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
.
1
4
,
No
.
1
,
Feb
r
u
ar
y
2
0
2
5
,
p
p
.
1
30
~
1
4
1
I
SS
N:
2252
-
8
9
3
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijai
.
v
1
4
.
i
1
.
pp
1
30
-
1
4
1
130
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e
.
co
m
Discrimi
na
tive d
e
ep learning
bas
ed
hybrid sp
ec
tro
-
te
mpo
ra
l
features
for sy
nth
etic
v
o
ice spo
o
fin
g
det
ec
tion
P
ra
nita
Nira
j
P
a
ls
a
pu
re
,
Ra
j
eswa
ri,
Sa
nd
ee
p K
um
a
r
K
em
peg
o
wda
,
K
um
bh
a
r
T
rupti
Ra
v
ik
um
a
r
D
e
p
a
r
t
me
n
t
o
f
El
e
c
t
r
o
n
i
c
s
a
n
d
C
o
mm
u
n
i
c
a
t
i
o
n
En
g
i
n
e
e
r
i
n
g
,
A
c
h
a
r
y
a
I
n
st
i
t
u
t
e
o
f
Te
c
h
n
o
l
o
g
y
,
V
i
s
v
e
s
v
a
r
a
y
a
T
e
c
h
n
o
l
o
g
i
c
a
l
U
n
i
v
e
r
si
t
y
,
B
e
l
a
g
a
v
i
,
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
Oct
16
,
2
0
2
3
R
ev
is
ed
J
an
11
,
2
0
2
4
Acc
ep
ted
Feb
21
,
2
0
2
4
Vo
ice
-
b
a
se
d
sy
ste
m
s
li
k
e
sp
e
a
k
e
r
id
e
n
ti
fica
ti
o
n
s
y
ste
m
s
(S
IS
)
a
n
d
a
u
to
m
a
ti
c
sp
e
a
k
e
r
v
e
rifi
c
a
ti
o
n
sy
ste
m
s
(AS
V)
a
re
p
ro
li
fe
ra
ti
n
g
a
c
ro
ss
i
n
d
u
str
ies
su
c
h
a
s
fin
a
n
c
e
a
n
d
h
e
a
lt
h
c
a
re
d
u
e
to
t
h
e
ir
u
ti
l
it
y
i
n
i
d
e
n
ti
t
y
v
e
rifi
c
a
ti
o
n
th
r
o
u
g
h
u
n
i
q
u
e
sp
e
e
c
h
p
a
tt
e
rn
a
n
a
ly
sis
.
De
sp
it
e
th
e
ir
a
d
v
a
n
c
e
m
e
n
ts,
ASVs
a
re
su
sc
e
p
ti
b
le
t
o
v
a
ri
o
u
s
s
p
o
o
fin
g
a
t
tac
k
s,
in
c
lu
d
i
n
g
l
o
g
ica
l
a
n
d
re
p
lay
a
tt
a
c
k
s,
p
o
sin
g
c
h
a
ll
e
n
g
e
s
d
u
e
t
o
th
e
s
o
p
h
isti
c
a
ted
a
c
o
u
st
ic
d
isti
n
c
ti
o
n
s
b
e
twe
e
n
a
u
th
e
n
ti
c
a
n
d
sp
o
o
fe
d
v
o
ice
s
.
To
c
o
u
n
tera
c
t,
th
is
stu
d
y
p
r
o
p
o
se
s
a
ro
b
u
st
y
e
t
c
o
m
p
u
t
a
ti
o
n
a
ll
y
e
fficie
n
t
c
o
u
n
ter
m
e
a
su
re
sy
ste
m
,
u
ti
li
z
in
g
a
sy
ste
m
a
ti
c
d
a
ta
p
ro
c
e
ss
in
g
p
i
p
e
li
n
e
c
o
u
p
led
with
a
h
y
b
ri
d
sp
e
c
tral
-
tem
p
o
ra
l
lea
rn
in
g
a
p
p
ro
a
c
h
.
T
h
e
a
im
is
t
o
id
e
n
ti
fy
e
ffe
c
ti
v
e
fe
a
tu
re
s
th
a
t
o
p
ti
m
ize
t
h
e
m
o
d
e
l'
s
d
e
tec
ti
o
n
a
c
c
u
ra
c
y
a
n
d
c
o
m
p
u
tatio
n
a
l
e
fficie
n
c
y
.
Th
e
m
o
d
e
l
a
c
h
iev
e
d
su
p
e
rio
r
p
e
rfo
rm
a
n
c
e
with
a
n
a
c
c
u
ra
c
y
o
f
9
9
.
4
4
%
a
n
d
a
n
e
q
u
a
l
e
rro
r
ra
te
(EE
R)
o
f
0
.
0
1
4
i
n
th
e
lo
g
ica
l
a
c
c
e
ss
sc
e
n
a
rio
o
f
th
e
A
S
Vs
p
o
o
f
2
0
1
9
c
h
a
ll
e
n
g
e
,
d
e
m
o
n
stra
ti
n
g
it
s
e
n
h
a
n
c
e
d
a
c
c
u
ra
c
y
a
n
d
re
li
a
b
il
i
ty
i
n
d
e
tec
ti
n
g
sp
o
o
f
in
g
at
tac
k
s with
m
in
imiz
e
d
e
rr
o
r
m
a
rg
in
.
K
ey
w
o
r
d
s
:
Au
to
m
atic
s
p
ea
k
er
v
e
r
if
icatio
n
Hy
b
r
id
f
ea
t
u
r
e
lear
n
i
n
g
L
STM
-
C
NN
Sp
ec
tr
al
-
tem
p
o
r
al
f
ea
t
u
r
e
Sp
o
o
f
in
g
attac
k
d
etec
tio
n
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
:
Pra
n
ita
Nir
aj
Pals
ap
u
r
e
Dep
ar
tm
en
t o
f
E
lectr
o
n
ics
an
d
C
o
m
m
u
n
icatio
n
E
n
g
in
ee
r
i
n
g
,
Ach
ar
y
a
I
n
s
titu
te
o
f
T
ec
h
n
o
lo
g
y
Vis
v
esv
ar
ay
a
T
ec
h
n
o
l
o
g
ical
Un
iv
er
s
ity
B
elag
av
i
5
9
0
0
1
8
,
Kar
n
atak
a,
I
n
d
ia
E
m
ail: p
r
an
itan
ir
ajp
alsap
u
r
e@
g
m
ail
.
co
m
1.
I
NT
RO
D
UCT
I
O
N
I
n
to
d
ay
'
s
d
ig
ital
wo
r
ld
,
wh
e
r
e
p
er
s
o
n
al
d
ata
an
d
s
ec
u
r
ity
ar
e
p
ar
a
m
o
u
n
t,
b
io
m
etr
ic
au
t
h
en
ticatio
n
d
eliv
er
s
a
u
n
iq
u
e
a
n
d
co
n
v
en
ien
t
way
to
id
en
tify
in
d
i
v
id
u
a
ls
b
ased
o
n
th
eir
p
h
y
s
io
lo
g
ical
tr
aits
r
ath
er
th
an
tr
ad
itio
n
al
p
ass
wo
r
d
s
an
d
PI
Ns
[
1
]
,
[
2
]
.
Am
o
n
g
v
a
r
io
u
s
b
io
m
etr
ic
m
o
d
alities
,
v
o
ice
-
b
ased
s
y
s
tem
s
ar
e
a
p
r
o
m
is
in
g
tech
n
o
lo
g
y
with
s
e
v
er
al
ad
v
a
n
tag
es c
o
m
p
ar
ed
to
o
th
er
b
i
o
m
etr
ic
m
o
d
alities
s
u
ch
as f
in
g
e
r
p
r
in
t a
n
d
f
ac
ial
r
ec
o
g
n
itio
n
[
3
]
.
V
o
ice
-
b
ased
au
th
en
ticatio
n
-
b
ased
a
p
p
r
o
ac
h
es
ar
e
m
o
r
e
u
s
e
r
-
f
r
ien
d
ly
b
ec
au
s
e
th
e
y
d
o
n
o
t
r
eq
u
ir
e
p
h
y
s
ical
co
n
tact
f
r
o
m
th
e
u
s
er
,
th
u
s
p
r
eser
v
in
g
p
r
iv
ac
y
an
d
allo
win
g
f
o
r
f
r
ictio
n
less
u
s
er
au
th
en
ticatio
n
in
r
e
m
o
te
en
v
ir
o
n
m
en
ts
[
4
]
.
