I
AE
S In
t
er
na
t
io
na
l J
o
urna
l o
f
Art
if
icia
l In
t
ellig
ence
(
I
J
-
AI
)
Vo
l.
14
,
No
.
4
,
A
u
g
u
s
t
2
0
2
5
,
p
p
.
2
6
8
9
~
2
7
0
0
I
SS
N:
2
2
5
2
-
8
9
3
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijai.v
14
.i
4
.
p
p
2
6
8
9
-
2
7
0
0
2689
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e.
co
m
Semi
-
a
utom
a
tic
v
o
ice compa
riso
n app
ro
a
ch using
spi
king
neura
l net
wo
rk
for forensics
K
rut
hik
a
Sid
da
na
k
a
t
t
e
G
o
p
a
la
ia
h
1,
2
,
T
risi
la
dev
i C
ha
nd
r
a
k
a
nt
N
a
g
a
v
i
1
,
P
a
ra
s
hiv
a
m
urt
hy
M
a
hes
ha
1
1
D
e
p
a
r
t
me
n
t
o
f
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
a
n
d
E
n
g
i
n
e
e
r
i
n
g
,
S
.
J
.
C
o
l
l
e
g
e
o
f
E
n
g
i
n
e
e
r
i
n
g
,
JS
S
S
c
i
e
n
c
e
a
n
d
Te
c
h
n
o
l
o
g
y
U
n
i
v
e
r
si
t
y
,
M
y
so
r
e
,
I
n
d
i
a
2
D
e
p
a
r
t
me
n
t
o
f
A
r
t
i
f
i
c
i
a
l
I
n
t
e
l
l
i
g
e
n
c
e
a
n
d
M
a
c
h
i
n
e
L
e
a
r
n
i
n
g
,
V
i
d
y
a
v
a
r
d
h
a
k
a
C
o
l
l
e
g
e
o
f
En
g
i
n
e
e
r
i
n
g
,
M
y
so
r
e
,
I
n
d
i
a
.
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
J
u
l
16
,
2
0
2
4
R
ev
is
ed
Mar
30
,
2
0
2
5
Acc
ep
ted
J
u
n
8
,
2
0
2
5
Th
is
p
a
p
e
r
e
x
p
lo
re
s
th
e
a
p
p
l
ica
ti
o
n
o
f
a
se
m
i
-
a
u
to
m
a
ti
c
tec
h
n
iq
u
e
u
sin
g
sp
ik
i
n
g
n
e
u
ra
l
n
e
two
rk
(S
NN
)
a
p
p
r
o
a
c
h
fo
r
fo
re
n
sic
v
o
ice
c
o
m
p
a
riso
n
(F
VC),
a
d
d
re
ss
in
g
th
e
li
m
it
a
ti
o
n
s
o
f
trad
it
io
n
a
l
m
e
th
o
d
s
th
a
t
a
re
ti
m
e
-
c
o
n
su
m
in
g
a
n
d
s
u
b
jec
ti
v
e
.
By
in
teg
ra
ti
n
g
m
a
c
h
i
n
e
lea
rn
in
g
with
h
u
m
a
n
e
x
p
e
rti
se
,
th
e
S
NN
,
w
h
ich
m
imic
s
th
e
b
ra
in
’s
p
r
o
c
e
ss
in
g
o
f
tem
p
o
ra
l
in
fo
rm
a
ti
o
n
,
is
a
p
p
l
ied
t
o
a
n
a
l
y
z
e
Au
stra
li
a
n
En
g
li
sh
v
o
ice
d
a
t
a
in
.
flac
fo
rm
a
t.
T
h
e
m
o
d
e
l
lev
e
ra
g
e
s
s
y
n
a
p
t
ic
c
o
n
n
e
c
ti
o
n
stre
n
g
t
h
s
m
o
d
ifi
e
d
b
y
sp
ik
e
ti
m
in
g
,
a
ll
o
w
in
g
fo
r
f
lex
ib
l
e
v
o
ice
fe
a
tu
re
re
p
re
se
n
tatio
n
.
P
e
rfo
rm
a
n
c
e
m
e
tri
c
s,
in
c
lu
d
in
g
c
o
n
f
u
sio
n
m
a
tri
c
e
s
a
n
d
re
c
e
iv
e
r
o
p
e
ra
ti
n
g
c
h
a
ra
c
teristic
(
ROC
)
a
n
a
ly
sis,
in
d
ica
te
th
e
m
o
d
e
l’s
a
c
c
u
ra
c
y
o
f
9
4
.
2
1
%
,
h
i
g
h
li
g
h
ti
n
g
t
h
e
e
ffe
c
ti
v
e
n
e
ss
o
f
th
e
S
NN
-
b
a
se
d
a
p
p
r
o
a
c
h
fo
r
F
VC
.
K
ey
w
o
r
d
s
:
Ar
tific
ial
n
eu
r
al
n
etwo
r
k
Dig
ital f
o
r
en
s
ics
Fo
r
en
s
ic
v
o
ice
co
m
p
ar
is
o
n
Fre
e
lo
s
s
les
s
au
d
io
co
d
ec
Sp
ik
in
g
n
eu
r
al
n
etwo
r
k
s
T
h
is i
s
a
n
o
p
e
n
a
c
c
e
ss
a
rticle
u
n
d
e
r th
e
CC B
Y
-
SA
li
c
e
n
se
.
C
o
r
r
e
s
p
o
nd
ing
A
uth
o
r
:
T
r
is
ilad
ev
i Ch
an
d
r
ak
an
t N
a
g
a
v
i
Dep
ar
tm
en
t o
f
C
o
m
p
u
ter
Scie
n
ce
an
d
E
n
g
in
ee
r
in
g
,
S.
J
.
C
o
l
leg
e
o
f
E
n
g
in
ee
r
in
g
J
SS
Scien
ce
an
d
T
ec
h
n
o
lo
g
y
Un
iv
er
s
ity
My
s
o
r
e
,
Kar
n
atak
a
5
7
0
0
0
6
,
I
n
d
ia
E
m
ail
:
tr
is
ilad
ev
i@
s
jce.
ac
.
in
1.
I
NT
RO
D
UCT
I
O
N
Dig
ital
f
o
r
en
s
ics
is
a
cr
u
cial
b
r
an
ch
o
f
f
o
r
en
s
ic
s
cien
ce
th
a
t
f
o
cu
s
es
o
n
th
e
r
ec
o
v
er
y
,
in
v
esti
g
atio
n
,
ex
am
in
atio
n
a
n
d
a
n
aly
s
is
o
f
ev
id
en
ce
f
o
u
n
d
in
d
ig
ital
d
e
v
ices
.
T
h
is
f
ield
p
lay
s
a
c
r
u
cial
p
ar
t
in
s
o
lv
in
g
cr
im
es
b
y
u
n
c
o
v
er
i
n
g
ev
i
d
en
c
e
f
r
o
m
v
ar
io
u
s
d
i
g
ital
s
o
u
r
ce
s
,
s
u
ch
as
co
m
p
u
ter
s
,
s
m
ar
tp
h
o
n
e
s
an
d
n
etwo
r
k
s
.
I
t
in
v
o
l
v
es
a
m
eticu
l
o
u
s
p
r
o
ce
s
s
to
ass
u
r
e
th
e
a
u
th
en
ticity
a
n
d
r
eliab
ilit
y
o
f
th
e
d
ata
co
lle
cted
,
wh
ich
ca
n
b
e
u
s
ed
in
leg
al
p
r
o
ce
ed
in
g
s
.
I
n
th
e
f
ield
o
f
d
ig
ital
f
o
r
en
s
ics,
f
o
r
en
s
ic
v
o
ice
c
o
m
p
ar
is
o
n
(
FVC
)
is
a
s
p
ec
ialized
b
r
an
ch
f
o
cu
s
ed
o
n
ex
am
i
n
in
g
th
e
d
if
f
er
en
t
v
o
ice
r
ec
o
r
d
in
g
s
.
T
h
e
p
r
im
ar
y
o
b
jectiv
e
o
f
F
VC
i
s
to
id
en
tify
a
s
u
s
p
ec
t
b
y
co
m
p
a
r
in
g
a
tr
ac
e
v
o
ice
s
am
p
le
with
k
n
o
wn
s
am
p
les,
p
r
o
v
id
in
g
an
ev
i
d
en
ce
-
b
ased
ass
ess
m
en
t
o
f
wh
eth
er
th
e
r
ec
o
r
d
in
g
s
co
m
e
f
r
o
m
th
e
s
am
e
s
p
ea
k
e
r
f
o
r
leg
a
l
o
r
in
v
esti
g
ativ
e
p
u
r
p
o
s
es
[
1
]
–
[
4
]
.
T
r
ad
itio
n
ally
,
FVC
h
as
r
elied
h
ea
v
ily
o
n
m
an
u
al
an
aly
s
is
,
in
wh
ich
t
r
ain
ed
ex
p
e
r
ts
lis
ten
to
an
d
c
o
m
p
ar
e
v
o
ice
s
am
p
les.
Alth
o
u
g
h
th
is
m
eth
o
d
ca
n
b
e
ef
f
ec
ti
v
e,
it
is
o
f
te
n
tim
e
-
co
n
s
u
m
in
g
an
d
p
r
o
n
e
to
h
u
m
an
er
r
o
r
an
d
b
ias.
T
o
ad
d
r
ess
th
ese
lim
itatio
n
s
,
t
h
er
e
h
as
b
ee
n
a
s
h
if
t
to
war
d
s
s
em
i
-
au
to
m
atic
ap
p
r
o
ac
h
es
th
at
in
teg
r
ate
h
u
m
a
n
ex
p
er
tis
e
with
m
ac
h
in
e
lear
n
i
n
g
an
d
d
ee
p
lear
n
i
n
g
tec
h
n
iq
u
es.
T
h
ese
ap
p
r
o
ac
h
es
ai
m
to
en
h
an
ce
b
o
t
h
th
e
ef
f
icien
cy
an
d
ac
c
u
r
ac
y
o
f
v
o
ice
co
m
p
ar
is
o
n
s
,
r
e
d
u
cin
g
r
elian
ce
o
n
s
u
b
jectiv
e
ju
d
g
m
e
n
t
wh
ile
im
p
r
o
v
in
g
r
eliab
ilit
y
in
f
o
r
e
n
s
ic
in
v
esti
g
atio
n
s
[
5
]
–
[
7
]
.
T
h
e
m
o
s
t
ch
allen
g
in
g
task
in
FVC
i
s
an
aly
zin
g
au
d
io
s
am
p
l
es
f
r
o
m
b
o
th
tr
ac
e
a
n
d
k
n
o
wn
s
o
u
r
ce
s
to
d
eter
m
in
e
th
e
s
im
ilar
ity
o
r
d
is
s
im
ilar
ity
in
th
e
s
u
s
p
ec
t's
v
o
ice,
wh
ich
is
cr
u
cial
f
o
r
i
d
en
tify
in
g
a
n
y
p
ast
cr
im
i
n
al
ac
tiv
ity
.
W
h
ile
ea
r
lier
s
tu
d
ies
h
av
e
ex
p
lo
r
e
d
th
e
im
p
ac
t
o
f
s
em
i
au
to
m
atic
ap
p
r
o
a
ch
ar
tific
ial
n
eu
r
al
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
14
,
No
.
4
,
Au
g
u
s
t
20
25
:
2
6
8
9
-
2
7
0
0
2690
n
etwo
r
k
(
ANN)
a
co
m
p
u
tatio
n
al
m
o
d
el
tr
ain
e
d
to
id
en
tif
y
an
d
class
if
y
s
p
ea
k
e
r
s
b
ased
o
n
p
atter
n
s
in
a
u
d
io
f
ea
tu
r
es
ex
tr
ac
ted
f
r
o
m
s
p
ee
c
h
s
am
p
les
.
An
AN
N
is
m
ad
e
u
p
o
f
n
etwo
r
k
e
d
n
o
d
es
o
r
n
e
u
r
o
n
s
,
th
at
p
r
o
ce
s
s
in
f
o
r
m
atio
n
s
im
ilar
to
b
io
lo
g
ical
n
eu
r
al
n
etwo
r
k
s
,
th
ey
h
av
e
n
o
t
ex
p
licitly
ad
d
r
ess
well.
L
ik
ewise,
i
ts
in
f
lu
en
ce
o
n
t
h
e
p
r
o
p
o
s
ed
s
tu
d
y
u
tili
ze
s
s
p
ik
in
g
n
eu
r
al
n
et
wo
r
k
(
SNN
)
,
it
o
f
f
er
s
an
in
n
o
v
ativ
e
m
eth
o
d
f
o
r
v
o
ice
p
atter
n
r
ec
o
g
n
itio
n
.
T
h
er
e
is
lo
t
o
f
o
n
g
o
in
g
r
esear
ch
in
ar
tific
ial
in
tellig
en
ce
,
m
ac
h
in
e
lear
n
in
g
,
an
d
n
eu
r
o
s
cien
ce
r
elate
d
to
SNN
.
I
n
o
r
d
e
r
to
co
m
p
r
eh
e
n
d
b
r
ain
f
u
n
ctio
n
an
d
cr
e
ate
in
n
o
v
ativ
e
co
m
p
u
tatio
n
al
m
o
d
els,
a
lar
g
e
n
u
m
b
er
o
f
ac
ad
em
ic
in
s
titu
tio
n
s
an
d
r
esear
ch
lab
s
in
v
esti
g
ate
SNN
.
T
h
e
f
ield
o
f
FVC
h
as
lim
ited
u
s
e
o
f
SNN
in
th
e
ex
is
tin
g
wo
r
k
s
.
T
h
is
r
esear
ch
aim
s
to
ad
d
r
ess
th
is
g
ap
b
y
p
r
o
p
o
s
in
g
a
SNN
m
o
d
el
th
at
ass
is
t
in
v
esti
g
ativ
e
ag
en
cies
in
th
e
ef
f
ec
tiv
e
id
e
n
tific
atio
n
o
f
t
h
e
s
u
s
p
ec
ts
.
