n
b
e
u
s
ed
t
o
ex
tract th
e
Hurst
param
e
ter. Howev
e
r Abr
y–Veitch
(AV) estimato
r
u
s
i
n
g wavelet d
eco
m
p
o
s
i
tio
n
is
l
e
ss t
i
m
e
cons
um
i
ng and
doe
s not
ass
u
m
e
the speec
h si
g
n
a
l
s
t
o
be fract
a
l
as i
n
form
er m
e
t
hods
hence
was a
natural choice.
The ste
p
s
of extrac
ting feature
s
are de
picted below.
Wav
e
let
d
ecom
p
o
s
itio
n
:
By ap
p
l
ying
d
i
screte wav
e
let tran
sfo
r
m
(DWT) th
e ap
prox
im
a
tio
n
,
and
d
e
tail
,
coe
ffici
ents
of t
h
e c
o
ncerne
d signal
is obtaine
d
.
Here
,
are t
h
e
dec
o
m
posi
t
i
on scal
e
and coe
fficie
n
t inde
x
of eac
h
scale res
p
ectively.
Hu
rst
e
x
po
ne
n
t
com
put
at
i
o
n
(HC
)
:
T
h
e
va
ri
ance
1
⁄∑
,
is calcul
a
ted
for each s
cale
j
,
being
num
ber of a
v
ailable c
o
efficients in e
ach scale
j
. By a weighted linear re
gre
ssion
The slope
α
of
t
h
e pl
ot
l
o
g
vers
us
j
is ob
tain
ed
an
d th
e
Hu
rst
p
a
ram
e
ter is co
mp
u
t
ed
u
s
ing
t
h
e relatio
n
s
(5
)
1
α
/2
(6
)
whe
r
e
is a co
nstan
t
.
Hurst pa
ram
e
ter is com
puted from
each segments of em
otional s
p
eec
h
si
gnal a
n
d from
the entire si
gnal.
pH
v
ect
o
r
co
m
p
o
s
ition
is
do
ne b
y
tak
i
n
g
(J+
1
)
val
u
es
o
f
,
,……,
].
LP_V
Q
C fe
atu
r
es
Speec
h sam
p
le can
be a
p
proxim
a
ted as a linear
co
m
b
i
n
atio
n of
p
r
ev
iou
s
sam
p
les in
LPC as
descri
bed
by
t
h
e rel
a
t
i
o
n
,
1
2
⋯
(7
)
Initially each
utterance
of e
m
otional signa
l
is se
gm
ented into
fram
e
s to m
a
ke a statistically non-
st
at
i
onary
si
g
n
a
l
t
o
nearl
y
st
at
i
onary
fol
l
o
wed
by
a
Ha
m
m
i
ng wi
n
d
o
w
t
o
c
o
m
p
ens
a
t
e
for
sha
r
p
bo
u
nda
ry
d
i
scon
tinu
ities. Nex
t
to it, LPC co
efficien
t
s
are ex
tr
acted u
s
i
n
g LP an
alysis. To
app
l
y v
ector
q
u
a
n
t
izatio
n
technique s
p
ee
ch signal is divide
d into num
ber of bl
oc
ks
from
which the
m
a
xim
u
m
of each bloc
k is deri
ved
to
ge
nerate
the
code book. A code book vec
t
or
is
se
lected
from
similar s
p
eech
sam
p
les. Based on m
i
nim
u
m
di
st
ance
fr
om
t
h
e m
a
xim
u
m
si
gnal
of a
bl
o
c
k i
s
sel
ect
ed
t
o
de
vel
o
p t
h
e
co
de b
o
o
k
i
n
dex
o
f
t
h
at
par
t
i
c
ul
ar
bl
oc
k. F
o
r e
a
c
h
bl
ock a
b
ove
pr
oce
d
u
r
e i
s
re
peat
ed.
A
ppl
i
c
at
i
on o
f
vect
or
qua
nt
i
zat
i
on t
o
LPC
feat
ure
s
gi
ves
the
LP_VQC feature vectors.