B
asically
,
v
o
ic
e
-
b
ased
b
io
m
etr
ic
a
p
p
licatio
n
s
em
p
lo
y
a
s
p
ea
k
er
r
ec
o
g
n
itio
n
s
y
s
tem
(
SIS)
an
d
au
to
m
atic
s
p
ea
k
er
v
er
if
icatio
n
s
y
s
tem
(
A
VS)
.
SIS
ar
e
d
esig
n
ed
to
d
eter
m
in
e
a
s
p
ea
k
er
'
s
id
en
tity
,
wh
ile
A
SV
is
u
s
ed
to
co
n
f
ir
m
th
at
th
e
p
r
esen
ted
s
p
ea
k
er
'
s
id
en
tity
i
s
w
h
o
h
e
claim
s
to
b
e
[
5
]
.
ASV
s
y
s
tem
s
ar
e
wid
ely
u
s
ed
in
s
ec
u
r
ity
-
s
en
s
itiv
e
in
d
u
s
tr
ies,
s
u
ch
as
ac
ce
s
s
co
n
tr
o
l,
u
s
er
au
th
en
ticatio
n
,
f
in
an
cial
s
er
v
ices,
v
ir
t
u
al
ass
is
tan
ts
,
telec
o
m
m
u
n
icatio
n
s
,
an
d
h
ea
lth
ca
r
e
[
6
]
.
B
y
a
n
aly
zin
g
s
p
ee
ch
p
atter
n
s
a
n
d
ch
ar
ac
ter
is
tics
,
ASVs
ca
n
p
r
ev
en
t
u
n
a
u
th
o
r
ize
d
ac
ce
s
s
,
r
ed
u
c
e
id
en
tity
f
r
au
d
,
an
d
im
p
r
o
v
e
u
s
er
ex
p
er
ien
ce
[
7
]
.
Alth
o
u
g
h
ASV
s
y
s
tem
s
h
av
e
p
r
o
m
is
in
g
b
en
ef
its
,
t
h
ey
ar
e
n
o
t
im
m
u
n
e
to
v
o
ice
s
p
o
o
f
in
g
attac
k
s
,
in
wh
ich
an
attac
k
er
attem
p
ts
to
tr
ick
th
e
s
y
s
tem
b
y
im
itatin
g
a
leg
itima
te
u
s
er
'
s
v
o
ice
[
8
]
.
R
ep
lay
attac
k
s
in
v
o
lv
e
r
ec
o
r
d
in
g
a
g
en
u
in
e
u
s
er
'
s
v
o
ice
an
d
p
lay
in
g
it
b
ac
k
to
tr
ick
th
e
ASV
s
y
s
tem
in
to
th
in
k
in
g
th
at
th
e
at
tack
er
is
th
e
g
en
u
in
e
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
Dis
crim
in
a
tive
d
ee
p
lea
r
n
in
g
b
a
s
ed
h
yb
r
id
s
p
ec
tr
o
-
temp
o
r
a
l fe
a
tu
r
es fo
r
…
(
P
r
a
n
ita
N
ir
a
j P
a
ls
a
p
u
r
e
)
131
s
p
ea
k
er
[
9
]
.
Ad
d
itio
n
ally
,
s
p
ee
ch
s
y
n
th
esis
u
s
es
tex
t
-
to
-
s
p
ee
c
h
(
T
T
S)
tech
n
o
lo
g
ies
to
cr
ea
te
s
p
ee
ch
th
at
m
im
ics
th
e
g
en
u
in
e
s
p
ea
k
e
r
'
s
v
o
ice
f
e
atu
r
es
to
cr
ea
te
r
ea
lis
tic
-
s
o
u
n
d
in
g
s
p
ee
c
h
to
s
p
o
o
f
th
e
ASV
s
y
s
tem
[
1
0
]
.
Vo
ice
tr
an
s
f
o
r
m
atio
n
attac
k
s
in
v
o
lv
e
m
o
d
i
f
y
i
n
g
an
attac
k
er
'
s
v
o
ice
to
r
esem
b
le
a
r
ea
l
s
p
ea
k
e
r
'
s
,
b
y
p
ass
in
g
th
e
ASV
s
y
s
tem
[
1
1
]
.
As tec
h
n
o
lo
g
y
co
n
tin
u
es to
ev
o
lv
e,
attac
k
e
r
s
will in
cr
ea
s
in
g
ly
ad
o
p
t m
o
r
e
s
o
p
h
is
ticated
m
eth
o
d
s
to
d
ec
eiv
e
ASV
s
y
s
tem
s
,
m
ak
in
g
it
ch
allen
g
in
g
f
o
r
an
ti
-
s
p
o
o
f
in
g
s
y
s
tem
s
t
o
r
em
ain
ef
f
ec
tiv
e
[
1
2
]
.
An
o
th
er
ch
allen
g
e
is
h
an
d
lin
g
th
e
q
u
alit
y
o
f
tr
ain
in
g
d
ata
s
ets,
wh
ich
ar
e
o
f
ten
af
f
ec
ted
b
y
h
i
g
h
d
im
en
s
io
n
s
an
d
tem
p
o
r
al
d
y
n
am
ics.
Fu
r
th
e
r
m
o
r
e,
au
th
e
n
tic,
an
d
s
p
o
o
f
ed
s
p
ee
c
h
ac
o
u
s
tic
f
ea
tu
r
es
m
ay
b
e
v
e
r
y
s
im
ilar
,
r
ep
r
esen
tin
g
a
n
i
n
ter
-
class
s
im
ilar
ity
p
r
o
b
lem
.
W
ith
th
ese
co
m
p
lex
ities
an
d
ad
v
an
ce
s
in
ar
tific
ial
in
tellig
en
c
e
tech
n
o
lo
g
y
,
v
o
ice
s
p
o
o
f
in
g
attac
k
s
s
er
io
u
s
ly
th
r
e
aten
th
e
r
eliab
ilit
y
o
f
ASV
s
y
s
tem
s
[
1
3
]
.
Hen
ce
,
r
o
b
u
s
t
an
ti
-
s
p
o
o
f
in
g
m
ea
s
u
r
es
ar
e
ess
en
tial
to
co
u
n
ter
v
ar
io
u
s
f
o
r
m
s
o
f
th
is
th
r
ea
t.
T
h
is
p
a
p
er
in
tr
o
d
u
ce
s
a
co
m
p
u
tatio
n
a
lly
ef
f
icien
t
d
esig
n
o
f
an
in
tellig
en
t
v
o
ice
s
p
o
o
f
in
g
co
u
n
ter
m
ea
s
u
r
e
s
y
s
tem
f
o
r
ASV
th
at
lev
er
ag
es
a
s
y
s
tem
atic
d
ata
p
r
o
ce
s
s
in
g
p
ip
elin
e,
m
u
lti
-
lev
el
au
d
io
f
ea
t
u
r
es
m
o
d
ellin
g
an
d
th
e
s
tr
en
g
t
h
s
o
f
a
h
y
b
r
i
d
s
p
ec
tr
al
-
tem
p
o
r
allea
r
n
in
g
m
o
d
el
to
class
if
y
th
e
g
en
u
in
e
a
n
d
s
p
o
o
f
ed
v
o
ice
ef
f
ec
tiv
ely
.
I
n
a
r
ec
en
t
s
tate
o
f
t
h
e
ar
t,
s
ig
n
if
ican
t
r
esear
ch
h
as
b
ee
n
d
o
n
e
to
ad
d
r
ess
th
e
ch
allen
g
es
p
o
s
e
d
b
y
v
o
ice
s
p
o
o
f
in
g
attac
k
s
.
R
esear
ch
er
s
h
av
e
ex
p
lo
r
ed
m
a
n
y
an
ti
-
s
p
o
o
f
in
g
tech
n
iq
u
es,
in
clu
d
i
n
g
f
e
atu
r
e
-
b
ased
,
d
ee
p
-
lear
n
in
g
ap
p
r
o
ac
h
es
a
n
d
s
o
p
h
is
ticated
m
o
d
els.
Ma
g
az
i
n
e
et
a
l.