T
h
e
S
NN
i
n
s
p
ir
ed
b
y
th
e
n
eu
r
al
ar
ch
itectu
r
e
o
f
th
e
b
io
l
o
g
ical
b
r
ain
,
p
r
o
ce
s
s
es
in
f
o
r
m
atio
n
th
r
o
u
g
h
d
is
cr
ete
s
p
ik
es,
m
im
ick
in
g
th
e
way
n
eu
r
o
n
s
co
m
m
u
n
icate
.
T
h
e
t
im
e
-
v
ar
y
in
g
p
r
o
ce
s
s
in
g
m
ak
e
s
SNN
p
ar
ticu
lar
ly
well
-
s
u
ited
with
ad
ju
s
tab
le
th
r
esh
o
ld
f
o
r
an
al
y
zin
g
c
o
m
p
l
ex
v
o
ice
r
ec
o
r
d
in
g
s
f
o
r
FVC
.
T
h
e
m
o
tiv
atio
n
o
f
ad
o
p
tin
g
th
e
SNN
f
o
r
FVC
is
d
esig
n
ed
f
o
r
task
s
in
v
o
lv
in
g
s
im
ilar
ity
o
r
d
is
s
im
ilar
ity
esti
m
atio
n
[
8
]
–
[
1
2
]
.
T
h
e
k
e
y
co
n
tr
ib
u
tio
n
s
o
f
th
is
r
esear
ch
ar
e
o
u
tlin
ed
as f
o
llo
w
s
:
−
T
o
p
r
e
-
p
r
o
ce
s
s
in
p
u
t
d
ata:
r
em
o
v
e
b
ac
k
g
r
o
u
n
d
n
o
is
e
f
o
r
a
cl
ea
r
s
p
ee
ch
r
ec
o
r
d
in
g
.
−
T
o
ex
tr
ac
t
a
n
d
c
o
m
p
ar
e
f
ea
tu
r
es:
u
s
e
an
ad
ju
s
tab
le
th
r
esh
o
ld
SNN
to
co
m
p
a
r
e
v
o
ice
s
am
p
les
an
d
an
aly
z
e
s
im
ilar
ities
o
r
d
if
f
er
en
ce
s
.
−
T
o
co
n
d
u
ct
p
e
r
f
o
r
m
an
ce
a
n
al
y
s
is
:
e
v
alu
ate
th
e
s
y
s
tem
u
s
in
g
a
co
n
f
u
s
io
n
m
atr
ix
an
d
its
e
x
ten
d
ed
m
etr
ics
f
o
r
co
m
p
r
eh
e
n
s
iv
e
ass
e
s
s
m
en
t
.
−
T
o
p
er
f
o
r
m
co
m
p
ar
ativ
e
an
aly
s
is
:
c
o
m
p
ar
e
th
e
r
esu
lts
o
f
th
e
p
r
o
p
o
s
ed
r
esear
ch
wo
r
k
with
ex
is
tin
g
s
tu
d
ies
to
h
ig
h
lig
h
t im
p
r
o
v
em
en
ts
an
d
co
n
tr
ib
u
tio
n
s
.
T
h
e
o
b
jectiv
e
o
f
th
is
wo
r
k
is
to
id
en
tify
s
u
s
p
ec
ts
in
FVC
u
s
in
g
SNN.
I
n
p
u
t
v
o
ice
s
am
p
les
ar
e
co
lle
cted
an
d
an
aly
ze
d
to
d
e
tect
s
im
ilar
itie
s
an
d
d
is
s
im
i
l
ar
ities
b
etwe
en
v
o
ices.
T
h
es
e
s
am
p
les
u
n
d
er
g
o
p
r
e
-
p
r
o
ce
s
s
in
g
with
a
s
tatio
n
ar
y
n
o
is
e
r
ed
u
ctio
n
alg
o
r
ith
m
t
o
en
h
an
ce
clar
ity
.
T
h
e
p
r
e
-
p
r
o
ce
s
s
ed
v
o
ices
ar
e
th
en
co
n
v
er
ted
in
to
d
is
cr
ete
s
p
ik
es,
m
im
ick
in
g
n
eu
r
o
n
al
co
m
m
u
n
icatio
n
.
T
h
is
tim
e
-
d
e
p
en
d
en
t
p
r
o
ce
s
s
in
g
m
ak
es
SNNs
p
ar
ticu
lar
ly
wel
l
-
s
u
ited
f
o
r
an
aly
zin
g
co
m
p
lex
v
o
ice
r
ec
o
r
d
in
g
s
in
FVC
,
with
an
ad
ju
s
tab
le
th
r
esh
o
ld
to
im
p
r
o
v
e
ac
cu
r
a
cy
.
T
h
e
ac
cu
r
ac
y
o
f
th
e
s
y
s
tem
is
ev
alu
ated
u
s
in
g
a
co
n
f
u
s
io
n
m
atr
ix
an
d
ex
ten
d
ed
p
er
f
o
r
m
an
ce
m
etr
i
cs.
Fin
ally
,
th
e
p
r
o
p
o
s
ed
a
p
p
r
o
ac
h
is
co
m
p
ar
ed
with
ex
is
tin
g
s
tu
d
ies
to
d
em
o
n
s
tr
ate
its
ef
f
ec
tiv
en
ess
.
T
h
e
s
tr
u
ctu
r
e
o
f
th
e
p
a
p
er
is
o
r
g
a
n
ized
as
f
o
llo
ws:
t
h
e
r
e
v
iew
o
f
liter
atu
r
e
is
g
iv
e
n
in
s
ec
tio
n
2
.
Sectio
n
3
d
etails
d
ata
co
l
lectio
n
an
d
ex
p
er
im
en
tal
s
etu
p
.
Sect
io
n
4
d
is
cu
s
s
es
FVC
u
s
in
g
a
S
NN
.
Ad
d
itio
n
ally
,
s
ec
tio
n
5
d
escr
ib
es
th
e
r
esu
lt
a
n
aly
s
is
an
d
d
is
cu
s
s
io
n
o
b
tain
e
d
f
r
o
m
SNN.
Fu
r
th
e
r
m
o
r
e,
th
e
s
ec
tio
n
6
p
r
esen
ts
a
co
m
p
ar
is
o
n
with
ex
is
tin
g
wo
r
k
.
T
h
e
p
ap
er
is
co
n
clu
d
e
d
in
s
ec
tio
n
7
.
2.
L
I
T
E
R
AT
U
RE
R
E
VI
E
W
R
esear
ch
in
FVC
h
as
ev
o
lv
ed
o
v
e
r
th
e
d
ec
a
d
es,
in
t
r
o
d
u
cin
g
v
ar
i
o
u
s
m
eth
o
d
o
l
o
g
ies
f
o
r
au
th
en
ticatin
g
an
d
v
er
if
y
i
n
g
s
p
ee
ch
.
W
ith
th
e
r
is
e
o
f
d
ig
ital te
ch
n
o
lo
g
y
,
v
o
ice
c
o
m
p
ar
is
o
n
p
lay
s
a
cr
u
cial
r
o
le
in
f
o
r
en
s
ics.
T
h
is
r
ev
iew
ex
am
in
es
k
ey
FVC
m
eth
o
d
s
,
f
o
cu
s
in
g
o
n
a
s
e
mi
-
au
to
m
atic
ap
p
r
o
ac
h
u
s
in
g
SNNs
k
n
o
wn
f
o
r
a
n
aly
zin
g
c
o
m
p
lex
s
p
ee
ch
p
atter
n
s
.
I
t
h
ig
h
lig
h
ts
th
e
n
ee
d
f
o
r
en
h
an
ce
d
v
o
ice
p
atter
n
r
ec
o
g
n
itio
n
,
d
r
iv
in
g
th
e
p
r
o
p
o
s
ed
SNN
-
b
ased
r
esear
ch
to
im
p
r
o
v
e
s
u
s
p
ec
t
id
en
tific
atio
n
th
r
o
u
g
h
v
o
ice
s
im
ilar
ity
co
m
p
ar
is
o
n
s
[
1
3
]
–
[
1
8
]
.
Sev
er
al
s
t
u
d
i
es
[
1
9
]
–
[
2
5
]
h
av
e
d
em
o
n
s
t
r
at
e
d
t
h
e
ef
f
ec
t
iv
en
ess
o
f
SN
Ns
f
o
r
s
p
a
ti
o
te
m
p
o
r
al
p
att
er
n
class
i
f
i
ca
t
io
n
.
M
o
r
ales
et
a
l
.
[
25
]
d
ev
el
o
p
e
d
a
m
u
l
til
ay
e
r
SNN
o
n
t
h
e
S
p
iNN
a
k
e
r
p
l
a
tf
o
r
m
,
u
s
i
n
g
le
a
k
y
in
t
eg
r
ate
-
a
n
d
-
f
i
r
e
n
eu
r
o
n
s
an
d
f
ir
i
n
g
r
at
e
-
b
as
e
d
al
g
o
r
i
t
h
m
s
t
o
t
r
ai
n
i
n
t
er
-
l
ay
er
co
n
n
e
ct
io
n
s
.
T
h
e
n
etw
o
r
k
ac
h
ie
v
e
d
o
v
er
8
5
%
h
it r
a
te
p
er
cl
ass
wi
th
a
s
i
g
n
al
-
to
-
n
o
is
e
r
a
tio
(
S
NR
)
a
b
o
v
e
3
d
B
,
d
e
m
o
n
s
tr
a
ti
n
g
its
e
f
f
ec
ti
v
e
co
n
f
i
g
u
r
a
ti
o
n
a
n
d
tr
ai
n
i
n
g
m
et
h
o
d
.
Si
m
il
ar
ly
,
W
u
et
a
l
.
[
19
]
p
r
o
p
o
s
e
d
t
h
e
s
el
f
-
o
r
g
a
n
iz
in
g
m
a
p
(
SOM
)
-
S
NN
,
a
b
i
o
l
o
g
ic
all
y
in
s
p
i
r
e
d
a
r
ti
f
i
cia
l
s
p
i
k
i
n
g
ci
r
c
u
it
(
ASC
)
f
r
am
e
wo
r
k
c
o
m
b
in
in
g
a
n
u
n
s
u
p
er
v
is
ed
SOM
wi
th
a
n
ev
en
t
-
b
as
e
d
S
NN
to
c
lass
i
f
y
s
p
at
io
te
m
p
o
r
a
l
p
a
tte
r
n
s
.
O
n
t
h
e
r
e
al
wo
r
l
d
c
o
m
p
u
ti
n
g
p
ar
tn
er
s
h
i
p
(
RWCP
)
a
n
d
T
I
D
I
G
I
T
S
d
atas
ets
,
SOM
-
SN
N
s
h
o
we
d
r
o
b
u
s
t
n
ess
to
n
o
is
e
an
d
e
ar
ly
d
ec
is
io
n
-
m
a
k
i
n
g
,
a
c
h
ie
v
i
n
g
9
7
.
4
0
%
a
n
d
9
9
.
6
0
%
ac
c
u
r
a
cy
,
r
esp
ec
ti
v
el
y
.
W
u
et
a
l
.
[
20
]
in
v
esti
g
ated
SNNs
f
o
r
ac
o
u
s
tic
m
o
d
elin
g
in
lar
g
e
-
v
o
ca
b
u
lar
y
a
u
to
m
at
ic
s
p
ee
ch
r
ec
o
g
n
itio
n
(
ASR
)
,
ac
h
iev
in
g
co
m
p
etitiv
e
ac
cu
r
ac
ies
with
o
n
ly
10
-
tim
e
s
tep
s
an
d
0
.
6
8
tim
es
th
e
s
y
n
ap
tic
o
p
er
atio
n
s
p
er
au
d
io
f
r
am
e.
T
h
is
co
m
b
in
atio
n
o
f
en
e
r
g
y
-
e
f
f
icien
t
n
eu
r
o
m
o
r
p
h
ic
h
ar
d
wa
r
e
an
d
d
ee
p
SNNs
s
h
o
ws p
o
ten
tial f
o
r
ASR
o
n
m
o
b
ile
an
d
e
m
b
ed
d
ed
d
e
v
ices,
with
r
ep
o
r
ted
ac
cu
r
ac
ies o
f
1
8
.
7
% a
n
d
3
6
.
9
%.
Au
g
e
e
t
a
l
.
[
21
]
e
x
p
lo
r
e
d
SNNs
f
o
r
en
er
g
y
-
e
f
f
ici
e
n
t
e
d
g
e
d
e
v
ic
es,
em
p
h
asi
zi
n
g
s
m
al
l
-
s
ca
l
e
n
e
u
r
o
m
o
r
p
h
i
c
i
m
p
le
m
e
n
t
ati
o
n
s
.
B
y
i
n
t
eg
r
a
ti
n
g
r
es
o
n
ati
n
g
n
e
u
r
o
n
s
as
t
h
e
SNN
i
n
p
u
t
l
a
y
e
r
f
o
r
e
n
d
-
to
-
e
n
d
o
n
li
n
e
a
u
d
i
o
c
lass
i
f
i
ca
t
io
n
,
t
h
e
y
en
ab
le
d
lo
w
-
p
o
we
r
c
o
n
ti
n
u
o
u
s
au
d
i
o
s
tr
ea
m
an
al
y
s
is
.
T
h
e
ap
p
r
o
ac
h
,
e
v
al
u
at
ed
u
s
i
n
g
a
k
e
y
w
o
r
d
s
p
o
tti
n
g
b
e
n
ch
m
a
r
k
,
d
e
m
o
n
s
t
r
a
te
d
s
tr
o
n
g
ac
c
u
r
a
cy
u
s
i
n
g
m
el
-
f
r
eq
u
en
cy
s
p
ec
t
r
a
l
f
e
at
u
r
es.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
S
emi
-
a
u
to
ma
tic
vo
ice
c
o
mp
a
r
i
s
o
n
a
p
p
r
o
a
ch
u
s
in
g
s
p
ikin
g
n
e
u
r
a
l
…
(
K
r
u
th
ika
S
id
d
a
n
a
k
a
tte
Go
p
a
la
ia
h
)
2691
Fu
r
th
e
r
,
Mu
k
h
o
p
a
d
h
y
ay
et
a
l
.