3.
R
E
SU
LTS AN
D ANA
LY
SIS
A dat
a
base of
50
0 ut
t
e
ran
ces fr
om
chi
l
d
ren has been
de
vel
ope
d.
The
dat
a
base has bee
n
pre
p
are
d
fo
r
f
o
u
r
em
o
tio
n
s
b
o
r
e
d
o
m
, an
gry, sad
and
h
a
pp
y of
f
i
v
e
ch
ild
r
e
n
(
t
h
r
ee boys an
d two
g
i
r
l
s)
in
t
h
e ag
e
gr
oup
of
six
to th
irteen
with
d
u
ration
of fi
v
e
secon
d
s
.
A cro
s
s va
lid
atio
n
app
r
o
a
ch
is prov
id
ed
b
y
ran
d
o
m
ly p
a
rtiti
o
n
i
n
g
th
e in
pu
t data in
to
trai
n
i
ng
,
v
a
lid
atio
n
an
d
testin
g
sets in
t
h
e
rat
i
o
0
.
6
,
0
.
2 a
nd
0.
2
respect
i
v
el
y
.
Test
i
n
g s
e
t
has
been
use
d
t
o
fi
nd t
h
e
per
f
o
r
m
a
nce
of t
h
e P
NN cl
assi
fi
er
whi
l
e
a
d
j
u
st
m
e
nt
o
f
net
w
or
k
desi
g
n
param
e
t
e
r has
been
pe
rf
o
r
m
e
d
by
t
h
e
val
i
d
at
i
on set
.
Th
e
cl
assi
fi
er pe
rf
o
r
m
a
nce has
be
en st
udi
e
d
wi
t
h
vari
ous
s
p
re
adi
n
g
p
a
r
a
m
e
ter
o
f
0.25
,
0
.
42
an
d
1
.
1
.
A
spr
ead
co
nstan
t
o
f
1
.
1 is selected
f
o
r ou
r stud
y as i
t
p
r
ov
id
ed
m
a
x
i
m
u
m
classification a
ccuracy
for all
the four classe
s of em
otions.
Tabl
e 1 s
h
ow
s t
h
e cl
assi
fi
cat
i
on er
ro
r i
n
perce
n
t
a
ge
fo
r
fo
ur cl
asses
of em
ot
i
ons
wi
t
h
di
ffe
ren
t
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
IJEC
E V
o
l
.
5, No
. 2, A
p
ri
l
20
15
:
31
1 – 3
1
7
31
5
feature set
.
PNN classifier indicates th
e erro
r
perce
n
tage
of
pH is lo
wer t
h
an the LP
_
V
Q
C
features
fo
r
all fou
r
classes of
spee
ch em
otions.
Howe
ve
r the
a
v
e
r
age
classifi
catio
n error is arou
nd 04
.89
p
e
rcen
t lower in case
o
f
pH
vectors as shown in Ta
ble 2. The sam
e
is
also
represe
n
ted
gra
p
hically in Figure 2. The a
v
erage
cl
assi
fi
cat
i
on e
r
r
o
r i
s
s
h
ow
n i
n
Fi
g
u
r
e 3
g
r
a
phi
cal
l
y
.
The
classification e
r
r
o
r
fo
u
n
d
lo
w
e
r f
o
r c
h
ild e
m
otion
with
PNN classifier. It
h
a
s b
een
te
sted
wi
th
EMO-DB data b
a
se also
,
th
ou
gh
th
ere is n
o
av
ailab
ility o
f
chi
l
d
re
n dat
a
ba
se.
Tabl
e
1. C
l
assi
fi
cat
i
on e
r
r
o
r i
n
%
Usi
n
g
P
N
N
Feature
Angry
Bore
Sad
Happy
L
P
_VQC
14.
17
17.
09
12.
43
16.
87
pH 09.
17
07.
19
12.
35
11.