[
1
4
]
h
ig
h
lig
h
ted
th
e
p
o
ten
tial
th
r
ea
t
o
f
d
ee
p
f
ak
es
t
o
ASV
tech
n
o
lo
g
i
es.
T
h
ey
in
tr
o
d
u
ce
d
a
d
etec
tio
n
a
p
p
r
o
ac
h
o
n
m
o
d
u
latio
n
-
s
p
ec
tr
o
g
r
a
m
f
ea
tu
r
es
th
at
ch
ar
ac
ter
izes
s
ess
io
n
id
e
n
tity
,
g
en
d
er
,
an
d
th
e
s
o
u
r
ce
o
f
g
en
er
atio
n
v
ar
iatio
n
to
d
if
f
e
r
en
t
iate
b
etwe
en
th
e
m
.
Z
h
an
g
et
a
l.
[
1
5
]
in
tr
o
d
u
ce
d
a
o
n
e
-
class
lear
n
in
g
m
eth
o
d
to
d
etec
t
T
T
S
-
g
e
n
er
ated
s
p
o
o
f
ed
v
o
ices.
T
h
eir
ap
p
r
o
ac
h
r
e
v
o
lv
es
ar
o
u
n
d
c
o
m
p
ac
tin
g
g
en
u
i
n
e
s
p
ee
ch
r
ep
r
esen
tatio
n
s
an
d
in
tr
o
d
u
cin
g
an
an
g
u
lar
m
ar
g
in
,
with
o
u
t
d
ata
au
g
m
en
tatio
n
,
to
d
is
tin
g
u
is
h
b
etwe
en
g
en
u
in
e
a
n
d
s
p
o
o
f
e
d
v
o
ices
in
th
e
em
b
e
d
d
in
g
s
p
ac
e.
W
h
en
test
ed
o
n
th
e
ASVs
p
o
o
f
2
0
1
9
d
ataset,
th
eir
m
o
d
el
ac
h
iev
e
d
a
n
eq
u
al
e
r
r
o
r
r
ate
(
E
E
R
)
o
f
2
.
1
9
%.
An
ap
p
licatio
n
o
f
en
s
em
b
le
lear
n
in
g
f
o
r
v
o
ic
e
an
ti
-
s
p
o
o
f
in
g
s
y
s
tem
is
p
r
esen
ted
b
y
Du
a
et
a
l.
,
[
1
6
]
wh
er
e
r
ec
u
r
r
en
t
n
e
u
r
al
n
etwo
r
k
(
R
NN)
an
d
co
n
v
o
lu
ti
o
n
al
n
eu
r
al
n
etwo
r
k
(
C
NN)
wer
e
tr
ain
ed
o
n
Mel
-
f
r
eq
u
e
n
cy
-
ce
p
s
tr
al
-
co
ef
f
icien
ts
(
MFC
C
)
ex
tr
ac
ted
f
r
o
m
th
e
A
SVs
p
o
o
f
d
ataset.
Z
h
an
g
et
a
l.
[
1
7
]
ex
p
lo
r
ed
a
v
o
ice
s
p
o
o
f
in
g
s
ce
n
ar
io
wh
er
e
f
a
k
e
v
o
ice
s
ig
n
als
g
en
er
ated
u
s
in
g
T
T
S
a
r
e
em
b
ed
d
e
d
i
n
to
l
eg
itima
te
u
tter
an
ce
s
.
T
h
e
au
th
o
r
s
d
e
v
elo
p
ed
a
co
u
n
ter
m
ea
s
u
r
e
s
ch
em
e
u
tili
zi
n
g
s
elf
-
p
r
e
-
tr
ain
ed
m
o
d
els
f
o
r
f
ea
tu
r
e
e
x
tr
ac
tio
n
.
A
d
d
itio
n
al
ly
,
th
ey
aim
e
d
f
o
r
s
im
u
ltan
eo
u
s
u
tter
an
ce
-
an
d
s
e
g
m
en
t
-
lev
el
d
etec
tio
n
.
E
x
p
er
i
m
en
tally
,
th
eir
ap
p
r
o
ac
h
ac
h
ie
v
ed
E
E
R
s
o
f
0
.
7
7
%
o
n
th
e
Par
tialSp
o
o
f
d
atab
ase
a
n
d
0
.
9
0
% o
n
ASVs
p
o
o
f
2
0
19.
I
n
th
e
s
tu
d
y
o
f
Xu
e
et
a
l.
[
1
8
]
,
an
iter
ativ
e
k
n
o
wled
g
e
d
is
till
atio
n
m
eth
o
d
is
ad
o
p
te
d
f
o
r
f
a
k
e
s
p
ee
ch
d
etec
tio
n
wh
er
e
a
d
ee
p
n
et
wo
r
k
as
th
e
i
n
s
tr
u
cto
r
is
m
o
d
elled
to
g
u
id
e
m
u
ltip
le
s
h
allo
w
class
if
ier
s
b
y
m
in
im
izin
g
f
ea
tu
r
e
d
if
f
e
r
en
ce
s
.
L
ei
et
a
l.
[
1
9
]
d
e
v
elo
p
ed
a
m
eth
o
d
to
d
etec
t
k
n
o
wn
an
d
u
n
k
n
o
wn
s
p
o
o
f
in
g
attac
k
s
u
s
in
g
1
-
D
C
NN,
Siam
ese
C
NN,
an
d
Gau
s
s
ian
m
ix
tu
r
e
m
o
d
el
(
GM
M)
co
m
p
o
n
e
n
ts
to
ca
p
tu
r
e
l
o
ca
l
an
d
g
lo
b
al
s
p
ee
ch
f
ea
tu
r
es.
W
u
et
a
l.
[
2
0
]
in
tr
o
d
u
ce
d
th
e
f
ea
tu
r
e
en
g
in
ee
r
in
g
tech
n
iq
u
e,
w
h
ich
u
s
es
a
tr
a
n
s
f
o
r
m
er
tr
ain
ed
o
n
a
g
en
u
in
e
s
p
ee
ch
f
r
o
m
th
e
ASVs
p
o
o
f
2
0
1
9
lo
g
ic
al
ac
ce
s
s
co
r
p
u
s
to
id
en
tify
an
d
r
em
o
v
e
s
p
o
o
f
in
g
ar
tef
ac
ts
.
J
av
ed
et
a
l.
[
2
1
]
d
ev
elo
p
ed
a
f
r
am
ewo
r
k
th
at
u
s
es
co
-
o
cc
u
r
r
en
ce
p
atter
n
s
an
d
ce
p
s
tr
al
co
ef
f
icien
ts
to
d
etec
t
d
is
to
r
tio
n
s
a
n
d
a
r
tef
ac
t
s
in
d
u
ce
d
b
y
d
if
f
er
en
t
s
p
o
o
f
i
n
g
m
et
h
o
d
s
,
p
r
o
v
id
in
g
co
m
p
r
eh
en
s
iv
e
p
r
o
tectio
n
ag
ain
s
t
ev
en
co
m
p
lex
s
p
o
o
f
in
g
attac
k
s
.
T
h
e
e
f
f
ec
tiv
en
ess
o
f
h
y
b
r
id
izin
g
d
if
f
er
e
n
t
lear
n
i
n
g
m
o
d
els,
n
am
el
y
C
NN
f
o
r
f
ea
tu
r
e
ex
tr
ac
tio
n
a
n
d
s
u
p
p
o
r
t
v
ec
to
r
m
ac
h
in
e
(
SVM
)
f
o
r
class
if
ica
tio
n
task
s
,
is
s
tu
d
ied
b
y
[
2
2
]
.
Kwa
k
et
a
l.
[
2
3
]
,
[
2
4
]
d
ev
el
o
p
ed
n
ew
m
o
d
els
th
at
ar
e
m
o
r
e
ef
f
icien
t
a
n
d
r
o
b
u
s
t
to
u
n
s
ee
n
s
p
o
o
f
attac
k
s
.
Gu
o
et
a
l.
[
2
5
]
u
s
ed
in
c
r
em
en
tal
lear
n
in
g
to
im
p
r
o
v
e
th
e
g
e
n
er
aliza
b
ilit
y
o
f
s
p
o
o
f
d
etec
tio
n
m
o
d
els
to
u
n
s
ee
n
s
p
o
o
f
alg
o
r
ith
m
s
.