[
2
2
]
s
tu
d
i
ed
h
u
m
a
n
f
o
o
ts
te
p
s
o
u
n
d
class
if
ica
ti
o
n
i
n
n
at
u
r
al
en
v
i
r
o
n
m
e
n
ts
u
s
in
g
a
w
ir
eless
s
en
s
o
r
n
etw
o
r
k
(
W
SN)
f
o
r
s
ec
u
r
it
y
s
u
r
v
e
illa
n
ce
.
B
y
em
p
l
o
y
i
n
g
an
SN
N
wit
h
s
im
p
le
t
im
e
-
d
o
m
ai
n
f
e
at
u
r
es,
t
h
e
y
ai
m
e
d
t
o
cr
ea
t
e
e
n
e
r
g
y
-
e
f
f
ic
ie
n
t
,
co
s
t
-
ef
f
ec
ti
v
e
s
e
n
s
o
r
n
o
d
es
.
Si
m
u
lat
io
n
s
s
h
o
w
ed
s
i
g
n
i
f
ic
a
n
t
p
o
we
r
s
a
v
i
n
g
s
wit
h
a
n
a
lo
g
S
NNs,
d
es
p
it
e
m
i
n
o
r
ac
cu
r
a
c
y
l
o
s
s
m
i
ti
g
at
e
d
b
y
r
ed
u
n
d
a
n
c
y
an
d
m
aj
o
r
it
y
v
o
t
in
g
.
Fu
t
u
r
e
r
esea
r
ch
m
a
y
f
o
c
u
s
o
n
l
o
w
-
p
o
w
er
f
e
at
u
r
e
ex
tr
ac
t
io
n
f
o
r
s
u
r
v
eil
la
n
c
e
s
y
s
t
em
s
[
2
3
]
.
E
ar
lier
,
Yam
az
ak
i
et
a
l
.
[
23
]
h
ig
h
lig
h
ted
th
e
lim
itatio
n
s
o
f
d
ee
p
n
eu
r
al
n
etwo
r
k
s
,
s
u
ch
as
h
ig
h
co
m
p
u
tatio
n
al
co
s
ts
an
d
en
er
g
y
co
n
s
u
m
p
tio
n
in
d
r
o
n
es
an
d
s
elf
-
d
r
iv
in
g
v
e
h
icles.
T
h
ey
p
r
o
p
o
s
ed
SNNs
a
s
ef
f
icien
t
alter
n
ativ
es,
m
im
ick
i
n
g
b
io
lo
g
ical
n
eu
r
o
n
s
th
r
o
u
g
h
s
p
ar
s
ity
an
d
tem
p
o
r
al
co
d
i
n
g
.
T
h
e
p
ap
er
r
e
v
iews
b
io
lo
g
ical
n
eu
r
o
n
th
e
o
r
ies,
s
p
i
k
e
-
b
ased
n
eu
r
o
n
m
o
d
els,
SN
N
tr
ain
in
g
m
eth
o
d
s
,
an
d
ap
p
licatio
n
s
in
co
m
p
u
ter
v
is
io
n
an
d
r
o
b
o
tics
,
o
f
f
er
in
g
f
u
tu
r
e
r
esear
ch
in
s
ig
h
ts
.
Kh
o
lk
in
et
a
l
.
[
24
]
d
is
cu
s
s
ed
th
e
r
is
in
g
i
n
ter
est
in
SNNs
d
esp
ite
ch
alle
n
g
es
with
v
o
n
Neu
m
an
n
ar
ch
itectu
r
es,
n
o
tin
g
th
at
h
a
r
d
war
e
ad
v
an
ce
m
en
ts
n
o
w
en
a
b
le
p
r
ac
tical
SNN
ap
p
licatio
n
s
.
T
h
eir
co
m
p
ar
is
o
n
o
f
SNN
an
d
ANN
r
eser
v
o
ir
c
o
m
p
u
tin
g
ar
c
h
itectu
r
es
u
s
in
g
th
e
R
C
Net
lib
r
ar
y
s
h
o
wed
S
NNs
h
ad
lo
n
g
er
r
u
n
tim
es
b
u
t
s
u
p
er
io
r
class
if
ica
tio
n
,
p
a
r
ticu
lar
ly
f
o
r
co
m
p
lex
d
atasets
lik
e
in
d
u
s
tr
ial
s
en
s
o
r
f
au
lts
.
I
n
b
all
b
ea
r
i
n
g
d
iag
n
o
s
is
,
SNNs
o
u
tp
er
f
o
r
m
e
d
ANNs,
wh
ich
ac
h
iev
e
d
o
n
l
y
6
1
%
ac
c
u
r
ac
y
.
T
ab
le
1
s
h
o
ws
r
ev
iews
o
f
SNN
tech
n
iq
u
es
f
o
r
s
p
ee
ch
a
n
aly
s
is
,
h
ig
h
lig
h
tin
g
th
eir
lim
ited
a
p
p
licatio
n
in
v
o
ice
an
aly
s
is
an
d
ab
s
en
ce
in
FVC
.
T
h
is
g
ap
m
o
tiv
ates
o
u
r
r
esear
ch
to
in
teg
r
ate
SNNs
f
o
r
e
n
h
a
n
ce
d
v
o
ice
a
n
aly
s
is
,
with
th
e
g
o
al
o
f
tr
a
n
s
f
o
r
m
in
g
FVC
in
leg
al
in
v
esti
g
atio
n
s
.
T
ab
le
1
.
L
iter
atu
r
e
r
ev
iew
o
f
SNN
m
eth
o
d
s
C
i
t
a
t
i
o
n
D
a
t
a
s
e
t
M
e
t
h
o
d
O
v
e
r
v
i
e
w
R
e
s
u
l
t
s i
n
(
%)
M
o
r
a
l
e
s
e
t
a
l
.
[
25
]
P
u
r
e
t
o
n
e
sam
p
l
e
s
S
N
N
,
S
p
i
N
N
a
k
e
r
R
o
b
u
st
n
e
ss
,
e
f
f
i
c
i
e
n
c
y
i
n
t
h
e
n
e
u
r
o
mo
r
p
h
i
c
f
i
e
l
d
85
W
u
e
t
a
l
. [
19
]
R
W
C
P
&
TI
D
I
G
I
TS
D
i
sag
r
e
e
S
O
M
-
S
N
N
S
O
M
f
o
r
f
r
e
q
u
e
n
c
y
r
e
p
r
e
s
e
n
t
a
t
i
o
n
S
N
N
f
o
r
sp
a
t
i
o
t
e
mp
o
r
a
l
p
a
t
t
e
r
n
&
9
7
.
4
0
,
9
9
.
6
0
W
u
e
t
a
l
. [
20
]
TI
M
I
T,
Li
b
r
i
s
p
e
e
c
h
,
F
A
M
E
S
N
N
,
M
F
C
C
,
F
B
A
N
K
,
F
M
LLR
La
r
g
e
v
o
c
a
b
u
l
a
r
y
r
e
c
o
g
n
i
t
i
o
n
3
6
.
9
,
1
8
.
7
M
u
k
h
o
p
a
d
h
y
a
y
e
t
a
l
. [
22
]
H
u
ma
n
f
o
o
t
st
e
p
s
o
u
n
d
s
S
N
N
,
W
S
N
e
n
e
r
g
y
e
f
f
i
c
i
e
n
c
y
Ti
me
d
o
m
a
i
n
f
o
r
a
c
o
u
s
t
i
c
c
l
a
ss
i
f
i
c
a
t
i
o
n
K
h
o
l
k
i
n
e
t
a
l
.
[
24
]
A
c
c
e
l
e
r
o
met
e
r
d
a
t
a
R
C
N
e
t
,
A
N
N
,
S
N
N
B
a
l
l
b
e
a
r
i
n
g
d
i
a
g
n
o
s
i
s
S
N
N
=
1
0
0
,
A
N
N
=
6
1
Y
a
maz
a
k
i
e
t
a
l
. [
23
]
R
o
b
o
t
i
c
s
d
o
ma
i
n
s
S
N
N
,
A
N
N
S
N
N
v
s
d
e
e
p
n
e
t
w
o
r
k
s
&
e
n
e
r
g
y
e
f
f
i
c
i
e
n
t
a
p
p
l
i
c
a
t
i
o
n
s
A
u
d
i
o
c
l
a
ss
i
f
i
c
a
t
i
o
n
3.
DATA CO
L
L
E
C
T
I
O
N
AND
E
XP
E
R
I
M
E
N
T
A
L
SE
T
UP
Fo
r
th
e
FVC
s
tu
d
y
,
k
n
o
wn
s
p
ee
ch
s
am
p
les
an
d
tr
ac
e
d
ata
wer
e
co
llected
f
r
o
m
th
e
Un
i
v
er
s
ity
o
f
New
So
u
th
W
ales
Facu
lty
o
f
E
lectr
ical
E
n
g
in
ee
r
in
g
an
d
T
elec
o
m
m
u
n
icatio
n
s
in
Sy
d
n
ey
,
Au
s
tr
alia.
T
h
e
b
e
n
c
h
m
a
r
k
d
atas
et
u
s
e
d
f
o
r
e
v
al
u
ati
o
n
c
o
n
s
is
ts
o
f
A
u
s
t
r
ali
an
E
n
g
l
is
h
r
e
c
o
r
d
i
n
g
s
f
r
o
m
o
v
e
r
3
8
9
9
s
p
ea
k
e
r
s
,
f
e
at
u
r
i
n
g
v
a
r
i
o
u
s
s
t
y
l
es su
c
h
a
s
ca
s
u
al
te
le
p
h
o
n
e
c
o
n
v
er
s
ati
o
n
s
,
in
f
o
r
m
ati
o
n
e
x
c
h
an
g
e
tas
k
s
,
a
n
d
p
s
e
u
d
o
-
p
o
lic
e
in
t
er
v
i
ews
.
T
h
is
d
ataset
was
d
iv
id
ed
in
to
tr
ai
n
in
g
a
n
d
te
s
tin
g
d
ata,
with
ac
ce
s
s
g
r
an
t
ed
u
p
o
n
o
b
tain
i
n
g
p
er
m
is
s
io
n
f
r
o
m
th
e
r
elev
a
n
t
au
th
o
r
ities
.
T
h
e
d
atasets
u
s
ed
in
th
is
FV
C
ex
p
er
im
en
t
wer
e
s
o
u
r
ce
d
f
r
o
m
th
e
FVC
d
ata
r
ep
o
s
ito
r
y
[
2
6
]
.
T
h
e
f
o
cu
s
o
n
A
u
s
tr
alian
E
n
g
lis
h
allo
ws
f
o
r
p
r
ec
is
e
a
n
aly
s
is
o
f
s
p
ee
c
h
p
atter
n
s
u
n
iq
u
e
to
Au
s
tr
alian
s
p
ea
k
e
r
s
,
ca
p
tu
r
in
g
v
ar
iatio
n
s
in
ac
c
en
t,
p
r
o
n
u
n
ciatio
n
,
an
d
o
th
er
lin
g
u
is
tic
f
ea
t
u
r
es
ess
en
tial
f
o
r
r
eliab
le
v
o
ice
co
m
p
ar
is
o
n
s
in
f
o
r
e
n
s
ic
co
n
tex
ts
.
T
h
e
d
ata
is
p
r
o
v
id
e
d
in
f
r
ee
l
o
s
s
les
s
au
d
io
co
d
ec
(
.
f
lac)
f
ile
f
o
r
m
at,
a
n
d
a
s
u
m
m
ar
y
o
f
t
h
e
ex
p
er
im
en
tal
d
ata
c
o
llectio
n
is
p
r
esen
ted
in
T
a
b
le
2
.
T
ab
le
2
.
Su
m
m
a
r
y
o
f
ex
p
er
im
en
tal
d
ata
co
llectio
n
D
a
t
a
s
e
t
n
a
me
N
u
mb
e
r
o
f
s
a
m
p
l
e
s
Tr
a
i
n
i
n
g
Te
st
i
n
g
G
e
n
d
e
r
A
u
d
i
o
f
o
r
ma
t
A
u
st
r
a
l
i
a
n
E
n
g
l
i
s
h
3
8
9
9
2
7
2
9
1
7
0
F
e
mal
e
&
M
a
l
e
.
f
l
a
c
4.
T
H
E
M
E
T
H
O
D
-
F
O
RE
NSI
C
VO
I
CE
CO
M
P
ARI
SO
N
USI
NG
SP
I
K
I
NG
NE
URA
L
NE
T
WO
RK
Fig
u
r
e
1
illu
s
tr
ates
th
e
p
r
o
p
o
s
ed
ar
ch
itectu
r
e
f
o
r
FVC
u
s
i
n
g
SNN.
T
h
e
p
r
o
p
o
s
ed
m
eth
o
d
in
th
is
s
tu
d
y
ten
d
e
d
to
h
a
v
e
an
in
o
r
d
i
n
ately
h
ig
h
er
p
r
o
p
o
r
tio
n
o
f
t
h
e
ex
p
er
im
e
n
tal
ap
p
r
o
ac
h
em
p
lo
y
s
an
SNN
m
o
d
el
,
wh
er
e
th
e
in
p
u
t
lay
er
r
ec
eiv
es
r
aw
s
p
ee
ch
s
am
p
les
in
th
e
f
o
r
m
o
f
au
d
io
f
iles
(
.
f
lac)
.
T
h
es
e
s
am
p
les
ar
e
p
r
e
-
p
r
o
ce
s
s
ed
to
r
em
o
v
e
n
o
is
e
an
d
o
th
er
in
ter
f
er
en
ce
s
.
Af
ter
p
r
e
-
p
r
o
ce
s
s
in
g
,
f
ea
tu
r
e
ex
tr
ac
tio
n
an
d
class
if
icatio
n
ar
e
p
er
f
o
r
m
ed
u
s
in
g
a
n
SNN.