25
Table
2.
Avera
g
e Classification error i
n
%
Using PNN
Feature
Average % Classi
fication error
L
P
_VQC
15.
14
pH 10.
25
Figure
2. Perce
n
tage cl
assi
fication e
r
ror for differe
nt em
otional s
p
eech
Fi
gu
re
3.
Perce
n
t
a
ge a
v
e
r
age
cl
assi
fi
cat
i
on e
r
r
o
r
f
o
r
va
ri
o
u
s
em
ot
i
onal
spe
ech
4.
CO
NCL
USI
O
N
pH
vect
or
feat
ures
p
r
o
v
e
t
o
be m
o
re r
o
bu
s
t
t
h
an
LP_VQC
spectral feat
ures
. It
can
be
ap
plied
fo
r
i
ndi
vi
dual
em
ot
i
o
n
f
o
r
f
u
t
h
er ve
ri
fi
cat
i
o
n.
B
y
appl
y
i
n
g
t
h
e p
r
o
p
o
se
d
m
e
t
hod,
t
h
e i
m
port
a
nce
of
feat
ure
s
0
2
4
6
8
10
12
14
16
18
Angry
B
ore
S
ad
H
a
ppy
14.
17
17.
09
12.
43
16.
87
9.
17
7.
19
12.
35
11.
25
% Classification error
Different speech
em
otions
LP_VQ
C
pH
0
2
4
6
8
10
12
14
16
LP_VQ
C
pH
15.
14
10.
25
% Average
classification
error
Di
fferent
feat
ure vect
ors for LP_VQC
and
pH
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Classification of
Emotional
Speech of C
h
ildren
Using
Pr
obabilistic Neural Network
(Mih
ir NM)
31
6
co
u
l
d
b
e
ev
al
u
a
ted alon
g with
th
e
p
e
rfo
r
man
ce of t
h
e
classifier. Th
e resu
lts will
be h
e
l
p
fu
l for t
h
e later
research
o
n
the em
o
tio
n
classificatio
n
.
Our wo
rk
is i
n
ten
d
e
d
t
o
recogn
ize ev
ery d
a
y activ
ities b
a
sed
o
n
em
ot
i
onal
si
g
n
a
l
s
. In t
h
i
s
pa
per a
num
ber
of
di
ffe
re
nt
t
echni
que
s we
re
i
nvest
i
g
at
e
d
and a
p
pl
i
e
d as
feat
ure
ext
r
act
o
r
s a
n
d
cl
assi
fi
cat
i
on
m
e
t
hod
was
u
s
ed a
n
d
c
o
m
p
ared.
Each
cl
assi
fi
cat
i
on
st
ep al
so
bec
o
m
e
s sim
p
l
e
r
and acc
urate a
s
it proce
sses
a lim
ited num
ber of
features
and classe
s.
Whe
n
desi
gni
ng t
h
ese system
s, it is
very
i
m
port
a
n
t
t
o
det
ect
si
t
u
at
i
ons t
h
at
m
i
ght
affect
sub
s
eq
ue
nt
pe
rf
orm
a
nce. I
n
fut
u
re
, au
di
o
-
bas
e
d
classification
schem
e
can be tested with the de
ve
l
o
p
m
ent
of a
u
t
o
nom
ous sy
st
e
m
s
t
o
reco
g
n
i
ze. Th
e
per
f
o
r
m
a
nce of
ne
wl
y
p
r
o
p
o
se
d
feat
ure
set
w
a
s com
p
ared
a
n
d
f
o
un
d t
o
be
effect
i
v
e
i
n
m
a
jo
ri
t
y
of
t
h
e
ca
ses.
R
e
duct
i
o
n
of f
eat
ure set
eve
n
an
ot
he
r i
m
port
a
nt
t
echni
qu
e can be t
e
st
e
d
f
o
r t
h
e p
r
o
p
o
se
d cl
assi
fi
ed
. Thes
e
m
a
y
be ke
pt
f
o
r f
u
t
u
re
w
o
r
k
.