T
h
ey
d
is
cu
s
s
h
o
w
to
en
h
an
ce
th
ese
m
o
d
els
'
em
b
ed
d
in
g
s
p
ac
e
an
d
d
ec
is
io
n
b
o
u
n
d
a
r
ies
to
ad
ap
t
to
n
ew
s
p
o
o
f
i
n
g
th
r
ea
ts
.
Ma
lik
et
a
l.
[
2
6
]
ad
d
r
ess
ed
th
e
v
u
ln
e
r
ab
ilit
y
o
f
v
o
ice
-
ac
ti
v
ated
s
er
v
ices
lik
e
ch
atb
o
ts
to
au
d
io
r
e
p
lay
attac
k
s
.
T
h
ey
in
tr
o
d
u
ce
d
a
co
u
s
tic
ter
n
ar
y
p
atter
n
s
-
g
am
m
ato
n
e
ce
p
s
tr
al
co
ef
f
icien
t
an
d
u
s
ed
a
m
u
lti
-
class
SVM
class
i
f
ier
tr
ain
e
d
th
r
o
u
g
h
er
r
o
r
-
co
r
r
ec
tin
g
o
u
tp
u
t
co
d
es
o
n
th
e
o
p
t
im
al
f
ea
tu
r
e
s
p
ac
e.
T
h
e
m
o
d
el
p
r
esen
ted
b
y
Ad
ib
an
et
a
l.
[
2
7
]
u
tili
ze
s
o
n
e
-
clas
s
lear
n
in
g
f
o
r
d
etec
tin
g
s
y
n
th
etic
v
o
ice
s
p
o
o
f
in
g
.
An
E
E
R
o
f
2
.
1
9
d
ep
icts
m
o
d
e
r
ate
ef
f
icien
cy
in
m
in
im
izin
g
f
alse
ac
ce
p
tan
ce
an
d
r
ejec
tio
n
r
ates.
Hen
ce
,
it
ca
n
b
e
s
ee
n
th
at
r
esear
ch
er
s
h
av
e
p
r
o
p
o
s
ed
v
ar
io
u
s
s
tr
ateg
ies
to
co
u
n
ter
ac
t
v
o
ice
s
p
o
o
f
in
g
attac
k
s
,
y
et
m
an
y
ex
is
tin
g
m
eth
o
d
s
i
n
v
o
lv
e
c
o
m
p
lex
an
d
s
o
p
h
is
ticated
m
o
d
els.
Mo
s
t
cu
r
r
en
t
ap
p
r
o
ac
h
es
p
r
ed
o
m
in
an
tly
f
o
c
u
s
o
n
ex
tr
ac
tin
g
m
el
-
s
p
ec
tr
o
g
r
am
f
e
atu
r
es a
n
d
o
f
ten
em
p
lo
y
C
NN
f
o
r
class
if
icatio
n
.
Ho
wev
er
,
e
x
p
lo
r
in
g
alter
n
ativ
e
f
ea
tu
r
es
r
ep
r
esen
tin
g
b
o
t
h
tem
p
o
r
al
an
d
s
p
ec
tr
al
s
ig
n
al
ch
a
r
a
cter
is
tics
co
u
ld
p
o
ten
tially
en
h
an
ce
th
e
ef
f
icie
n
cy
o
f
d
etec
tin
g
v
o
ice
s
p
o
o
f
in
g
attac
k
s
o
n
ASV
s
y
s
tem
s
.
2.
M
E
T
H
O
D
T
h
is
s
ec
tio
n
d
etails
th
e
p
r
o
p
o
s
ed
s
y
s
tem
d
esig
n
an
d
elab
o
r
a
tes
o
n
th
e
im
p
lem
en
tatio
n
p
r
o
ce
d
u
r
e
f
o
r
v
o
ice
s
p
o
o
f
co
u
n
ter
m
ea
s
u
r
es
f
o
r
ASV
s
y
s
tem
s
.
Firs
t,
th
e
a
d
o
p
ted
d
ataset
is
b
r
ief
ly
d
escr
ib
ed
,
an
d
f
o
llo
win
g
th
is
,
th
e
ex
tr
ac
tio
n
o
f
t
h
e
ess
en
tial
f
ea
tu
r
es
th
at
co
n
tr
ib
u
te
to
war
d
s
lev
er
ag
in
g
m
o
r
e
wei
g
h
tag
e
in
th
e
laten
t
f
ea
tu
r
e
ex
tr
ac
tio
n
f
o
r
th
e
p
r
o
p
o
s
ed
h
y
b
r
id
s
p
ec
tr
al
-
s
p
atial
le
ar
n
in
g
m
o
d
el.
Fu
r
th
er
,
an
im
p
l
em
en
tatio
n
s
tr
ateg
y
was a
d
o
p
ted
f
o
r
f
ea
tu
r
e
m
atr
ix
p
r
ep
ar
atio
n
an
d
tr
ain
in
g
th
e
h
y
b
r
id
lear
n
in
g
m
o
d
el
is
d
is
cu
s
s
ed
.
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
.
1
4
,
No
.
1
,
Feb
r
u
ar
y
2
0
2
5
:
130
-
1
4
1
132
2
.
1
.
P
r
o
po
s
ed
s
y
s
t
em
T
h
e
r
esear
ch
wo
r
k
r
ep
o
r
ted
i
n
th
is
p
ap
er
is
p
r
im
ar
ily
f
o
cu
s
ed
o
n
in
tr
o
d
u
cin
g
a
u
n
iq
u
e
an
d
li
g
h
tweig
h
t
ap
p
r
o
ac
h
to
d
ev
elo
p
in
g
a
s
u
s
tain
ab
le
v
o
ice
s
p
o
o
f
co
u
n
ter
m
ea
s
u
r
e
s
y
s
tem
f
o
r
r
ea
l
-
wo
r
ld
ASV
s
y
s
tem
ap
p
licatio
n
s
.
T
h
e
p
r
o
p
o
s
ed
s
y
s
tem
is
d
esig
n
ed
to
d
y
n
am
ically
ad
ap
t
to
th
e
s
p
ea
k
er
'
s
v
o
ice
s
ig
n
al
d
ev
iatio
n
s
,
r
eg
ar
d
less
o
f
t
h
e
in
ter
-
class
s
im
ilar
ity
b
etwe
en
th
e
s
p
o
o
f
a
n
d
th
e
r
ea
l
au
d
io
s
ig
n
al.
I
n
t
h
is
ca
s
e,
th
e
s
tu
d
y
b
eliev
es th
at
i
f
an
attac
k
er
s
p
o
o
f
s
o
r
m
im
ics th
e
au
d
io
o
f
th
e
r
ea
l sp
ea
k
er
,
h
e
also
h
as so
m
e
laten
t f
ea
tu
r
e
th
at,
if
id
en
tifie
d
,
ca
n
h
elp
h
im
lear
n
th
e
p
atter
n
o
f
th
e
s
p
o
o
f
e
d
s
ig
n
al.
C
an
an
d
b
e
ab
le
t
o
cr
ea
t
e
a
lear
n
in
g
m
o
d
el
f
o
r
e
f
f
icien
tly
.
R
ed
u
ce
o
p
p
o
s
in
g
s
p
ee
ch
.
T
h
er
ef
o
r
e,
t
h
e
p
r
o
p
o
s
ed
s
tu
d
y
f
o
c
u
s
es
o
n
e
x
tr
ac
ti
n
g
m
u
ltip
le
f
ea
tu
r
es
f
r
o
m
th
e
au
d
io
f
iles
,
as
d
if
f
er
en
t
f
ea
tu
r
es
p
r
o
v
i
d
e
d
if
f
er
e
n
t
v
o
ice
attr
ib
u
tes,
o
f
ten
in
th
e
f
o
r
m
o
f
s
p
ec
tr
al
an
d
tem
p
o
r
al
r
ep
r
esen
tatio
n
s
.
T
h
e
p
r
o
p
o
s
ed
s
tu
d
y
im
p
lem
en
t
s
a
h
y
b
r
id
s
p
ec
tr
al
-
tem
p
o
r
al
lear
n
in
g
m
o
d
el
b
y
lev
er
ag
in
g
C
NN
an
d
l
o
n
g
s
h
o
r
t
-
ter
m
m
em
o
r
y
(
L
STM
)
d
ee
p
l
ea
r
n
in
g
m
o
d
els.