O
n
ce
th
e
p
r
e
-
p
r
o
ce
s
s
ed
d
ata
is
f
ed
in
to
th
e
SNN
m
o
d
el,
f
e
atu
r
es
s
u
ch
as
tim
e
an
d
f
r
e
q
u
en
c
y
d
if
f
e
r
en
ce
s
o
f
s
p
ik
es
ar
e
ex
tr
ac
ted
an
d
e
n
co
d
ed
in
to
a
f
o
r
m
at
s
u
itab
le
f
o
r
th
e
SNN.
T
h
e
en
co
d
ed
d
ata
is
th
en
p
ass
ed
th
r
o
u
g
h
th
e
SNN,
wh
ich
co
n
s
i
s
ts
o
f
m
u
ltip
le
lay
er
s
o
f
in
te
r
co
n
n
ec
ted
n
eu
r
o
n
s
,
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
14
,
No
.
4
,
Au
g
u
s
t
20
25
:
2
6
8
9
-
2
7
0
0
2692
allo
win
g
f
o
r
d
etailed
an
al
y
s
is
an
d
co
m
p
ar
is
o
n
o
f
v
o
ice
p
atte
r
n
s
.
E
ac
h
n
e
u
r
o
n
in
th
e
SNN
in
teg
r
ates
in
co
m
in
g
s
p
ik
es f
r
o
m
o
th
er
n
eu
r
o
n
s
an
d
em
its
its
o
wn
s
p
ik
e
wh
en
its
m
em
b
r
an
e
p
o
ten
tial r
ea
ch
es a
s
p
ec
if
ied
th
r
esh
o
ld
.
T
h
e
tim
in
g
o
f
th
ese
s
p
ik
es
en
co
d
es
in
f
o
r
m
atio
n
a
b
o
u
t
th
e
in
p
u
t
s
p
ee
ch
.
Fin
ally
,
th
e
o
u
tp
u
t
lay
er
o
f
th
e
SNN
g
en
er
ates
a
s
et
o
f
s
p
ik
es
r
ep
r
e
s
en
tin
g
th
e
r
ec
o
g
n
ize
d
s
p
ee
ch
,
wh
ich
is
th
en
d
ec
o
d
ed
b
ac
k
in
to
a
co
n
v
en
tio
n
al
f
o
r
m
at,
s
u
ch
as
a
s
im
ilar
ity
s
co
r
e
b
etwe
en
th
e
in
p
u
t
s
p
ee
ch
an
d
a
r
ef
er
en
ce
s
am
p
le
.
T
h
e
p
e
r
f
o
r
m
an
ce
is
ev
alu
ate
d
th
r
o
u
g
h
a
d
ju
s
tab
le
th
r
esh
o
ld
s
p
ik
e
tim
in
g
a
n
d
ass
ess
ed
u
s
in
g
a
co
n
f
u
s
io
n
m
atr
ix
,
wh
ich
i
n
clu
d
es
m
etr
i
cs
s
u
ch
as
ac
cu
r
ac
y
,
p
r
ec
is
io
n
,
r
ec
all,
F1
s
co
r
e
,
an
d
F2
s
co
r
e.
T
h
is
s
ec
tio
n
o
u
tlin
es
th
e
p
r
o
p
o
s
ed
r
esear
ch
m
eth
o
d
o
lo
g
y
f
o
r
FVC
u
s
in
g
SNNs
.
A
d
etailed
d
escr
ip
tio
n
o
f
ea
ch
s
u
b
s
ec
tio
n
,
in
clu
d
in
g
p
re
-
p
r
o
ce
s
s
in
g
an
d
t
h
e
SNNs
m
o
d
el,
is
p
r
o
v
i
d
ed
i
n
th
e
s
u
b
s
eq
u
e
n
t sectio
n
s
.
Fig
u
r
e
1
.
Pro
p
o
s
ed
ar
c
h
itectu
r
e
f
o
r
FVC
u
s
in
g
SNN
4
.
1
.
P
re
pro
ce
s
s
ing
T
h
e
s
tatio
n
ar
y
n
o
is
e
r
ed
u
ctio
n
m
eth
o
d
is
em
p
lo
y
ed
to
elim
in
ate
b
ac
k
g
r
o
u
n
d
n
o
is
e
f
r
o
m
t
h
e
f
o
r
en
s
ic
v
o
ice
s
am
p
les,
p
ar
ticu
lar
ly
w
ith
in
th
e
Au
s
tr
alian
E
n
g
lis
h
d
ataset,
wh
ich
is
s
to
r
ed
in
th
e
.
f
lac
f
ile
f
o
r
m
at.
T
h
ese
ar
e
p
r
o
v
id
e
d
to
th
e
m
o
d
el
alo
n
g
with
a
n
o
is
e
s
am
p
le,
en
co
m
p
ass
in
g
th
e
ty
p
ical
b
a
ck
g
r
o
u
n
d
n
o
is
e
f
o
r
th
e
s
am
p
le.
T
h
is
n
o
is
e
s
am
p
l
e
is
co
m
b
in
ed
with
a
s
ig
n
al
c
lip
co
n
tain
in
g
b
o
t
h
th
e
n
o
is
e
an
d
th
e
s
ig
n
al
th
at
n
ee
d
s
to
b
e
r
e
m
o
v
e
d
,
as
illu
s
tr
ated
in
Fig
u
r
e
2
(
a)
n
o
is
y
s
p
e
ec
h
in
p
u
t
d
ata
Fig
u
r
e
2
(
b
)
n
o
i
s
e
r
ed
u
ce
d
s
p
ee
ch
o
u
tp
u
t
d
ata.
T
h
e
f
o
llo
win
g
p
r
o
v
id
es a
n
ex
p
la
n
atio
n
o
f
th
e
s
ta
tio
n
ar
y
n
o
is
e
r
ed
u
ctio
n
A
lg
o
r
i
th
m
1
Alg
o
r
ith
m
1
: Statio
n
er
y
n
o
is
e
r
ed
u
ctio
n
al
g
o
r
ith
m
Input: Australian English dataset audio recording samples of voice are used.
Output: Noise Reduced Speech Data.
Step 1: Spectrogram is calculated for the noisy audio clip.
Step 2: In frequency statistics are measured using the noise spectrogram.
Step 3: On the basis of noise statistics a threshold is created.
Step 4: Through th
e signals spectrogram is calculated.
Step 5: By the signal spectrogram threshold is determined and compared.
Step 6: To smooth the mask over time and frequency the linear filter is used.
Step 7: The mask is applied to the signals spectrogram and inverts
the noise signal to
produce positive results.
(
a)
(
b
)
Fig
u
r
e
2
.
Pre
p
r
o
ce
s
s
in
g
r
esu
lt
s
(
a)
n
o
is
y
s
p
ee
ch
i
n
p
u
t
d
ata
(
b
)
n
o
is
e
r
e
d
u
ce
d
s
p
ee
c
h
o
u
tp
u
t d
ata
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
S
emi
-
a
u
to
ma
tic
vo
ice
c
o
mp
a
r
i
s
o
n
a
p
p
r
o
a
ch
u
s
in
g
s
p
ikin
g
n
e
u
r
a
l
…
(
K
r
u
th
ika
S
id
d
a
n
a
k
a
tte
Go
p
a
la
ia
h
)
2693
4
.
2
.
Sp
ik
ing
neura
l net
wo
rk
T
h
e
p
r
o
p
o
s
ed
m
o
d
el
u
tili
ze
s
SNN
in
s
p
ir
ed
b
y
th
e
s
tr
u
ctu
r
e
an
d
f
u
n
ctio
n
ality
o
f
b
io
lo
g
ical
n
eu
r
a
l
n
etwo
r
k
s
f
o
u
n
d
in
th
e
h
u
m
an
b
r
ain
.
I
n
co
n
tr
ast
to
co
n
v
e
n
t
io
n
al
n
eu
r
al
n
etwo
r
k
s
,
wh
ich
d
ep
en
d
o
n
s
ig
n
als
with
co
n
tin
u
o
u
s
v
alu
es,
SNN
em
p
lo
y
d
is
cr
ete
s
p
ik
es
o
r
p
u
ls
es
to
co
m
m
u
n
icate
in
f
o
r
m
atio
n
b
etwe
en
n
eu
r
o
n
s
,
r
esem
b
lin
g
th
e
tr
an
s
m
is
s
io
n
o
f
s
ig
n
als
th
r
o
u
g
h
ac
tio
n
p
o
ten
tials
in
b
io
lo
g
ical
s
y
s
tem
s
.
Ad
o
p
tin
g
a
s
em
i
-
au
to
m
atic
ap
p
r
o
ac
h
b
a
s
ed
SNN
ar
e
co
n
s
id
er
e
d
f
o
r
F
VC
in
th
e
p
r
o
p
o
s
ed
r
esear
ch
m
o
d
el.
T
h
e
p
r
o
ce
s
s
b
eg
in
s
w
ith
lo
ad
in
g
an
d
p
r
ep
r
o
ce
s
s
in
g
th
e
au
d
i
o
d
ata,
e
x
tr
a
ctin
g
r
elev
an
t
f
ea
tu
r
es
an
d
co
n
v
er
tin
g
co
n
tin
u
o
u
s
au
d
io
s
ig
n
als
in
to
s
p
ik
e
tr
ai
n
s
u
s
in
g
en
co
d
in
g
tech
n
iq
u
e
s
lik
e
r
ate
co
d
in
g
o
r
tim
e
-
to
-
f
ir
s
t
-
s
p
ik
e
co
d
in
g
.
Su
b
s
eq
u
en
tly
,
s
y
n
a
p
tic
weig
h
ts
o
f
th
e
SNN
ar
e
in
i
tialized
ei
th
er
r
an
d
o
m
ly
o
r
u
s
in
g
p
r
e
-
tr
a
in
ed
weig
h
ts
f
r
o
m
a
n
eu
r
al
n
etwo
r
k
.
A
th
r
esh
o
ld
is
ap
p
lied
to
d
eter
m
in
e
th
e
s
i
m
ilar
ity
o
r
d
is
s
im
ilar
ity
o
f
s
p
i
k
e
tr
ain
s
.
W
ith
in
th
e
SNN
m
o
d
el,
n
eu
r
o
n
s
co
m
m
u
n
icate
th
r
o
u
g
h
d
is
cr
ete
s
p
ik
es
v
ia
th
e
m
em
b
r
an
e
p
o
te
n
tial.
Neu
r
o
n
s
ac
cu
m
u
late
in
p
u
t o
v
er
tim
e
an
d
em
it a
s
p
i
k
e
o
n
ce
t
h
e
th
r
esh
o
l
d
is
r
ea
ch
ed
.
Vo
ice
p
atter
n
s
ar
e
en
c
o
d
ed
in
th
e
tim
in
g
o
f
s
p
ik
es,
an
d
s
y
n
a
p
tic
weig
h
ts
ar
e
u
p
d
ated
b
ase
d
o
n
s
p
ik
e
tim
in
g
,
co
m
m
o
n
ly
th
r
o
u
g
h
s
p
ik
e
-
tim
in
g
-
d
e
p
en
d
e
n
t
p
last
icity
(
STDP)
o
r
v
ar
ian
ts
,
to
id
e
n
tify
v
o
ice
p
atter
n
s
.
T
o
d
ec
o
d
e
a
n
d
p
r
ed
ict
k
n
o
wn
an
d
tr
ac
e
v
o
ice
s
am
p
les
an
ad
ju
s
tab
le
th
r
esh
o
ld
is
ap
p
li
ed
to
d
eter
m
in
e
th
e
s
im
ilar
ity
o
r
d
is
s
im
ilar
ity
o
f
s
p
ik
e
tr
ai
n
s
.
Sev
er
al
p
ar
a
m
eter
s
s
ig
n
if
ican
tly
i
n
f
lu
en
ce
SNN
b
eh
av
io
r
an
d
lear
n
in
g
,
i
n
clu
d
in
g
in
p
u
t
s
p
ik
e,
m
em
b
r
a
n
e
p
o
ten
tial,
s
p
ik
e
g
en
er
atio
n
,
STDP,
an
d
o
u
tp
u
t
s
p
ik
e.
I
n
p
u
t
s
p
ik
es
r
ec
eiv
e
in
f
o
r
m
atio
n
f
r
o
m
p
r
o
ce
s
s
ed
au
d
io
s
am
p
les
wh
ic
h
th
en
p
ass
th
r
o
u
g
h
th
e
m
em
b
r
an
e
f
u
n
ctio
n
.
Me
m
b
r
an
e
p
o
ten
tial
g
en
er
ate
s
s
p
ik
e
tim
es
an
d
o
n
ce
th
e
s
p
ik
e
t
im
e
s
u
r
p
ass
es
a
th
r
es
h
o
ld
SNN
n
etwo
r
k
id
en
tifie
s
v
o
ice
p
atter
n
s
f
o
r
class
if
icatio
n
b
ased
o
n
s
im
ilar
ity
o
r
d
is
s
im
ilar
ity
u
s
in
g
ST
DP.
Ou
tp
u
t
s
p
ik
es
g
en
er
ated
b
y
n
eu
r
o
n
s
aid
in
ev
alu
atin
g
o
r
id
e
n
tify
in
g
s
im
ilar
o
r
d
is
s
im
ilar
v
o
ice
s
am
p
l
es.
T
h
e
s
u
b
s
ec
tio
n
s
4
.
2
.
1
t
h
r
o
u
g
h
4
.
2
.
5
elab
o
r
ate
o
n
th
ese
k
ey
p
ar
am
eter
s
.
4
.
2
.
1
.
I
np
ut
s
pik
es a
nd
m
em
bra
ne
po
t
ent
ia
l
T
h
e
in
p
u
t
s
p
ik
es
ar
e
d
e
r
iv
ed
f
r
o
m
p
r
e
-
p
r
o
ce
s
s
ed
au
d
i
o
s
a
m
p
les,
ca
p
tu
r
i
n
g
i
n
f
o
r
m
atio
n
r
elate
d
to
b
o
th
tim
e
a
n
d
f
r
eq
u
en
cy
.
C
o
n
s
eq
u
en
tly
,
i
n
p
u
t
s
p
ik
es
p
lay
a
p
iv
o
tal
r
o
le
in
e
n
co
d
i
n
g
to
ac
h
iev
e
th
e
d
esire
d
o
u
tp
u
t
th
r
o
u
g
h
m
em
b
r
a
n
e
p
o
t
en
tial.