REFERE
NC
ES
[1]
N. Fragopanago
s, J.G.
Tay
l
or
. “Emotion recognition in
human–computer interaction”.
Neural N
e
tworks
. vol. 18, p
p
.
389-405, Mar
.
2
005.
[2]
D. Ververid
is,
C. Kotropoulos. “
E
m
o
tional sp
eech
r
ecogn
itio
n: Resourc
e
s, f
eatur
es and m
e
thods”.
Sp
eec
h
Communication, Elsevier
. vol
. 48
, no
. 9
,
pp
. 1162
-1181, Apr. 200
6.
[3]
M.E. A
y
adi, M.S. Kamel,
and F. Karray
.
“Survey
on speech reco
gniti
on: R
e
sources, feat
ures and
methods”.
Pat
t
e
r
n
Recogn
ition
. vo
l. 44
, pp
. 572-58
7, Mar
.
2011
.
[4]
B. S
c
huller
et.
a
l
.
“
T
im
ing Level
s
in S
e
gm
ent-Based S
p
eech Em
otion Recogn
ition
”
.
ICSLP, INTERSPEECH 2006
.
pp. 1818-1821
,
23-27 Sept. Pen
n
s
y
lvan
ia 2006.
[5]
E. Hurst. “Long-
term storage
cap
acity
of
reservoirs”.
T
r
ans
. Am
er
. Soc
.Ci
v
il
Eng
.
pp. 770-799
, Ap
r. 1951
.
[6]
T. Higu
chi.
“
A
pproach to
an
irre
gular t
i
m
e
s
e
ri
es
on the b
a
s
i
s
of t
h
e fra
cta
l
th
eor
y
”.
Physics D
. vol. 31, pp
. 277-28
3,
1988.
[7]
Ricardo
S
a
nt’An
a
, Ros
â
ng
el
a Co
elho
,
and
Abrah
a
m Alcaim. “Text-Independ
en
t
Speaker R
ecogn
ition B
a
sed on
the
Hurst Param
e
ter
and th
e Mult
id
im
ensiona
l Fractional Brownian
Motion Model”.
I
EEE Transactions on
Audio
,
Speech
, and
Language Proc
essing
. vol. 14
, no
. 3
,
pp.931-940, May
2006.
[8]
L
.
Z
a
o
e
t
.
a
l.
“T
ime
-Fre
quency
Featur
e and AMS-GMM
Mask fo
r Acoustic Emotion Classificatio
n”.
IEEE Signal
Processing Letters
. vol. 21, no. 5, pp.620-624
, May
2014.
[9]
Thomas F. Quatieri. “
Disc
re
te
-Time
Spe
ec
h S
i
gna
l Processing
”. Prenti
ce-Hal
l
, 3rd
edition
.
1996
.
[10]
H.K. Palo, Mihir Naray
a
n Mohanty
,
Ma
hesh Chandra. “Design of Neural
Network Model for
Emotional Speech
Recognition”.
A
r
tificial Int
e
ll
igence and Evo
l
uti
onary
Algorithm
s
in Engineerin
g Systems
. vol.
325, pp. 291-30
0,
Springer India,
Nov. 2014.
[11]
H.K. Palo, Mihir Naray
a
na Mohanty
,
Mahesh C
h
andra. “Novel
Feature Ex
traction Techniqu
e for Child Emotion
Recognition”.
I
n
ternational Co
nference on Electrical,
Elec
tro
n
ics, Signals
,
Communication
and Optimization
(
EESCO)
-
2015, IEEE
. Jan
.
2015
.
[12]
H.K. Palo, Mih
i
r Naray
a
n Mohanty
,
Ma
hesh Ch
andra. “Use of Different Fe
atur
es for Emotion
Recognition Using
MLP Network”.
Advances in
Intelligen
t syst
ems and computing
. v
o
l. 332
, pp
.7-15, Springer In
dia,
Jan. 2015
.