T
h
e
C
NN
m
o
d
el
is
u
s
ed
to
lear
n
an
d
ca
p
tu
r
e
th
e
s
p
ec
tr
al
d
ata,
wh
ile
L
STM
h
as
b
ee
n
im
p
le
m
en
ted
to
lear
n
tem
p
o
r
al
f
ea
t
u
r
es
an
d
ca
p
t
u
r
e
t
h
e
lo
n
g
-
ter
m
d
ep
e
n
d
en
cies.
T
h
e
p
r
o
p
o
s
ed
s
y
s
tem
d
esig
n
is
lig
h
tweig
h
t
b
ec
au
s
e
th
e
s
tu
d
y
p
er
f
o
r
m
s
d
im
en
s
io
n
ality
r
ed
u
ctio
n
o
p
e
r
atio
n
s
o
n
s
o
m
e
ex
tr
ac
ted
f
ea
tu
r
es
an
d
f
o
c
u
s
es
o
n
im
p
lem
e
n
tin
g
a
h
ig
h
l
y
o
p
tim
ized
lear
n
i
n
g
m
o
d
el
with
o
p
tim
al
lear
n
a
b
le
p
ar
am
eter
s
,
with
lo
n
g
er
tr
ain
i
n
g
ep
o
ch
s
an
d
o
p
tim
al
h
y
p
er
p
ar
am
eter
s
s
elec
tio
n
.
T
h
e
s
ch
em
atic
ar
ch
itectu
r
e
o
f
th
e
p
r
o
p
o
s
ed
s
y
s
tem
is
s
h
o
wn
in
Fig
u
r
e
1
.
Fig
u
r
e
1
.
A
p
r
o
p
o
s
ed
h
y
b
r
id
s
p
ec
tr
al
-
tem
p
o
r
al
m
o
d
el
f
o
r
s
p
o
o
f
co
u
n
ter
m
ea
s
u
r
e
Fig
u
r
e
1
d
ep
icts
th
e
b
lo
ck
-
b
a
s
ed
wo
r
k
f
lo
w
o
f
th
e
p
r
o
p
o
s
ed
h
y
b
r
id
s
p
ec
tr
al
-
tem
p
o
r
al
m
o
d
el,
wh
ich
co
m
p
r
is
es
f
o
u
r
c
o
r
e
m
o
d
u
less
u
ch
as
:
i)
e
x
p
lo
r
at
o
r
y
d
ata
an
aly
s
is
,
ii)
m
eta
-
d
ata
co
n
s
tr
u
ctio
n
,
iii)
f
ea
tu
r
e
ex
tr
ac
tio
n
,
iii)
f
ea
tu
r
e
co
n
ca
te
n
atio
n
,
a
n
d
iv
)
lear
n
in
g
s
p
ec
tr
al
an
d
tem
p
o
r
al
f
ea
tu
r
es
f
o
r
s
p
o
o
f
c
o
u
n
te
r
m
ea
s
u
r
e.
T
h
e
n
o
v
elty
o
f
th
e
p
r
o
p
o
s
ed
s
y
s
tem
is
th
at
it
u
s
es
a
lig
h
tweig
h
t
ap
p
r
o
a
ch
th
at
is
n
o
t
co
m
p
u
t
atio
n
ally
ex
p
e
n
s
iv
e.
T
h
is
f
ea
tu
r
e
m
ak
es
it
m
o
r
e
p
r
a
ctica
l
f
o
r
r
ea
l
-
wo
r
ld
a
p
p
licatio
n
s
.
T
h
e
s
y
s
tem
is
s
u
s
tain
ab
le
b
ec
au
s
e
it
ca
n
ad
ap
t
to
th
e
s
p
ea
k
er
'
s
v
o
ice
s
ig
n
al
ch
an
g
es
o
v
e
r
tim
e.
T
h
is
asp
ec
t
is
im
p
o
r
tan
t
b
ec
au
s
e
th
e
c
h
ar
ac
ter
is
tics
o
f
a
s
p
ea
k
er
'
s
v
o
ice
ca
n
c
h
an
g
e
d
u
e
to
f
ac
to
r
s
s
u
c
h
as
ag
e
,
illn
ess
,
o
r
en
v
ir
o
n
m
e
n
tal
co
n
d
iti
o
n
s
.
T
h
e
s
y
s
tem
is
p
r
ac
tical
f
o
r
r
ea
l
-
wo
r
ld
ASV
s
y
s
tem
ap
p
licatio
n
s
b
ec
a
u
s
e
it
c
an
b
e
im
p
lem
en
ted
o
n
v
ar
i
o
u
s
h
ar
d
war
e
p
latf
o
r
m
s
an
d
d
o
es n
o
t r
eq
u
ir
e
m
u
ch
tr
ai
n
in
g
d
ata.
2
.
2
.
Da
t
a
s
et
des
cr
iptio
n
T
h
e
d
ataset
co
n
s
id
er
ed
in
th
is
s
tu
d
y
is
ASV
Sp
o
o
f
2
0
1
9
[
2
8
]
,
a
lar
g
e
-
s
ca
le
p
u
b
lic
d
atab
ase
o
f
s
y
n
th
esized
,
co
n
v
er
ted
,
an
d
r
ep
lay
ed
s
p
ee
ch
.
I
t
was
cr
ea
t
ed
f
o
r
th
e
th
ir
d
au
t
o
m
atic
s
p
ea
k
er
v
er
if
icatio
n
s
p
o
o
f
in
g
a
n
d
c
o
u
n
ter
m
ea
s
u
r
es
ch
allen
g
e
(
ASVSp
o
o
f
2
0
1
9
)
,
wh
ich
was
h
eld
in
2
0
1
9
.
T
h
e
d
ataset
is
d
esig
n
ed
to
h
elp
r
esear
ch
er
s
d
ev
el
o
p
an
d
ev
alu
ate
tech
n
iq
u
es
f
o
r
d
ete
ctin
g
s
p
o
o
f
in
g
attac
k
s
ag
ain
s
t
ASV
s
y
s
tem
s
[
2
9
]
.
T
h
e
ASVs
p
o
o
f
2
0
1
9
d
ataset
co
n
s
is
ts
o
f
o
v
er
1
3
,
0
0
0
a
u
d
io
r
ec
o
r
d
in
g
s
,
d
iv
id
e
d
in
to
two
s
ce
n
a
r
io
s
s
u
ch
a
s
lo
g
ical
ac
ce
s
s
attac
k
s
an
d
p
h
y
s
ical
ac
ce
s
s
attac
k
s
.
T
h
e
p
r
o
p
o
s
ed
s
tu
d
y
co
n
s
id
er
s
a
lo
g
ical
ac
ce
s
s
attac
k
s
d
ataset
as
a
ca
s
e
s
tu
d
y
,
in
clu
d
in
g
g
en
u
in
e
an
d
s
p
o
o
f
ed
r
ec
o
r
d
in
g
s
.
T
h
e
g
en
u
in
e
r
ec
o
r
d
i
n
g
s
wer
e
cr
ea
te
d
b
y
r
ec
o
r
d
in
g
th
e
tar
g
et
s
p
ea
k
er
s
in
a
co
n
tr
o
lled
en
v
ir
o
n
m
en
t
with
o
v
er
2
0
,
0
0
0
a
u
d
io
r
ec
o
r
d
in
g
s
,
cr
ea
ted
u
s
in
g
v
ar
io
u
s
tech
n
iq
u
es,
s
u
ch
as v
o
ice
co
n
v
er
s
io
n
an
d
s
p
ee
ch
s
y
n
th
esis
.
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
Dis
crim
in
a
tive
d
ee
p
lea
r
n
in
g
b
a
s
ed
h
yb
r
id
s
p
ec
tr
o
-
temp
o
r
a
l fe
a
tu
r
es fo
r
…
(
P
r
a
n
ita
N
ir
a
j P
a
ls
a
p
u
r
e
)
133
2
.
3
.
M
et
a
-
da
t
a
c
o
ns
t
ruct
io
n
T
h
is
s
ec
tio
n
d
is
cu
s
s
es
th
e
p
r
o
ce
s
s
o
f
m
etad
ata
p
r
e
p
ar
ati
o
n
,
wh
ich
i
n
v
o
lv
es
ex
tr
ac
tin
g
ess
en
tial
in
f
o
r
m
atio
n
f
r
o
m
p
r
o
to
c
o
l
f
il
es
an
d
o
r
g
an
izin
g
it
in
t
o
C
SV
f
iles
f
o
r
e
f
f
icien
t
d
ata
m
an
ag
em
en
t
an
d
m
o
d
el
tr
ain
in
g
.