On
ce
th
e
in
p
u
t
s
p
i
k
e
tim
e
is
g
en
er
ated
,
th
e
m
em
b
r
a
n
e
p
o
ten
tial
in
d
icate
s
th
e
elec
tr
ic
p
o
ten
tial
t
h
r
o
u
g
h
o
u
t
th
e
v
o
ice
p
atter
n
s
.
I
t
g
o
v
er
n
s
th
e
n
eu
r
o
n
'
s
g
en
er
atio
n
o
f
an
ac
tio
n
p
o
ten
tial
s
p
ik
e,
with
c
r
itical
f
ac
to
r
s
in
clu
d
in
g
its
u
p
d
ate
o
v
er
tim
e
an
d
r
esp
o
n
s
e
to
in
c
o
m
in
g
s
p
i
k
es.
T
h
ese
f
ac
to
r
s
d
eter
m
in
e
h
o
w
s
p
ik
es a
r
e
g
e
n
er
ated
o
v
e
r
tim
e.
Up
d
ate
o
v
er
t
im
e:
t
h
e
m
em
b
r
an
e
p
o
ten
tial
is
d
y
n
a
m
ic
an
d
c
h
an
g
es
o
v
er
tim
e
i
n
r
esp
o
n
s
e
to
in
co
m
in
g
s
ig
n
als
o
r
s
p
ik
es.
T
h
e
d
y
n
am
ics
o
f
th
is
ch
an
g
e
a
r
e
ty
p
ically
d
escr
ib
ed
b
y
a
s
e
t
o
f
e
q
u
atio
n
s
t
h
at
m
o
d
el
h
o
w
th
e
n
eu
r
o
n
in
te
g
r
ates
in
co
m
in
g
in
f
o
r
m
atio
n
.
T
h
e
r
esp
o
n
s
e
t
o
in
co
m
in
g
s
p
i
k
es:
t
h
e
m
em
b
r
an
e
p
o
ten
tial
is
in
f
lu
en
ce
d
b
y
th
e
s
y
n
ap
tic
in
p
u
ts
r
ec
eiv
ed
f
r
o
m
co
n
n
ec
ted
n
eu
r
o
n
s
,
with
ea
ch
in
co
m
in
g
s
p
ik
e
co
n
tr
ib
u
tin
g
to
th
e
c
h
an
g
e
in
t
h
e
m
em
b
r
a
n
e
p
o
te
n
tial.
I
n
(
1
)
r
ep
r
esen
ts
th
e
m
em
b
r
an
e
p
o
te
n
tial o
v
er
tim
e.
V
(
t
)
=
Σ
iw
i
∗
si
(
t
−
ti
)
(
1
)
W
h
er
e
:
−
T
h
e
m
em
b
r
an
e
p
o
ten
tial f
u
n
ctio
n
at
tim
e
t is d
en
o
te
d
b
y
V
(
t
)
.
−
∑i
d
en
o
tes th
e
s
u
m
m
atio
n
o
v
e
r
th
e
in
d
e
x
i.
−
wi
r
ep
r
esen
ts
th
e
weig
h
t a
s
s
o
ciate
d
with
ea
ch
f
u
n
ctio
n
.
−
s
i(
t
-
ti)
is
th
e
s
p
ik
e
tr
ain
f
r
o
m
th
e
p
r
esy
n
a
p
tic
n
eu
r
o
n
i a
t tim
e
t
-
ti
4
.
2
.
2
.
Sp
ik
e
g
ener
a
t
io
n
Af
ter
a
s
p
ik
e
is
g
en
er
ated
b
y
th
e
m
em
b
r
an
e
p
o
ten
tial,
it r
ea
ch
es a
th
r
esh
o
ld
th
at
allo
ws th
e
n
eu
r
o
n
to
id
e
n
tify
b
o
t
h
d
is
s
im
ilar
ity
an
d
s
im
ilar
ity
in
v
o
ice
p
atter
n
s
.
T
h
e
p
ar
am
eter
s
o
f
s
p
ik
e
g
en
er
a
tio
n
ar
e
in
f
lu
en
ce
d
b
y
th
r
esh
o
ld
cr
o
s
s
in
g
,
ac
tio
n
p
o
ten
tial,
an
d
n
eu
r
o
n
al
r
esp
o
n
s
e.
T
h
r
esh
o
ld
c
r
o
s
s
in
g
r
ef
e
r
s
to
th
e
p
r
o
ce
s
s
wh
er
e
n
eu
r
o
n
s
p
o
s
s
ess
a
s
p
ec
if
ic
th
r
esh
o
ld
lev
el
o
f
m
em
b
r
an
e
p
o
ten
tial.
W
h
en
th
e
m
em
b
r
an
e
p
o
ten
tial
s
u
r
p
ass
es
th
is
th
r
esh
o
ld
,
th
e
n
eu
r
o
n
g
e
n
er
ates
a
s
p
ik
e,
also
k
n
o
w
n
as
an
ac
tio
n
p
o
ten
tial.
T
h
e
ac
tio
n
p
o
ten
tial
is
a
b
r
ief
elec
tr
ical
p
u
ls
e
th
at
tr
av
els
al
o
n
g
th
e
n
eu
r
o
n
'
s
ax
o
n
,
s
ig
n
al
in
g
th
e
n
e
u
r
o
n
'
s
ac
tiv
atio
n
to
o
th
er
n
e
u
r
o
n
s
o
r
tar
g
et
ce
lls
.
T
h
is
s
p
ik
e
r
e
p
r
esen
ts
an
all
-
or
-
n
o
th
in
g
r
es
p
o
n
s
e:
if
th
e
m
em
b
r
an
e
p
o
t
en
tial
ex
ce
ed
s
th
e
th
r
esh
o
ld
,
a
s
p
ik
e
is
g
e
n
er
ated
; o
th
er
wis
e,
n
o
s
p
ik
e
o
cc
u
r
s
.
Ma
th
em
atica
lly
,
th
e
n
eu
r
o
n
g
e
n
er
ates sp
ik
es wh
en
its
m
em
b
r
a
n
e
p
o
ten
tial c
r
o
s
s
es a
th
r
esh
o
ld
,
r
e
p
r
esen
ted
b
y
θ
in
(
2
)
.
If
V
(
t
)
≥
θ
,
the
n
the
n
e
uro
n
e
mits
a
s
pik
e
(
2
)
4
.
2
.
3
.
Sp
ik
e
t
im
e
depend
ent
pla
s
t
icit
y
As
th
e
s
p
ik
e
r
ea
ch
es
th
e
th
r
e
s
h
o
ld
,
STDP
is
u
tili
ze
d
to
ad
ju
s
t
th
e
s
y
n
ap
tic
weig
h
ts
,
ass
ess
in
g
th
e
s
tr
en
g
th
an
d
wea
k
n
ess
o
f
c
o
n
n
ec
tio
n
s
in
th
e
SNN
n
et
wo
r
k
to
d
is
ce
r
n
s
im
ilar
ity
a
n
d
d
is
s
im
ilar
ity
in
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
14
,
No
.
4
,
Au
g
u
s
t
20
25
:
2
6
8
9
-
2
7
0
0
2694
r
ec
o
g
n
izin
g
v
o
ice
p
atter
n
s
.
T
h
e
s
tr
en
g
th
o
f
a
co
n
n
ec
tio
n
(
s
y
n
ap
s
e)
in
an
SNN
n
etwo
r
k
s
h
o
u
ld
alter
b
ased
o
n
th
e
r
elativ
e
tim
in
g
o
f
s
p
ik
es
b
etwe
en
th
e
p
r
esy
n
ap
tic
an
d
p
o
s
ts
y
n
ap
tic
n
eu
r
o
n
s
.
I
f
a
p
r
esy
n
ap
tic
n
eu
r
o
n
co
n
s
is
ten
tly
f
ir
es
b
ef
o
r
e
a
p
o
s
ts
y
n
ap
tic
n
eu
r
o
n
,
th
e
co
n
n
ec
tio
n
b
etwe
en
th
em
s
tr
en
g
th
e
n
s
.
C
o
n
v
er
s
ely
,
i
f
th
e
p
o
s
ts
y
n
ap
tic
n
eu
r
o
n
f
ir
es
f
ir
s
t,
th
e
co
n
n
ec
tio
n
wea
k
en
s
.
T
h
is
p
r
o
ce
s
s
en
a
b
les
th
e
n
etwo
r
k
t
o
ad
ap
t
to
p
atter
n
s
in
th
e
in
p
u
t
d
ata.
T
h
e
s
y
n
a
p
tic
weig
h
ts
u
n
d
e
r
g
o
p
last
icity
ad
ju
s
tm
en
ts
b
ased
o
n
th
e
tim
in
g
o
f
p
r
e
an
d
p
o
s
ts
y
n
ap
tic
s
p
ik
es a
s
r
ep
r
esen
ted
in
(
3
)
.
Δ
wi
=
η
⋅
si
(
t
−
ti
)
⋅
PostSyn
a
pti
c
S
pik
e
(
t
)
(
3
)
W
h
er
e
:
−
Δ
wi
r
ep
r
esen
ts
th
e
ch
an
g
e
in
t
h
e
s
y
n
ap
tic
weig
h
t w
i.
−
η
is
th
e
lear
n
in
g
r
ate,
co
n
tr
o
lli
n
g
th
e
m
a
g
n
itu
d
e
o
f
weig
h
t a
d
ju
s
tm
en
ts
.
−
s
i(
t−ti
)
is
th
e
f
u
n
ctio
n
ass
o
ciate
d
with
th
e
tim
in
g
o
f
th
e
p
r
esy
n
ap
tic
s
p
ik
e
at
tim
e
ti.
−
Po
s
tSy
n
ap
ticSp
ik
e(
t)
is
a
f
u
n
c
tio
n
th
at
in
d
icate
s
if
a
s
p
ik
e
o
c
cu
r
r
ed
i
n
th
e
p
o
s
ts
y
n
ap
tic
n
eu
r
o
n
at
tim
e
t.
4
.
2
.
4
.
O
utput
s
pik
e
T
h
e
o
u
tp
u
t
s
p
ik
e
aid
s
in
d
ec
o
d
in
g
v
o
ice
s
im
il
ar
ity
o
r
d
is
s
im
ilar
ity
,
with
p
er
f
o
r
m
a
n
ce
ev
alu
ated
u
s
in
g
ac
cu
r
ac
y
,
p
r
ec
is
io
n
,
r
ec
all,
F1
s
co
r
e
,
an
d
F2
s
co
r
e.
M
em
b
r
an
e
p
o
ten
tial
u
p
d
ates
in
teg
r
ate
s
ig
n
als,
wh
ile
s
p
ik
e
g
en
er
atio
n
in
d
icate
s
n
eu
r
o
n
ac
tiv
atio
n
.
Sy
n
ap
tic
p
last
icity
en
ab
les
ad
ap
tiv
e
le
ar
n
in
g
,
m
im
ick
i
n
g
b
io
lo
g
ical
n
e
u
r
al
s
y
s
tem
s
.
T
h
e
s
e
s
tep
s
ar
e
o
u
tlin
ed
in
Alg
o
r
ith
m
2.
Alg
o
r
ith
m
2
: T
o
id
en
tif
y
th
e
s
im
ilar
ity
o
r
d
is
s
im
ilar
ity
o
f
v
o
ices th
r
o
u
g
h
th
e
p
r
e
p
r
o
ce
s
s
ed
d
ata
Input: Preprocessed forensic voice samples.
Output: Prediction for FVC based on the evaluation on Confusion matrix.
Step 1: Initialization
−
Synaptic
Weights,
Membra
ne
Potentials,
and
Thres
holds:
Set
the
initial
values
for
synaptic
weights,
membrane
potenti
al
s,
an
d
th
re
sh
ol
ds
fo
r
al
l
ne
ur
on
s.
In
th
is
ca
se
,
they are initialized to 0.5.
−
Learning
Rate
(η):
Choose
a
learning
rate
parame
ter
(η)
to
control
the
m
agnitude
of
weight adjustments during the learning process. Here, it is set to 0.001.
Step 2: Training
−
A
djusting
Synaptic
Weigh
ts:
Utilize
a
learning
r
ule
based
on
spike
time
to
update
synaptic weights. The spike timing difference based on the adjustable threshold.
spi
ke
_
t
i
m
e
s
=
(
y
>
0
.
5
)
.
n
o
n
z
e
r
o
(
)
[
0
]
∗
0
.
001
#
if
s
p
i
ke
a
m
p
li
t
ude
>
0
.
5
(4)
In
th
e
pr
ov
id
ed
co
de
fo
r
vo
ic
e
co
mp
ar
is
on
,
th
e
ex
ac
t
ti
me
of
a
sp
ik
e
is
d
et
er
mi
ne
d
ba
se
d
on the threshold condition.
Where,
•
y is the audio signal.
•
(y
>
0.
5)
cr
ea
te
s
a
bi
na
ry
ma
sk
wh
er
e
th
e
am
pl
it
ud
e
va
lu
es
gr
ea
te
r
th
an
0.
5
ar
e
marked as True and others as `False`.
•
.nonzero () returns the indices where the condition is True.
•
*0
.0
01
sc
al
es
th
e
in
di
ce
s
to
re
pr
es
en
t
ti
me
in
se
co
nd
s
(a
ss
um
in
g
th
e
au
di
o
is
sampled at 1
,
000 Hz
).
•
Th
e
re
su
lt
in
g
sp
ik
e_
ti
me
s
va
ri
ab
le
co
nt
ai
ns
th
e
ti
me
s
(i
n
se
co
nd
s)
wh
e
n
th
e
amplitude
of
the
audio
signal
exceeds
the
threshold
of
0.5.
These
times
correspond
to
the
occurre
nces
of
spikes
in
the
aud
io
signal,
as
determined
by
the
ch
os
en
th
re
sh
ol
d.
Th
e
s
p
ec
if
ic
va
lu
e
of
0.