[13]
K
w
on, O
.
W
., Chan, K
., H
a
o, J
.
, Lee
,
T.W
.
“
Emotion recognit
i
on by speech signal
s
”. In: Proc. Inte
rspeech. pp
. 125–
128, 2003
.
[14]
Batlin
er, A
., Fis
c
her,
K., Huber
,
R., Spilk
er,
J.,
No
¨
th
, E
.
“
Desperately seeking
emoti
ons: acto
rs, wizards, and
human beings
”
.
In: Proc. ISCA
W
o
rkshop on Speech
and
Em
otio
n,
Newc
astle
, N
o
rthern Ir
el
and.
pp. 195–200
, 20
00.
[15]
S
h
am
i, M., Ve
rh
elst, W
.
“
A
utomati
c
c
l
assific
a
tio
n of e
xpr
essiven
e
ss in speech
:
a
m
u
lti-corpus stu
d
y
”
.
In
: Mu
¨
l
l
er
,
C. (Ed.)
,
Speak
e
r
Classification
II, Lecture Not
e
s in
Computer Scien
ce/
Artif
icia
l
Intell
igen
ce
. vol. 4441. Springer
,
Heidelb
e
rg–Berlin–New York. p
p
. 43–56
, 2007
.
[16]
Chuang,
Z.J., W
u
, C.H
.
“
Emotio
n recognition us
ing acousti
c
features and te
xtual conten
t
”. In: Proc. ICME,
Taip
ei,
Taiwan
, pp
. 53–
56. 2004
.
[17]
T.
I
liou
et
.a
l.
“
C
lassific
a
tion
o
n
Speech Em
otion Recognition
-A Com
p
arative
Study
”.
International Journal on
Advances in
Life Sciences
. vol 2
no 1 & 2, pp. 18
-28 , 2010.
[18]
Nerm
ine Ahm
e
d Hend
y
and Ha
nia Farag
.
“
E
m
o
tion Recogn
ition
Using Neural Network: A Com
p
arative Stud
y”.
World Acad
emy
of Science,
Eng
i
neering and
Technology
. vol.7. p
p
. 1149-1155
, M
a
r. 2013
.
[19]
M.M. Javidi an
d E.F. Roshan.
“
S
peech Em
otion Recogn
ition b
y
Using Com
b
inations of C5.0, Neural Network
(NN), and Support Vector Machines
(SVM)
Classifi
cation Methods”.
Jour
nal of mathematics and computer
Scien
c
e
. vol. 6
,
pp. 191-200
, Ap
r. 2013
.
[20]
M
i
hir Nara
yan
M
ohant
y, A Routra
y. “
M
achin
e
Learing Approa
ch for Em
otiona
l S
p
eech Clas
s
i
f
i
ca
tion”
.
SEMC
O-
14
, Book
chap
ter, Springer
/Ver
lag Ber
lin
Heidelberg, 2014
.
[21]
K.Z. Mao et.
a
l. “
P
robabilisti
c
Neural-Networ
k
St
ructure Determ
inat
ion for Pattern Classi
fica
tion”
.
IEEE
Transactions on
neural networks
. vol. 11, no. 4, p
p
. 1009-1016
. Ju
l. 2000
.
[22]
Mihir Naray
a
n
Mohanty
,
V.
ku
mar, A Routray, P. Kabi
sa
tpa
t
h
y
.
“
C
lassific
a
tio
n of Power Qua
lit
y Disturb
a
nce
s
Us
ing P
a
rzen K
e
rnels
”
.
International Journal of Emer
ging Electric Power S
y
stems
.
vol.11
,
Issue 1, ISSN (Onlin
e)
1553-779X, DOI: 10.2202
/1553-
779X.2335, pp
.
1-13, Jan
.
2010
.
[23]
Chuang, Z.J
., Wu, C.H. “
Emotio
n recognition using acoustic
features and textual content
”. In P
r
oc. ICM
E
, T
a
ipe
i
,
Taiwan
. pp
. 53-
56, 2004
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
IJEC
E V
o
l
.