Alg
o
r
ith
m
1
is
d
esig
n
ed
th
at
in
itiates
with
th
r
ee
in
p
u
ts
:
p
r
o
to
co
l
f
ile
(
P),
r
ep
r
esen
tin
g
th
e
f
ile
co
n
tain
in
g
th
e
r
eq
u
ir
ed
p
r
o
to
c
o
l
d
ata;
au
d
io
d
ir
ec
to
r
y
(
A)
,
t
h
e
d
ir
ec
to
r
y
co
n
tain
i
n
g
au
d
io
d
ata;
an
d
C
SV
f
ile
(
C
)
,
s
ig
n
if
y
in
g
t
h
e
f
ile
wh
er
e
th
e
ex
tr
ac
ted
m
eta
d
ata
is
to
b
e
s
to
r
ed
.
T
h
e
al
g
o
r
ith
m
'
s
o
u
tp
u
t
is
th
e
s
tr
u
ctu
r
ed
lis
t,
m
etad
ata
(
M)
,
co
n
tain
in
g
th
e
p
r
o
ce
s
s
ed
in
f
o
r
m
atio
n
ex
tr
ac
ted
f
r
o
m
th
e
p
r
o
to
co
l
f
ile.
Alg
o
r
ith
m
1
.
Me
tad
ata
c
o
n
s
tr
u
ctio
n
f
o
r
ASV
I
n
p
u
t:
P
(
p
r
o
t
o
co
l f
ile)
,
A
(
au
d
i
o
d
ir
ec
to
r
y
)
,
C
(
C
SV f
ile)
Ou
tp
u
t:
M
(
m
etad
ata)
Star
t
1.
I
n
itialize
M
= {}
2.
Op
en
p
r
o
to
co
l f
ile
P
f
o
r
r
ea
d
i
n
g
.
3.
Fo
r
ea
ch
lin
e
p
i
in
P
:
4.
L
et
Par
ts
b
e
th
e
r
esu
lt o
f
s
p
litt
in
g
th
e
lin
e
b
y
wh
ites
p
ac
e
5.
Par
ts
=
s
p
lit(
p
i,
‘
’
)
6.
//P
ar
ts
ϵ
[
S
_
I
d
(
s
p
ea
k
e
r
id
en
tity
)
,
F
_
I
d
(
f
ile
id
en
tity
)
,
…
La
b
el
]
7.
E
x
tr
ac
t th
e
f
o
llo
win
g
in
f
o
r
m
atio
n
f
r
o
m
Par
ts
:
8.
Fo
r
ea
ch
p
i
,
e
x
tr
ac
t
S_
I
D
→
Par
ts
[
0
]
F_
I
D
→
Par
ts
[
1
]
L
ab
el
→
Par
ts
[
4
]
9.
C
o
n
s
tr
u
ct
Au
d
io
Path
Ap
as:
Ap
=
A
+
”f
lac/”
+
F
_
I
D
10.
C
r
ea
te
a
m
etad
ata
r
ec
o
r
d
with
:
11.
Fo
r
ea
ch
p
i,
let
r
i
b
e
a
r
ec
o
r
d
:
r
i
=
{S
_
I
d
,
F
_
I
d
,
La
b
el,
Ap
}
12.
Ap
p
en
d
R
ec
o
r
d
to
Me
tad
ata
M
13.
M =
M
∪
{
r
i
}
14.
C
lo
s
e
p
r
o
to
co
l f
ile
15.
Op
en
csv
f
ile
C
f
o
r
wr
itin
g
16.
W
r
ite
a
h
ea
d
er
r
o
w:
S
_
I
D
,
F
_
I
D,
Ap
17.
Fo
r
ea
ch
r
i
in
M
:
18.
W
r
ite
a
r
o
w
with
v
alu
es
19.
wr
iteR
o
w(
C
,
r
i
[
S
_
I
d
],
r
i
[
F
_
I
d
],
r
i
[
La
b
el
],
r
i
[
Ap
]
)
20.
C
lo
s
e
C
21.
R
etu
r
n
M
E
n
d
An
em
p
ty
lis
t,
M,
is
in
itially
in
itialized
to
s
to
r
e
th
e
ex
tr
a
cted
m
etad
ata.
T
h
e
alg
o
r
ith
m
b
eg
in
s
b
y
o
p
en
in
g
th
e
p
r
o
to
c
o
l f
ile
P
f
o
r
r
ea
d
in
g
.
T
h
e
alg
o
r
ith
m
p
e
r
f
o
r
m
s
a
s
er
ies o
f
ex
tr
ac
tio
n
an
d
p
r
o
ce
s
s
in
g
s
tep
s
f
o
r
ea
ch
lin
e
p
i
p
r
esen
t
in
th
e
p
r
o
to
co
l
f
ile
P.
T
h
e
lin
e
is
s
p
lit
b
y
wh
ites
p
ac
e
in
t
o
co
n
s
titu
en
t
p
ar
ts
ca
lled
Par
ts
.
Sev
er
al
p
iece
s
o
f
in
f
o
r
m
atio
n
ar
e
ex
tr
ac
ted
f
r
o
m
th
ese
p
a
r
ts
,
s
u
ch
as
s
p
ea
k
er
id
en
tity
(
S_
I
D)
d
er
iv
ed
f
r
o
m
th
e
f
ir
s
t
elem
en
t,
f
ile
id
e
n
tity
(
F
_
I
D
)
f
r
o
m
t
h
e
s
ec
o
n
d
elem
e
n
t,
an
d
lab
el
f
r
o
m
th
e
f
if
th
e
lem
en
t
o
f
th
e
p
ar
ts
.
Fo
llo
win
g
th
e
ex
tr
ac
tio
n
,
th
e
alg
o
r
ith
m
co
n
s
tr
u
cts
th
e
Au
d
i
o
Path
.
I
t
f
o
r
m
s
th
is
b
y
co
n
ca
t
en
atin
g
th
e
A
with
a
s
tr
in
g
"f
lac/"
a
n
d
th
e
F
_
I
D
.
Su
b
s
eq
u
en
tly
,
a
m
etad
ata
r
ec
o
r
d
,
d
en
o
ted
as
r
i
,
is
cr
ea
te
d
with
t
h
e
ex
tr
ac
te
d
S
_
I
D
,
F
_
I
D
,
lab
el,
an
d
th
e
c
o
n
s
tr
u
c
ted
Au
d
io
Path
,
d
en
o
ted
as
Ap
.
Fu
r
th
e
r
,
ea
ch
r
ec
o
r
d
is
th
e
n
ap
p
e
n
d
ed
to
t
h
e
m
etad
ata
lis
t,
wh
ich
ac
cu
m
u
lates
all
th
e
m
eta
d
ata
r
ec
o
r
d
s
p
r
o
ce
s
s
ed
f
r
o
m
th
e
P
.
Af
ter
p
r
o
ce
s
s
in
g
all
lin
es
in
th
e
P
,
th
e
f
ile
is
clo
s
ed
,
an
d
th
e
C
SV
f
il
e
is
o
p
en
ed
f
o
r
wr
itin
g
.
A
h
ea
d
er
r
o
w
is
wr
itten
f
ir
s
t
in
to
th
e
csv
f
ile,
co
n
tain
in
g
th
e
co
lu
m
n
s
:
'
S_
I
D
'
,
'
F_
I
D
'
,
Ap
.
Su
b
s
eq
u
en
tly
,
f
o
r
ea
ch
r
ec
o
r
d
r
i
p
r
esen
t
in
t
h
e
m
etad
ata
M
,
a
n
ew
r
o
w
is
wr
itten
to
th
e
C
SV
f
ile,
with
v
alu
es
co
r
r
esp
o
n
d
in
g
to
th
e
elem
en
ts
o
f
th
e
r
ec
o
r
d
(
lin
e
1
9
)
an
d
u
p
o
n
wr
itin
g
all
th
e
m
etad
ata
r
ec
o
r
d
s
to
th
e
C
SV f
ile,
it is
f
in
ally
clo
s
ed
.