5
ca
n
be
ad
ju
st
ed
ac
co
rd
in
g
to
th
e
characteristics of voice samples and the desired sensitivity of spike detection.
−
Presenting
Training
Sampl
es:
Introduce
training
sa
mples
to
the
network.
The
se
samples
represent
patterns
or
d
ata
points
th
at
th
e
ne
tw
or
k
wi
ll
le
ar
n
to
re
co
gn
iz
e
or
classify.
By
using
the
library
functions
such
a
s
tensor
flow
and
pytor
ch
the
SN
N
network
is
built.
Where
the
optimization
function
used
is
Adam
and
b
inary
cross
entropy is the loss function.
Step 3: Testing
Th
e
fo
ll
ow
in
g
ps
eu
do
co
de
is
ut
il
iz
ed
to
id
en
ti
fy
an
d
ev
al
ua
te
vo
ic
e
sa
mp
le
s
be
tw
ee
n
th
e
kn
ow
n
an
d
tr
ac
e.
In
th
is
co
nt
ex
t,
0
re
pr
es
en
ts
f
al
se
,
an
d
1
re
pr
es
en
ts
t
ru
e,
in
di
ca
ti
ng
whether
the
suspect
is
id
entified
through
the
voic
e
samples.
This
evaluatio
n
is
performed
using accuracy to assess the similarity and dissimilarity in the voice.
defevaluate_voice_samples (known_sample, trace_sample):
If known_sample == trace_sample:
return 1 # True, suspect identified
else:
return 0 # False, suspect not identified
#
Example usage
known_sample = "voice_sample_1"
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
S
emi
-
a
u
to
ma
tic
vo
ice
c
o
mp
a
r
i
s
o
n
a
p
p
r
o
a
ch
u
s
in
g
s
p
ikin
g
n
e
u
r
a
l
…
(
K
r
u
th
ika
S
id
d
a
n
a
k
a
tte
Go
p
a
la
ia
h
)
2695
trace_sample = "voice_sample_2"
result =evaluate_voice_samples (known_sample, trace_sample)
print ("Result:", result)
Step 4: Inference of proposed research model
The proposed research model identifies the sus
pect based on similarity or dissimilarity.
5.
RE
SU
L
T
AND
DI
SCUS
SI
O
N
T
h
e
p
er
f
o
r
m
a
n
ce
ev
al
u
atio
n
o
f
th
e
p
r
o
p
o
s
ed
f
r
a
m
ewo
r
k
is
co
n
d
u
cte
d
u
s
in
g
v
ar
io
u
s
m
etr
ics,
with
a
d
etailed
an
aly
s
is
p
r
esen
ted
in
th
e
f
o
r
m
o
f
a
co
n
f
u
s
io
n
m
atr
i
x
an
d
r
ec
eiv
er
o
p
er
atin
g
ch
a
r
a
cter
is
tic
-
ar
ea
u
n
d
er
th
e
cu
r
v
e
(
R
OC
-
AUC
)
an
aly
s
is
.
T
h
ese
p
er
f
o
r
m
an
ce
ass
ess
m
en
ts
ar
e
s
y
s
tem
atica
lly
d
is
cu
s
s
ed
in
s
ec
tio
n
s
5
.
1
an
d
5
.
2
,
w
h
er
e
th
e
co
n
f
u
s
io
n
m
atr
ix
p
r
o
v
id
es
in
s
ig
h
ts
i
n
to
class
if
icatio
n
ac
cu
r
ac
y
b
y
d
is
p
lay
in
g
tr
u
e
p
o
s
itiv
es,
f
alse
p
o
s
itiv
es,
tr
u
e
n
eg
ativ
es,
an
d
f
alse
n
eg
ativ
e
s
.
Me
an
wh
ile,
th
e
R
OC
-
AUC
an
aly
s
is
ev
alu
ates
th
e
m
o
d
el'
s
d
is
cr
im
in
ativ
e
ab
ilit
y
,
illu
s
tr
atin
g
its
ef
f
ec
tiv
en
ess
in
d
is
tin
g
u
is
h
in
g
b
etwe
en
d
if
f
er
en
t c
lass
es.
5
.
1
.
Co
nfusi
o
n m
a
t
rix
a
nd
p
er
f
o
rma
nce
a
na
ly
s
is
T
h
e
Fig
u
r
e
3
(
a
)
d
is
p
lay
s
t
h
e
co
n
f
u
s
io
n
m
atr
ix
p
er
f
o
r
m
a
n
ce
ev
alu
atio
n
o
f
th
e
SNN
-
b
ase
d
ap
p
r
o
ac
h
in
class
if
y
in
g
th
e
Au
s
tr
alian
E
n
g
lis
h
d
ataset
f
o
r
FVC
.
P
er
f
o
r
m
a
n
ce
m
etr
ics
s
u
ch
as
ac
cu
r
ac
y
,
p
r
ec
is
io
n
,
r
ec
all,
F1
s
co
r
e
,
an
d
F2
s
co
r
e
ar
e
u
s
ed
to
ev
alu
ate
th
e
class
if
ier
'
s
ef
f
ec
tiv
en
ess
.
T
h
e
co
n
f
u
s
io
n
m
atr
ix
d
is
p
lay
s
th
e
n
u
m
b
er
o
f
co
r
r
ec
tly
id
e
n
tifie
d
m
atch
i
n
g
v
o
ices
as
tr
u
e
p
o
s
itiv
es
(
T
P)
en
tr
ies
in
th
e
b
o
t
to
m
r
ig
h
t
q
u
ad
r
a
n
t
o
f
m
atr
ix
in
d
icate
ac
cu
r
ate
m
atch
p
r
e
d
ictio
n
s
.
T
h
e
tr
u
e
n
e
g
ativ
e
(
T
N)
r
ef
er
s
t
o
th
e
en
tr
i
es
in
th
e
u
p
p
er
-
lef
t
q
u
ad
r
a
n
t
o
f
th
e
m
atr
ix
th
at
ac
cu
r
ately
p
r
ed
ict
n
o
n
-
m
atch
es
.
C
o
m
p
ar
ab
ly
,
f
alse
p
o
s
itiv
es
r
ep
r
esen
ted
b
y
th
e
en
tr
ies
in
th
e
u
p
p
er
-
r
ig
h
t
q
u
a
d
r
an
t
o
f
th
e
m
atr
i
x
,
wh
ich
in
d
icate
in
ac
cu
r
ate
m
atch
p
r
e
d
ictio
n
s
,
wh
ile
f
alse
n
eg
ativ
es
r
ep
r
esen
ted
b
y
th
e
en
tr
ies
in
th
e
b
o
tto
m
-
lef
t
q
u
ad
r
an
t
o
f
th
e
m
atr
ix
,
wh
ich
in
d
icate
in
ac
cu
r
a
te
p
r
ed
ictio
n
s
o
f
n
o
n
-
m
atch
es.
I
n
th
e
c
o
n
tex
t
o
f
FVC
u
s
in
g
SNN
th
e
Fig
u
r
e
3
(
a)
r
e
p
r
esen
ted
th
e
ac
tu
al
co
n
f
u
s
io
n
m
atr
ix
f
o
r
co
n
s
id
er
in
g
2
,
7
2
9
s
am
p
les
f
o
r
tr
ai
n
in
g
an
d
1
,
1
7
0
s
am
p
les
f
o
r
test
in
g
in
th
e
c
o
n
tex
t
o
f
id
en
tific
atio
n
o
f
m
atc
h
in
g
o
r
n
o
n
-
m
atch
in
g
o
f
th
e
v
o
ice.
A
v
alu
e
o
f
0
s
ig
n
if
ies
d
is
s
im
ilar
ity
in
v
o
ice
r
ec
o
g
n
itio
n
,
wh
ile
a
v
alu
e
o
f
1
1
4
s
ig
n
if
ies
s
im
ilar
ity
,
in
d
icatin
g
th
e
id
e
n
tific
atio
n
o
f
a
s
u
s
p
ec
t
u
s
in
g
1
,
1
7
0
test
in
g
s
am
p
les
as
r
ep
r
e
s
en
ted
in
co
n
f
u
s
io
n
m
at
r
ix
d
e
p
icted
in
th
e
Fig
u
r
e
3
(
a)
.
I
n
c
o
n
f
u
s
io
n
m
atr
ix
th
e
v
alu
es
in
ea
ch
ce
ll
in
d
icate
th
e
n
u
m
b
er
o
f
in
s
tan
ce
s
.
T
h
e
h
ea
tm
ap
v
is
u
aliza
tio
n
with
`
s
n
s
.
h
ea
tm
ap
`
p
r
o
v
id
es
a
co
lo
r
-
co
d
ed
r
e
p
r
esen
tatio
n
,
wh
er
e
d
ar
k
e
r
s
h
ad
es
in
d
icat
e
h
ig
h
er
c
o
u
n
ts
,
h
elp
in
g
to
q
u
ick
ly
id
e
n
tify
th
e
p
er
f
o
r
m
an
ce
o
f
th
e
m
o
d
el.
T
h
e
Fig
u
r
e
3
(
a
)
is
v
is
u
al
r
ep
r
ese
n
ts
o
f
v
alu
es tab
u
lated
in
T
ab
l
e
3
(
a)
.
C
o
m
p
ar
ed
th
e
cu
r
r
en
t
s
tate
o
f
ar
t
to
th
e
ex
is
tin
g
wo
r
k
s
[
1
]
-
[
3
]
,
id
en
tifie
d
th
e
d
ee
p
lear
n
in
g
tech
n
iq
u
es
f
o
r
o
u
r
p
r
o
p
o
s
ed
m
eth
o
d
wh
ic
h
d
em
o
n
s
tr
ates
a
p
o
ten
tial
en
h
an
ce
m
en
t
in
cla
s
s
if
icatio
n
ac
cu
r
ac
y
,
ac
h
iev
in
g
an
im
p
r
ess
iv
e
9
4
.
2
1
%
o
n
th
e
Au
s
tr
alian
E
n
g
lis
h
d
ataset.
Ad
d
itio
n
ally
,
th
e
m
o
d
e
l's
ef
f
ec
tiv
en
ess
is
f
u
r
th
er
s
u
p
p
o
r
te
d
b
y
its
s
tr
o
n
g
p
er
f
o
r
m
an
ce
ac
r
o
s
s
m
u
ltip
le
ev
alu
atio
n
m
etr
ics,
with
a
p
r
ec
is
io
n
o
f
8
5
.
2
1
%,
r
ec
all
o
f
8
2
.
1
6
%,
F1
s
co
r
e
o
f
8
1
.
1
1
%,
a
n
d
F2
s
co
r
e
o
f
8
0
.
1
0
%
ac
r
o
s
s
v
ar
ied
s
am
p
les.
T
h
ese
r
esu
lts
in
d
icate
th
e
r
o
b
u
s
tn
ess
an
d
r
eliab
ilit
y
o
f
o
u
r
a
p
p
r
o
ac
h
in
FVC
.
T
h
e
f
in
d
in
g
s
ar
e
ef
f
ec
tiv
el
y
p
r
es
en
ted
an
d
v
is
u
ally
d
ep
icted
in
Fig
u
r
e
3
(
b
)
,
ac
c
o
m
p
an
ied
b
y
d
etailed
n
u
m
e
r
ical
d
ata
in
T
ab
le
3
(
b
)
.
(
a)
(
b
)
Fig
u
r
e
3
.
Gr
a
p
h
ical
r
ep
r
esen
tatio
n
o
f
(
a)
co
n
f
u
s
io
n
m
atr
ix
an
d
(
b
)
p
er
f
o
r
m
an
ce
m
ea
s
u
r
es
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
14
,
No
.
4
,
Au
g
u
s
t
20
25
:
2
6
8
9
-
2
7
0
0
2696
T
h
e
SNN
ar
e
ANN
in
s
p
ir
ed
b
y
b
io
l
o
g
ical
n
e
u
r
o
n
s
,
u
tili
zin
g
s
p
ik
es
f
o
r
co
m
m
u
n
icatio
n
r
ath
er
th
an
weig
h
ted
s
u
m
s
o
f
in
p
u
ts
i
n
tr
ad
itio
n
al
ANN.
T
h
e
SNN
p
e
r
f
o
r
m
an
ce
is
ass
ess
ed
u
s
in
g
th
e
co
n
f
u
s
io
n
m
atr
ix
;
en
co
m
p
ass
in
g
ac
cu
r
ac
y
,
p
r
ec
i
s
io
n
,
r
ec
all,
F1
s
co
r
e
,
an
d
F2
s
co
r
e.
Mo
r
eo
v
e
r
,
th
e
g
r
a
p
h
p
o
r
tr
ay
s
co
n
s
is
ten
t
p
er
f
o
r
m
an
ce
ac
r
o
s
s
all
m
etr
ics,
in
d
icatin
g
r
o
b
u
s
tn
ess
an
d
g
en
er
aliza
tio
n
o
f
th
e
SNN
m
o
d
el
to
u
n
s
ee
n
d
ata.
T
h
is
s
u
g
g
ests
th
at
th
e
SNN
ef
f
ec
tiv
ely
av
o
i
d
s
o
v
er
f
itti
n
g
an
d
m
ai
n
tain
s
g
o
o
d
p
er
f
o
r
m
an
ce
o
n
v
ar
ied
d
atasets
.
I
n
s
u
m
m
ar
y
,
Fig
u
r
e
3
(
b
)
u
n
d
e
r
s
co
r
es
th
e
SNNs
p
o
ten
tial
as
p
r
o
m
is
in
g
n
eu
r
al
n
etwo
r
k
ar
ch
itectu
r
e
f
o
r
p
r
o
f
icien
tly
i
d
en
tify
in
g
v
o
i
ce
p
atter
n
s
in
FVC
task
s
.
T
ab
le
3
.