5, No
. 2, A
p
ri
l
20
15
:
31
1 – 3
1
7
31
7
[24]
Mc Gillowa
y, S
., Cowie
,
R.
, Doulas-Cowie,
E
.
,
Gielen
, S., W
e
s
terdijk
, M., Str
o
eve, S. “
Appro
a
ching automatic
recognition
of
emotion from
vo
ice:
A rough
b
e
nchmark
”. In
Proc. ISCA Workshop on Speech and
Emotion
,
Newcastle, Northern Ir
eland
.
pp
. 207-212, 2000.
BIOGRAP
HI
ES OF
AUTH
ORS
Hemanta Kum
a
r Palo
has
co
m
p
leted h
i
s ‘A.
M
.I.E
.’ from
“
I
nstitute
of
Engi
neers”,
Indi
a in
1997 and h
i
s Master of
Engin
e
ering from “Bir
la
Institute of
Technolog
y
”
, Mesra, Ran
c
hi in
2011. He completed his ‘Diplo
ma in
Rail Tran
sport and Management’ from
‟
Institute of Ra
il
Transport”, India in 2003
. He is
having 20
y
e
ar
s of
experience
in the
field of Electronics
and
Communication Engineer
ing fr
om 1990 to 20
10 in
Indian Air Force and was an Assistant
Professor in Gandhi Academ
y
of Techno
log
y
and
Engin
eer
ing
,
Odisha,
in th
e Department of
ECE from 2010 to 2012. He
is the life member
of IEI
,
India and is the memberof IEEE.
Curre
ntly
he
is se
rving a
s
a
n
Assista
n
t Prof
essor in the d
e
partment of
Electronics and
Com
m
unication Engine
ering of Ele
c
troni
cs in the
Institute of Te
c
hnica
l Educa
tion
and Research,
Siksha „O
‟
Anu
s
andhan Univers
i
ty
, Bhubaneswar, Odisha, Ind
i
a.
Mih
i
r Nara
yan
Moh
a
n
t
y
is
pre
s
entl
y working a
s
an As
s
o
ciat
e Professor in the Department of
Ele
c
troni
cs and
Com
m
unication Engine
ering
,
I
n
stitute of
Te
ch
nica
l Educ
ation
and Resear
ch,
Siksha „O
‟
Anusandhan University
, Bhuban
e
swar, Odis
ha. He has published over
100 papers in
International/National Journals
and Conferen
ces
along with app
r
oximately
20
y
e
ars of teaching
experi
enc
e
. He i
s
the active m
e
m
b
er of
m
a
n
y
profes
s
i
onal s
o
cie
t
ies
like IEE
E
, IE
T, IET
E
, EM
C
& EMI
Engine
e
r
s India,
IE
(I)
, I
S
CA, ACEEE,
I
A
Eng et
c.
He h
a
s rece
ived
his M
.
Te
ch.
degr
ee
in Communication S
y
stem Engineering from th
e S
a
m
b
alpur Univers
i
t
y
,
S
a
m
b
alpur, Odis
ha
.
Now he has do
ne his Ph.D. work in Applied Si
gnal Processing. He is currently
working
as
As
s
o
ciate P
r
ofe
s
s
o
r and was
Head in
th
e D
e
pa
rtment
of Electronics and
I
n
strumentation
Engine
ering,
Ins
titut
e
of
Te
chni
c
a
l
Educa
tion
an
d Resear
ch,
Siksha O
‟
Anusandh
an University
,
Bhubaneswar,
Odisha. His area of resear
ch
in
terests inclu
d
es Applied Signal and image
Processing, Digital Sign
al/Im
a
ge Processing,
Biomedical Signal Pr
ocessing, Microwave
Communication Engineeri
ng and
Bioinformatics.
Evaluation Warning : The document was created with Spire.PDF for Python.