Fig
u
r
e
s
2
an
d
3
o
f
f
er
co
m
p
r
eh
en
s
iv
e
v
is
u
al
an
aly
s
es
o
f
b
o
n
a
f
id
e
an
d
s
p
o
o
f
v
o
ice
s
ig
n
als,
r
es
p
ec
tiv
ely
.
Fig
u
r
e
2
d
is
p
lay
s
th
e
wa
v
ef
o
r
m
an
d
s
p
ec
tr
o
g
r
am
o
f
a
g
en
u
i
n
e
o
r
b
o
n
af
id
e
v
o
ice
s
ig
n
al,
ill
u
m
in
atin
g
its
u
n
iq
u
e
tem
p
o
r
al
a
n
d
s
p
ec
tr
al
ch
ar
ac
t
er
is
tics
,
wh
ich
ar
e
cr
u
cial
f
o
r
ASV
s
y
s
tem
s
.
I
n
co
n
tr
ast,
Fi
g
u
r
e
3
o
u
tlin
es
t
h
e
ch
ar
ac
ter
is
tics
o
f
a
s
p
o
o
f
o
r
a
r
tific
ially
g
en
er
ated
v
o
ice
s
ig
n
al,
em
p
h
asizin
g
t
h
e
an
o
m
alies
an
d
d
is
cr
ep
an
cies
in
its
s
p
ec
tr
al
r
ep
r
esen
tatio
n
c
o
m
p
ar
ed
to
b
o
n
af
id
e
s
ig
n
als.
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
.
1
4
,
No
.
1
,
Feb
r
u
ar
y
2
0
2
5
:
130
-
1
4
1
134
Fig
u
r
e
2
.
I
ll
u
s
tr
ativ
e
d
ep
ictio
n
o
f
b
o
n
af
id
e
s
p
ea
k
er
s
ig
n
al
c
h
ar
ac
ter
is
tics
Fig
u
r
e
3
.
A
g
r
ap
h
ic
r
ep
r
esen
ta
tio
n
o
f
d
is
tin
ctiv
e
s
p
o
o
f
v
o
ice
s
ig
n
al
attr
ib
u
tes
2
.
4
.
F
e
a
t
ure
ex
t
r
a
ct
io
n
T
h
e
f
ea
tu
r
e
e
x
tr
ac
tio
n
p
r
o
ce
s
s
is
cr
u
cial
in
th
e
ar
ch
itectu
r
e
o
f
s
p
ee
ch
-
s
p
o
o
f
in
g
c
o
u
n
ter
m
ea
s
u
r
es
s
y
s
tem
s
.
As
a
tr
an
s
f
o
r
m
atio
n
s
tep
,
it
co
n
v
er
ts
th
e
o
r
ig
in
al
s
p
ea
k
er
au
d
io
d
ata
in
to
a
s
tr
u
ct
u
r
ed
f
o
r
m
,
wh
ich
is
ad
v
an
tag
e
o
u
s
to
th
e
d
etailed
an
aly
s
is
an
d
ap
p
licatio
n
o
f
m
ac
h
in
e
lear
n
in
g
alg
o
r
ith
m
s
.
G
iv
en
th
e
in
h
er
en
tly
d
is
cr
im
in
ativ
e
n
atu
r
e
o
f
au
d
i
o
s
ig
n
als,
ca
p
tu
r
in
g
th
e
ess
en
ti
al
f
ea
tu
r
es
th
at
f
ac
ilit
ate
ac
cu
r
ate
d
is
cr
im
in
atio
n
b
etwe
en
r
ea
l
(
b
o
n
a
f
id
e)
an
d
s
p
o
o
f
ed
v
o
ices
is
im
p
o
r
tan
t.
T
h
er
ef
o
r
e,
th
e
s
tu
d
y
f
o
c
u
s
es
o
n
ex
tr
ac
tin
g
th
r
ee
ess
en
tial
f
ea
tu
r
es,
s
u
ch
as
MFC
C
s
,
s
p
ec
tr
al
co
n
tr
ast,
an
d
t
o
n
n
etzp
itch
f
ea
tu
r
es
th
at
ar
e
cr
u
c
ial
f
o
r
d
is
tin
g
u
is
h
in
g
b
o
n
af
id
e
an
d
s
y
n
th
etic
s
p
o
o
f
e
d
s
p
ee
ch
s
ig
n
als.
Ho
wev
e
r
,
th
e
v
ar
iab
ilit
y
in
a
u
d
io
s
ig
n
al
d
u
r
atio
n
o
f
ten
led
to
in
co
n
s
is
ten
t
f
ea
tu
r
e
len
g
t
h
s
.
T
o
co
r
r
ec
t
t
h
is
in
co
n
s
is
ten
cy
,
th
e
in
tr
o
d
u
ce
d
m
eth
o
d
in
co
r
p
o
r
ates
p
ad
d
in
g
,
a
tech
n
iq
u
e
th
at
alig
n
s
f
ea
tu
r
e
l
en
g
t
h
s
b
y
e
x
ten
d
in
g
o
r
tr
u
n
c
atin
g
th
em
to
a
p
r
ed
ete
r
m
in
e
d
len
g
th
.
A
s
am
p
le
v
is
u
aliza
tio
n
an
d
d
escr
ip
ti
o
n
o
f
th
e
ex
tr
ac
ted
f
ea
tu
r
es a
r
e
d
e
p
icted
in
Fig
u
r
e
s
4
an
d
5
.
Fig
u
r
e
4
d
is
p
lay
s
th
e
two
m
o
s
t
p
o
p
u
lar
a
u
d
io
f
ea
tu
r
es,
wit
h
s
u
b
f
ig
u
r
e
Fig
u
r
e
4
(
a
)
r
e
p
r
esen
tin
g
th
e
s
h
o
r
t
-
ter
m
p
o
wer
s
p
ec
tr
u
m
o
f
an
a
u
d
io
s
ig
n
al
th
at
ca
p
tu
r
es
th
e
s
p
ec
tr
al
ch
a
r
ac
ter
is
tics
o
f
th
e
s
ig
n
al
in
a
way
th
at
ap
p
r
o
x
im
ates
th
e
h
u
m
an
a
u
d
ito
r
y
s
y
s
tem
'
s
r
esp
o
n
s
e
to
s
o
u
n
d
.
T
h
is
r
ep
r
esen
tatio
n
s
h
o
wca
s
es
th
e
v
ar
iatio
n
o
f
MFC
C
s
o
v
er
tim
e,
p
r
o
v
id
i
n
g
in
s
ig
h
ts
in
to
t
h
e
au
d
io
s
ig
n
al
'
s
s
p
ec
tr
al
co
n
ten
t a
n
d
ac
o
u
s
tic
p
r
o
p
er
ties
,
wh
ich
ar
e
v
alu
ab
le
f
o
r
id
e
n
tify
in
g
t
r
a
n
s
ien
t
ev
en
ts
,
s
p
ee
ch
p
atter
n
s
,
an
d
ch
a
n
g
es
in
f
r
e
q
u
en
c
y
co
m
p
o
n
e
n
ts
with
in
th
e
au
d
io
s
ig
n
al.
Fig
u
r
e
4
(
b
)
d
em
o
n
s
tr
ates
th
e
s
p
ec
tr
al
co
n
tr
ast
o
f
th
e
au
d
io
s
ig
n
al,
a
r
ep
r
esen
tatio
n
th
at
q
u
a
n
tifie
s
th
e
d
if
f
e
r
en
ce
i
n
am
p
litu
d
e
b
e
twee
n
s
p
ec
tr
al
p
ea
k
s
an
d
v
alle
y
s
in
an
au
d
io
s
p
ec
tr
u
m
.
T
h
is
ca
n
b
e
p
ar
ticu
lar
l
y
us
ef
u
l
f
o
r
d
is
tin
g
u
is
h
in
g
b
etwe
en
d
if
f
er
e
n
t
ty
p
es
o
f
au
d
i
o
s
,
as
it
em
p
h
asizes
v
ar
iatio
n
s
in
s
p
ec
tr
al
s
tr
u
ctu
r
e,
s
u
ch
as th
e
p
r
esen
ce
o
f
h
a
r
m
o
n
ics o
r
f
o
r
m
an
ts
.