T
a
b
u
latio
n
o
f
r
esu
lts
: (
a)
co
n
f
u
s
io
n
m
atr
ix
an
d
(
b
)
p
er
f
o
r
m
a
n
ce
m
ea
s
u
r
e
(
a)
P
r
e
d
i
c
t
e
d
n
o
n
ma
t
c
h
P
r
e
d
i
c
t
e
d
ma
t
c
h
A
c
t
u
a
l
n
o
n
ma
t
c
h
T
r
u
e
n
e
g
a
t
i
v
e
s
F
a
l
se
p
o
si
t
i
v
e
s
A
c
t
u
a
l
ma
t
c
h
F
a
l
se
n
e
g
a
t
i
v
e
s
r
u
e
p
o
si
t
i
v
e
s
(
b
)
D
a
t
a
s
e
t
n
a
me
P
e
r
f
o
r
ma
n
c
e
me
a
su
r
e
s
R
e
s
u
l
t
s (%)
A
u
st
r
a
l
i
a
n
E
n
g
l
i
s
h
A
c
c
u
r
a
c
y
9
4
.
2
1
P
r
e
c
i
s
i
o
n
8
5
.
2
1
R
e
c
a
l
l
8
2
.
1
6
F
1
sc
o
r
e
8
1
.
1
1
F
2
sc
o
r
e
8
0
.
1
0
5
.
2
.
Rec
eiv
er
o
pera
t
ing
cha
r
a
ct
er
is
t
ic
a
nd
a
re
a
un
der
t
he
curv
e
a
na
ly
s
is
Fu
r
th
er
m
o
r
e
,
th
e
R
OC
cu
r
v
e
s
er
v
es
as
a
to
o
l
f
o
r
ass
ess
in
g
th
e
p
er
f
o
r
m
a
n
ce
o
f
th
e
class
if
icatio
n
m
o
d
el,
s
p
ec
if
ically
th
e
SNN.
T
h
e
d
eg
r
ee
o
f
s
ep
a
r
ab
ilit
y
is
in
d
icate
d
b
y
th
e
AUC
o
f
th
e
R
OC
cu
r
v
e.
T
h
e
R
OC
-
AU
C
cu
r
v
e
p
r
o
v
i
d
es
a
g
r
ap
h
ical
r
ep
r
esen
tatio
n
i
llu
s
tr
atin
g
th
e
class
if
icatio
n
p
er
f
o
r
m
a
n
ce
m
etr
ics
at
v
ar
io
u
s
th
r
esh
o
ld
s
.
T
h
e
R
OC
cu
r
v
e
is
a
g
r
a
p
h
ical
p
lo
t
th
at
d
ep
icts
th
e
p
er
f
o
r
m
an
ce
o
f
b
i
n
ar
y
class
if
ier
s
y
s
tem
as
its
th
r
esh
o
ld
v
ar
ie
s
.
I
t
is
ex
ten
s
iv
ely
u
tili
ze
d
in
m
ac
h
in
e
lear
n
in
g
to
ev
alu
ate
th
e
d
iag
n
o
s
ti
c
ca
p
ab
ilit
y
o
f
test
s
,
esp
ec
ially
in
s
ce
n
ar
io
s
with
im
b
alan
ce
d
o
u
tco
m
es.
I
n
th
is
co
n
tex
t,
th
e
b
in
ar
y
class
if
ier
is
r
ep
r
esen
ted
b
y
th
e
SNN,
wh
ich
aim
s
to
d
if
f
er
en
tiate
b
etwe
en
two
class
e
s
:
tr
u
e
p
o
s
itiv
es
an
d
f
alse
p
o
s
itiv
es,
as r
ep
r
esen
ted
in
Fig
u
r
e
4
(
a
)
.
T
h
e
tr
u
e
p
o
s
itiv
e
r
ate
(
T
PR
)
,
d
is
p
lay
ed
o
n
th
e
y
-
a
x
is
,
s
ig
n
if
ies
th
e
p
r
o
p
o
r
tio
n
o
f
co
r
r
ec
tly
class
if
ied
p
o
s
itiv
e
ca
s
es.
C
o
n
v
er
s
ely
,
th
e
f
alse
p
o
s
itiv
e
r
ate
(
FP
R
)
,
d
e
p
icted
o
n
th
e
x
-
ax
is
,
r
ep
r
esen
t
s
th
e
p
r
o
p
o
r
ti
o
n
o
f
in
co
r
r
ec
tly
class
if
ied
n
eg
ativ
e
ca
s
es.
T
h
e
ar
ea
u
n
d
er
t
h
e
R
OC
cu
r
v
e
(
AUC)
s
er
v
es
as
a
m
e
asu
r
e
o
f
th
e
o
v
e
r
all
p
er
f
o
r
m
an
ce
o
f
th
e
class
if
ier
.
A
p
er
f
ec
t
d
is
cr
im
in
atio
n
is
r
e
p
r
esen
ted
b
y
an
AUC
o
f
1
,
w
h
ile
an
AUC
o
f
0
.
9
4
in
d
icate
s
th
e
d
eg
r
ee
o
f
d
is
s
im
i
lar
ity
an
d
s
im
ilar
ity
in
th
e
v
o
i
ce
p
atter
n
s
.
T
h
e
r
ed
d
o
t
o
n
th
e
cu
r
v
e
d
en
o
tes
th
e
p
o
in
t
wh
er
e
th
e
T
PR
eq
u
als
th
e
FP
R
.
T
h
is
p
o
in
t
is
co
m
m
o
n
ly
r
ef
e
r
r
ed
t
o
as
th
e
"o
p
e
r
atin
g
p
o
in
t"
o
f
th
e
class
if
ier
,
wh
er
e
th
e
clas
s
if
ie
r
s
tr
ik
es
a
b
alan
ce
b
etwe
en
t
r
u
e
p
o
s
itiv
es
an
d
f
alse
p
o
s
iti
v
es,
m
ak
in
g
it
th
e
o
p
tim
al
th
r
esh
o
ld
f
o
r
class
if
ic
atio
n
.
T
h
e
b
lu
e
lin
e
in
t
h
e
g
r
a
p
h
r
ep
r
esen
ts
th
e
R
OC
f
o
r
a
r
an
d
o
m
class
if
ier
.
A
r
an
d
o
m
class
if
ier
ty
p
ically
p
r
o
d
u
ce
s
an
AUC
o
f
0
.
5
,
r
es
u
ltin
g
in
a
d
iag
o
n
al
lin
e
o
n
th
e
R
OC
cu
r
v
e.
Ho
wev
er
,
in
th
is
ca
s
e,
d
u
e
to
a
s
p
ec
if
ic
s
ce
n
ar
io
,
th
e
r
a
n
d
o
m
class
if
ier
'
s
AUC
is
m
en
tio
n
ed
as
0
.
9
4
.
T
h
e
o
b
s
er
v
atio
n
t
h
at
th
e
SNN
R
O
C
cu
r
v
e
lies
ab
o
v
e
th
e
R
OC
cu
r
v
e
o
f
t
h
e
r
an
d
o
m
class
if
ier
im
p
lies
th
at
th
e
SNN
ex
h
ib
its
s
u
p
er
io
r
p
er
f
o
r
m
an
ce
co
m
p
ar
ed
t
o
r
an
d
o
m
ch
a
n
ce
.
Nev
er
th
eless
,
th
e
m
ar
g
in
o
f
im
p
r
o
v
em
e
n
t
is
r
e
lativ
ely
m
o
d
est.
Ov
er
all,
th
e
g
r
ap
h
in
d
icate
s
th
at
th
e
SNN
p
er
f
o
r
m
s
s
atis
f
ac
to
r
ily
i
n
v
o
ice
co
m
p
ar
is
o
n
task
s
,
with
an
AUC
clo
s
e
to
9
4
.
2
1
%.
T
h
e
AUC
r
ep
r
esen
ts
th
e
en
tire
two
-
d
im
en
s
io
n
al
ar
ea
b
en
ea
th
th
e
R
OC
cu
r
v
e,
s
ig
n
if
y
in
g
t
h
e
class
if
ier
'
s
o
v
er
all
p
er
f
o
r
m
a
n
c
e.
Ma
th
em
atica
lly
,
th
e
AUC
i
s
d
eter
m
in
ed
b
y
ca
lcu
latin
g
t
h
e
d
e
f
in
ite
in
teg
r
al
o
f
th
e
f
u
n
ctio
n
f
(
x
)
with
r
esp
e
ct
to
th
e
v
er
tical
b
o
u
n
d
ar
ies,
a
s
d
escr
ib
ed
b
y
(
5
)
.
AUC
=
∫
a
b
f
(
x
)
dx
=
F
(
b
)
−
F
(
a
)
(
5
)
W
h
er
e
:
−
∫
ab
f
(
x
)
d
x
d
en
o
tes th
e
d
ef
in
ite
in
teg
r
al
o
f
th
e
f
u
n
ctio
n
f
(
x
)
o
v
er
th
e
in
ter
v
al
f
r
o
m
a
an
d
b
−
F(x
)
r
ep
r
esen
ts
th
e
a
n
tid
er
iv
at
iv
e
o
f
f
(
x
)
o
f
ten
r
ef
er
r
e
d
to
as
th
e
cu
m
u
lativ
e
d
is
tr
ib
u
tio
n
f
u
n
ctio
n
(
C
DF)
.
−
F(b
)
-
F(a
)
−
I
t
co
m
p
u
tes
th
e
d
if
f
e
r
en
ce
b
etwe
en
th
e
a
n
tid
er
iv
ativ
e
v
al
u
es
at
th
e
u
p
p
er
(
b
)
an
d
lo
wer
(
a)
b
o
u
n
d
s
,
r
ep
r
esen
tin
g
th
e
ac
cu
m
u
late
d
ar
ea
u
n
d
e
r
th
e
c
u
r
v
e
with
in
t
h
e
g
iv
en
in
ter
v
al.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
S
emi
-
a
u
to
ma
tic
vo
ice
c
o
mp
a
r
i
s
o
n
a
p
p
r
o
a
ch
u
s
in
g
s
p
ikin
g
n
e
u
r
a
l
…
(
K
r
u
th
ika
S
id
d
a
n
a
k
a
tte
Go
p
a
la
ia
h
)
2697
A
lar
g
er
R
OC
-
Ar
ea
s
ig
n
if
ies
im
p
r
o
v
e
d
ac
c
u
r
ac
y
o
f
th
e
class
if
ier
in
id
e
n
tify
in
g
in
d
iv
id
u
als.
Fig
u
r
e
4
(
a)
d
is
p
lay
s
s
ev
er
al
s
u
b
s
eq
u
en
t
R
OC
an
d
AU
C
g
r
ap
h
s
g
en
er
ated
d
u
r
in
g
th
e
an
al
y
s
is
.
T
h
e
g
r
ap
h
ical
h
is
to
g
r
am
r
e
p
r
esen
tatio
n
s
o
f
t
h
e
v
o
ice
co
m
p
ar
is
o
n
i
n
SNN
ar
e
s
h
o
w
n
in
Fig
u
r
e
4
(
b
)
.
T
h
e
v
o
ice
ca
lcu
lated
d
is
tr
ib
u
tio
n
in
r
elatio
n
to
am
p
litu
d
e
an
d
tim
e.
Peak
s
in
th
e
h
is
to
g
r
am
in
d
icate
th
e
r
an
g
e
an
d
d
is
tr
ib
u
tio
n
o
f
am
p
litu
d
es
in
a
v
o
ice
p
atter
n
o
r
au
d
io
s
ig
n
al.
I
n
Fig
u
r
e
4
(
b
)
,
au
d
io
A
an
d
au
d
io
B
r
e
p
r
ese
n
t
th
e
v
o
ice
s
am
p
le.
W
h
er
e
th
e
am
p
litu
d
e
r
ep
r
esen
ts
th
e
s
tr
en
g
th
o
r
i
n
ten
s
ity
o
f
th
e
s
o
u
n
d
wa
v
e
with
h
ig
h
er
a
m
p
litu
d
es
co
r
r
esp
o
n
d
in
g
to
lo
u
d
er
v
o
ic
e.
T
h
e
v
o
ice
p
atter
n
is
e
n
co
d
ed
in
S
NN
wh
e
n
th
e
f
r
e
q
u
en
cy
o
f
th
e
p
atter
n
r
ea
ch
es
a
ce
r
tain
th
r
esh
o
ld
.
Her
e
th
e
f
r
eq
u
e
n
cy
o
f
th
e
th
r
esh
o
ld
is
0
.
5
s
ec
o
n
d
s
is
tak
en
f
o
r
th
e
ex
p
er
im
en
tatio
n
.
Du
r
in
g
th
is
s
p
ik
e
tim
e
th
e
v
o
ice
p
atter
n
s
ar
e
id
en
tifie
d
to
k
n
o
w
th
e
s
im
il
ar
ity
o
f
th
e
k
n
o
wn
an
d
tr
ac
e.
(
a)
(
b
)
Fig
u
r
e
4
.
Gr
a
p
h
ical
r
ep
r
esen
tatio
n
o
f
SNN
r
esu
lts
(
a)
R
OC
-
AUC
g
r
ap
h
an
d
(
b
)
a
u
d
io
s
ig
n
al
h
is
to
g
r
am
p
lo
t
6.
CO
M
P
ARI
SO
N
WI
T
H
E
X
I
ST
I
NG
WO
RK
T
h
e
T
ab
le
4
p
r
esen
ts
a
co
m
p
r
eh
en
s
iv
e
co
m
p
ar
ativ
e
a
n
aly
s
is
o
f
ex
is
tin
g
r
esear
ch
s
tu
d
ies
alo
n
g
s
id
e
th
e
p
r
o
p
o
s
ed
SNN
ap
p
r
o
ac
h
f
o
r
th
e
Au
s
tr
alian
E
n
g
lis
h
d
ata
s
ets.
T
h
e
m
ain
ad
v
an
tag
e
o
f
S
NN
f
o
r
FVC
i
s
th
eir
ab
ilit
y
to
p
r
o
ce
s
s
s
p
ee
ch
d
ata
in
a
wa
y
th
at
cl
o
s
ely
r
esem
b
les
h
o
w
th
e
h
u
m
a
n
b
r
ain
p
r
o
c
ess
es
s
o
u
n
d
.
SNNs
ar
e
h
ig
h
ly
e
f
f
icien
t
at
ca
p
tu
r
in
g
th
e
tim
in
g
d
etails
in
s
p
ee
ch
,
s
u
ch
as
p
au
s
es,
p
itch
,
an
d
in
t
o
n
atio
n
,
wh
ic
h
ar
e
ess
en
tial
f
o
r
d
is
tin
g
u
is
h
in
g
b
etwe
en
s
p
ea
k
er
s
ac
cu
r
ately
.
T
h
ey
also
u
s
e
less
en
er
g
y
th
an
tr
ad
itio
n
al
n
e
u
r
al
n
etwo
r
k
s
,
m
a
k
in
g
th
em
id
ea
l
f
o
r
r
ea
l
-
tim
e
ap
p
licatio
n
s
.
Mo
r
eo
v
er
,
SNNs
ar
e
r
o
b
u
s
t
in
n
o
is
y
en
v
ir
o
n
m
e
n
ts
,
wh
ich
is
b
en
ef
icial
wh
e
n
an
al
y
zin
g
lo
w
-
q
u
ality
a
u
d
io
r
ec
o
r
d
in
g
s
o
f
ten
en
co
u
n
ter
e
d
in
f
o
r
en
s
ic
ca
s
es.
I
n
o
u
r
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
14
,
No
.
4
,
Au
g
u
s
t
20
25
:
2
6
8
9
-
2
7
0
0
2698
co
m
p
ar
ativ
e
s
tu
d
y
,
th
e
p
r
o
p
o
s
ed
f
r
am
ewo
r
k
ac
h
ie
v
ed
a
s
im
ilar
ity
ac
cu
r
ac
y
o
f
9
4
.
2
1
%,
o
u
tp
er
f
o
r
m
in
g
ex
is
tin
g
m
eth
o
d
s
.
T
h
ese
f
ea
tu
r
es
m
ak
e
SNNs
p
ar
ticu
lar
ly
w
ell
-
s
u
ited
f
o
r
th
e
c
o
m
p
lex
ch
a
llen
g
es
o
f
f
o
r
en
s
ic
v
o
ice
an
aly
s
is
.
Ov
er
all,
a
SNN
m
o
d
el
m
a
y
s
u
f
f
ice,
p
ar
ticu
l
ar
ly
f
o
r
ca
p
tu
r
in
g
tem
p
o
r
al
p
a
tter
n
s
in
s
p
ik
e
tim
es
f
o
r
v
o
ice
p
atter
n
r
ec
o
g
n
itio
n
.
T
h
e
co
m
p
a
r
is
o
n
u
n
d
er
s
co
r
es
th
e
v
er
s
atility
o
f
SNN
m
o
d
e
ls
ac
r
o
s
s
d
if
f
er
en
t
ap
p
licatio
n
s
,
ex
h
ib
itin
g
v
ar
y
i
n
g
lev
els
o
f
ac
cu
r
ac
y
,
ef
f
icien
cy
,
an
d
r
o
b
u
s
tn
ess
co
n
tin
g
e
n
t
u
p
o
n
th
e
s
p
ec
if
ic
d
ataset
a
n
d
ap
p
licatio
n
co
n
tex
t.
E
x
is
tin
g
r
esear
ch
s
tu
d
ies
u
tili
ze
d
iv
er
s
e
d
atasets
,
r
an
g
in
g
f
r
o
m
e
n
v
ir
o
n
m
en
tal
s
o
u
n
d
d
atasets
lik
e
R
W
C
P
an
d
s
p
o
k
en
d
ig
i
ts
d
ataset
s
s
u
ch
as
T
I
DI
GI
T
S
to
m
o
r
e
s
p
ec
ialized
d
atase
ts
lik
e
th
o
s
e
f
r
o
m
in
d
u
s
tr
ial
s
en
s
o
r
d
ata
a
n
aly
s
is
.
Fo
r
in
s
tan
ce
,
m
eth
o
d
o
l
o
g
ie
s
lik
e
th
e
SOM
-
SNN
m
o
d
el
ac
h
iev
e
im
p
r
ess
iv
e
ac
cu
r
ac
ies
o
f
9
7
.
4
0
%
an
d
9
9
.
6
0
%
o
n
th
e
T
I
DI
G
I
T
S
s
p
o
k
en
d
ig
its
an
d
R
W
C
P
en
v
ir
o
n
m
e
n
tal
s
o
u
n
d
d
atasets
,
r
esp
ec
tiv
ely
.
C
o
n
v
er
s
ely
,
SNNs
em
p
lo
y
ed
f
o
r
ac
o
u
s
tic
m
o
d
elin
g
d
em
o
n
s
tr
ate
lo
wer
ac
cu
r
ac
y
r
ates
o
f
3
6
.
9
%
an
d
1
8
.
7
%
o
n
th
e
T
I
MI
T
C
o
r
p
u
s
an
d
L
ib
r
is
p
ee
ch
d
atasets
.
I
n
a
s
tu
d
y
f
o
cu
s
ed
o
n
b
all
b
e
ar
in
g
d
iag
n
o
s
is
,
an
ANN
ac
h
iev
es
6
1
%
ac
cu
r
ac
y
.
Fu
r
th
er
m
o
r
e,
a
m
u
ltil
ay
er
SNN
d
esig
n
ed
f
o
r
au
d
i
o
s
am
p
le
class
if
icatio
n
u
s
in
g
Sp
iNNak
er
ex
h
ib
its
r
o
b
u
s
tn
es
s
an
d
e
f
f
icien
cy
in
n
eu
r
o
m
o
r
p
h
ic
en
g
in
ee
r
in
g
,
ac
h
iev
in
g
a
n
ac
cu
r
ac
y
o
f
8
5
%.
Ho
wev
er
,
th
ese
m
eth
o
d
o
lo
g
ie
s
o
f
ten
in
teg
r
ate
a
d
v
an
ce
d
SNN
tech
n
o
lo
g
y
with
o
th
e
r
m
o
d
els.
I
n
co
n
tr
ast,
th
e
p
r
o
p
o
s
ed
r
esear
ch
ex
cl
u
s
iv
ely
em
p
lo
y
s
SNN,
wh
ich
is
ad
v
an
tag
e
o
u
s
co
n
s
id
er
in
g
f
a
cto
r
s
lik
e
r
eso
u
r
ce
co
n
s
tr
ain
ts
an
d
th
e
u
n
a
v
aila
b
ilit
y
o
f
a
d
v
an
ce
d
s
y
s
tem
s
lik
e
GPU.
T
h
e
SNN
with
a
d
ju
s
tab
le
th
r
esh
o
ld
ef
f
ec
tiv
ely
d
ete
r
m
in
es
th
e
s
im
ilar
ity
o
r
d
is
s
im
ilar
ity
o
f
s
p
ik
e
tr
ain
s
in
v
o
ice
s
am
p
les.
Per
f
o
r
m
an
ce
ev
al
u
atio
n
u
tili
zin
g
a
co
n
f
u
s
io
n
m
at
r
ix
with
its
ex
ten
d
ed
m
etr
ic
v
alu
es
lik
e
ac
cu
r
ac
y
9
4
.
2
1
%,
p
r
ec
is
io
n
8
5
.
2
1
%,
r
ec
all
8
2
.
1
6
%,
F1
s
co
r
e
8
1
.
1
1
%,
an
d
F2
s
co
r
e
8
0
.
1
0
%
ar
e
ac
h
iev
ed
.
T
h
e
r
e
ar
e
ch
allen
g
es
in
a
p
r
o
p
o
s
ed
f
r
a
m
ewo
r
k
o
f
SNNs
f
o
r
FVC
s
u
ch
as c
o
m
p
lex
an
d
tim
e
-
c
o
n
s
u
m
in
g
tr
ain
in
g
d
u
e
to
s
p
ik
e
-
b
ased
lear
n
i
n
g
m
ec
h
an
is
m
s
lik
e
STDP.
T
h
ey
also
r
eq
u
ir
e
s
p
ec
ialized
n
eu
r
o
m
o
r
p
h
ic
h
ar
d
war
e
,
lim
itin
g
ac
ce
s
s
ib
ilit
y
.
T
ab
le
4
.
C
o
m
p
a
r
is
o
n
o
f
th
e
p
r
o
p
o
s
ed
ap
p
r
o
ac
h
with
ex
is
tin
g
wo
r
k
C
i
t
a
t
i
o
n
D
a
t
a
s
e
t
M
e
t
h
o
d
R
e
s
u
l
t
s i
n
(
%)
M
o
r
a
l
e
s
e
t
a
l
. [
25
]
P
u
r
e
t
o
n
e
sam
p
l
e
s
S
N
N
,
S
p
i
N
N
a
k
e
r
85
W
u
e
t
a
l
.
[
19
]
R
W
C
P
,
TI
D
I
G
I
TS D
i
sag
r
e
e
S
O
M
-
S
N
N
S
N
N
f
o
r
s
p
a
t
i
o
t
e
mp
o
r
a
l
p
a
t
t
e
r
n
& 9
7
.
4
0
,
9
9
.
6
0
W
u
e
t
a
l
.
[
20
]
TI
M
I
T,
Li
b
r
i
s
p
e
e
c
h
,
F
A
M
E
S
N
N
,
M
F
C
C
,
F
B
A
N
K
,
F
M
L
LR
3
6
.
9
,
1
8
.
7
A
u
g
e
e
t
a
l
. [
21
]
k
e
y
w
o
r
d
s
p
o
t
t
i
n
g
S
N
N
,
M
F
C
C
80
K
h
o
l
k
i
n
e
t
a
l
.
[
24
]
A
c
c
e
l
e
r
o
met
e
r
D
a
t
a
R
C
N
e
t
,
A
N
N
,
S
N
N
S
N
N
=
1
0
0
%
,
A
N
N
=
6
1
P
r
o
p
o
se
d
a
p
p
r
o
a
c
h
m
o
d
e
l
A
u
st
r
a
l
i
a
n
E
n
g
l
i
s
h
S
N
N
9
4
.
2
1
7.
CO
NCLU
SI
O
N
T
h
e
wo
r
k
p
r
o
p
o
s
ed
ex
p
lo
r
es
th
e
p
o
ten
tial
o
f
FVC
to
e
n
h
a
n
ce
s
u
s
p
ec
t
id
en
tific
atio
n
u
s
i
n
g
f
o
r
en
s
ic
s
p
ee
ch
r
ec
o
r
d
i
n
g
s
.
I
t
ap
p
lies
an
SNN
m
o
d
el
to
an
aly
ze
an
Au
s
tr
alian
E
n
g
lis
h
d
ataset
o
f
3
,
8
9
9
.
f
lac
f
ile
r
ec
o
r
d
in
g
s
,
u
tili
zin
g
s
tatio
n
ar
y
n
o
is
e
r
ed
u
ctio
n
f
o
r
p
r
e
-
p
r
o
c
ess
in
g
.
T
h
e
SNN
m
o
d
e
l
u
s
es
a
th
r
esh
o
ld
to
ass
ess
s
p
ik
e
tr
ain
s
im
ilar
ities
,
wh
er
e
n
eu
r
o
n
s
co
m
m
u
n
icate
v
ia
d
is
cr
ete
s
p
ik
es
th
r
o
u
g
h
m
e
m
b
r
an
e
p
o
ten
tials
.
Sy
n
ap
tic
weig
h
ts
,
u
p
d
ated
u
s
i
n
g
STDP
o
r
its
v
ar
ian
ts
,
h
elp
r
ec
o
g
n
ize
an
d
d
ec
o
d
e
v
o
ice
p
at
ter
n
s
.
Ou
r
f
in
d
in
g
s
p
r
o
v
id
e
co
n
clu
s
iv
e
e
v
id
en
ce
th
at
th
is
p
h
en
o
m
en
o
n
is
ass
o
ciate
d
with
a
SNN
m
o
d
el
ac
h
iev
es
9
4
.
2
1
%
ac
cu
r
ac
y
.
Fo
r
f
u
t
u
r
e
s
tu
d
ies
m
ay
in
v
esti
g
ate
o
n
r
ef
in
i
n
g
th
e
SNN
ar
ch
itectu
r
e
to
en
h
an
ce
r
ea
l
-
wo
r
ld
f
o
r
en
s
ic
ap
p
licatio
n
s
.
ACK
NO
WL
E
DG
E
M
E
NT
S
T
h
e
au
t
h
o
r
s
g
r
atef
u
lly
to
R
o
s
e
P.,
Z
h
an
g
C
.
,
an
d
Ge
o
f
f
r
ey
Stewar
t
M
o
r
r
is
o
n
f
r
o
m
th
e
FVC
L
ab
o
r
ato
r
y
,
UNSW
,
Sy
d
n
ey
,
Au
s
tr
alia,
f
o
r
p
r
o
v
id
in
g
ac
ce
s
s
to
th
eir
d
atab
ase
f
o
r
th
is
s
tu
d
y
.
F
UNDING
I
NF
O
R
M
A
T
I
O
N
T
h
is
r
esear
ch
is
s
u
p
p
o
r
ted
b
y
t
h
e
Dep
ar
tm
en
t o
f
Scien
ce
an
d
T
ec
h
n
o
lo
g
y
,
New
Delh
i,
I
n
d
i
a,
th
r
o
u
g
h
th
e
f
u
n
d
am
en
tal
r
esear
c
h
g
r
an
t
s
ch
em
e
DST
W
I
SE
-
Kir
an
f
ello
wsh
ip
f
o
r
Ph
.
D
.
Fil
e
No
:
DST/WISE
-
Ph
D/E
T
/2
0
2
3
/4
(
G)
.
AUTHO
R
CO
NT
RI
B
UT
I
O
NS ST
A
T
E
M
E
N
T
T
h
is
jo
u
r
n
al
u
s
es
th
e
C
o
n
tr
ib
u
to
r
R
o
les
T
ax
o
n
o
m
y
(
C
R
ed
iT)
to
r
ec
o
g
n
ize
in
d
iv
id
u
al
au
th
o
r
co
n
tr
ib
u
tio
n
s
,
r
ed
u
ce
au
th
o
r
s
h
ip
d
is
p
u
tes,
an
d
f
ac
ilit
ate
co
llab
o
r
atio
n
.
Evaluation Warning : The document was created with Spire.PDF for Python.