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
Dis
crim
in
a
tive
d
ee
p
lea
r
n
in
g
b
a
s
ed
h
yb
r
id
s
p
ec
tr
o
-
temp
o
r
a
l fe
a
tu
r
es fo
r
…
(
P
r
a
n
ita
N
ir
a
j P
a
ls
a
p
u
r
e
)
135
(
a)
(
b
)
Fig
u
r
e
4
.
A
n
an
aly
tical
v
is
u
ali
za
tio
n
o
f
(
a)
MFC
C
f
ea
tu
r
e
attr
ib
u
tes
an
d
(
b
)
s
p
ec
tr
o
g
r
am
r
e
p
r
esen
tatio
n
f
ea
tu
r
e
attr
ib
u
tes
Fig
u
r
e
5
d
is
p
lay
s
th
e
two
cr
iti
ca
l
au
d
ito
r
y
f
ea
tu
r
es;
F
ig
u
r
e
5
(
a)
d
is
p
lay
s
a
s
p
ec
tr
o
g
r
am
v
is
u
aliza
tio
n
f
o
r
au
d
i
o
an
aly
s
is
,
a
2
D
r
ep
r
es
en
tatio
n
o
f
an
a
u
d
io
s
ig
n
al
wh
er
e
tim
e
is
p
lo
tted
o
n
th
e
x
-
a
x
i
s
,
f
r
eq
u
en
c
y
o
n
th
e
y
-
ax
is
,
an
d
th
e
co
l
o
u
r
i
n
ten
s
ity
r
ep
r
esen
ts
th
e
m
ag
n
itu
d
e
o
r
p
o
wer
o
f
d
if
f
e
r
en
t
f
r
eq
u
e
n
cy
co
m
p
o
n
en
ts
at
ea
ch
p
o
in
t
in
tim
e.
T
h
is
v
is
u
al
r
ep
r
esen
tatio
n
en
ab
les
u
s
to
o
b
s
er
v
e
h
o
w
th
e
s
p
ec
tr
al
co
n
ten
t
o
f
th
e
au
d
io
ev
o
lv
es
o
v
er
tim
e.
I
n
n
ex
t
Fig
u
r
e
5
(
b
)
,
th
e
to
n
n
etz
f
ea
tu
r
e
o
f
a
n
au
d
io
s
ig
n
al
is
d
ep
i
cted
.
T
h
is
f
ea
tu
r
e
ca
p
tu
r
es
th
e
to
n
al
ch
ar
ac
ter
is
tics
o
f
an
au
d
io
s
ig
n
al,
r
ep
r
esen
te
d
as
a
s
et
o
f
v
alu
es
o
r
co
ef
f
icien
ts
th
at
d
escr
ib
e
th
e
en
er
g
y
d
is
tr
ib
u
tio
n
in
d
if
f
e
r
en
t
to
n
al
r
eg
io
n
s
.
Mo
r
e
o
v
er
,
to
n
n
etz
f
ea
t
u
r
es
ar
e
s
en
s
itiv
e
to
p
itch
v
ar
ia
tio
n
s
an
d
h
ar
m
o
n
ic
co
n
ten
t,
m
a
k
in
g
th
e
m
v
alu
a
b
l
e
f
o
r
id
e
n
tify
in
g
to
n
al
p
atter
n
s
an
d
d
if
f
e
r
en
ce
s
b
etwe
en
g
e
n
u
in
e
an
d
s
p
o
o
f
ed
v
o
ices
in
ASV
s
y
s
tem
s
.
T
h
e
ex
tr
ac
ted
f
ea
tu
r
esar
e
em
p
ir
ically
r
ev
iewe
d
,
an
d
it
is
id
en
tifie
d
th
at
MFC
C
f
ea
tu
r
es,
s
p
ec
tr
al
co
n
tr
ast,
an
d
T
o
n
n
etz
f
ea
tu
r
es
ar
e
c
r
itical
f
o
r
d
is
tin
g
u
is
h
in
g
th
e
d
if
f
e
r
en
ce
b
etwe
en
s
p
o
o
f
e
d
an
d
r
ea
l
v
o
ices,
p
r
o
v
id
in
g
u
n
iq
u
e
in
s
ig
h
ts
in
to
t
h
e
in
h
er
en
t
c
h
ar
ac
t
er
is
tics
o
f
a
s
p
ea
k
er
'
s
v
o
ice.
I
n
co
n
tr
ast,
MFC
C
f
ea
tu
r
es,
an
d
Me
l sp
ec
tr
o
g
r
am
s
ex
h
ib
it si
m
ilar
f
ea
tu
r
es d
er
iv
ed
f
r
o
m
t
h
e
in
p
u
t a
u
d
i
o
d
ata
b
ec
au
s
e
MFC
C
s
ar
e
d
er
iv
ed
f
r
o
m
Me
l sp
ec
tr
o
g
r
am
s
.
T
h
er
ef
o
r
e,
b
ased
o
n
e
m
p
ir
ical
r
esear
ch
,
th
is
s
tu
d
y
m
ain
ly
f
o
cu
s
es
o
n
MFC
C
,
s
p
ec
tr
al
co
n
tr
ast
an
d
T
o
n
n
etz
ch
a
r
ac
ter
is
tics
.
Prin
cip
al
co
m
p
o
n
en
t
an
aly
s
is
(
PC
A)
wa
s
d
ep
lo
y
ed
o
n
s
p
ec
tr
al
co
n
tr
ast
an
d
T
o
n
n
etzf
ea
tu
r
es
to
o
p
tim
ize
an
aly
s
is
ef
f
icien
cy
,
aim
in
g
to
r
e
d
u
ce
d
im
en
s
io
n
ality
wh
ile
r
e
tain
in
g
9
9
%
o
f
th
e
ex
p
lain
ed
v
ar
iatio
n
f
ac
to
r
s
.
T
h
e
r
ea
s
o
n
wh
y
MFC
C
is
e
x
em
p
ted
f
r
o
m
PC
A
is
b
ased
o
n
th
e
em
p
ir
ical
o
b
s
er
v
atio
n
th
at
its
co
n
tr
ib
u
tio
n
to
y
ield
im
p
r
o
v
e
m
en
t
is
m
in
im
al.
T
h
e
en
tire
co
m
p
u
tin
g
p
r
o
ce
s
s
in
v
o
lv
ed
i
n
ex
tr
a
ctin
g
an
d
s
elec
tin
g
au
d
i
o
f
ea
tu
r
es is
d
is
cu
s
s
ed
in
Alg
o
r
ith
m
2
.
Alg
o
r
ith
m
2
.
Featu
r
e
ex
tr
ac
tio
n
an
d
s
elec
tio
n
I
n
p
u
t:
A
(
Set o
f
r
aw
au
d
io
d
ata
)
;
L
max
(
Ma
x
im
u
m
allo
wab
le
f
e
atu
r
e
len
g
th
)
Ou
tp
u
ts
:
F
(
T
r
an
s
f
o
r
m
ed
f
ea
t
u
r
e
m
atr
ix
)
;
S
(
Selecte
d
f
ea
tu
r
e
s
et)
Star
t
1.
E
x
tr
ac
t in
itial f
ea
tu
r
e
s
et
F
={
MFC
C
S,
Sp
ec
tr
o
g
r
am
,
Sp
ec
tr
a
l Co
n
tr
ast,
T
o
n
n
etz}
f
r
o
m
A
2.
L
et
F
’
=
∅
: I
n
itialize
th
e
tr
an
s
f
o
r
m
ed
f
ea
tu
r
e
m
atr
ix
3.
Dim
en
s
io
n
al
C
o
n
s
is
ten
cy
:
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
.
1
4
,
No
.
1
,
Feb
r
u
ar
y
2
0
2
5
:
130
-
1
4
1
136
4.
Fo
r
ea
ch
f
ea
tu
r
e
F
i
in
F
:
5.
C
o
m
p
u
te
n
u
m
_
tim
e_
f
r
a
m
es =
len
g
th
(
F
i
)
6.
I
f
n
u
m
_
tim
e_
f
r
am
es>
L
max
:
7.
T
r
u
n
ca
te:
F
i
’
=
F
i
[
:,
:
L
max
]
8.
E
ls
e
I
f
n
u
m
_
tim
e
_
f
r
am
es<
L
max
:
9.
E
x
ten
d
:
F
i
'=
Pad
(
F
i
,
(
0
,
0
)
,
(
0
,
L
ma
x
-
n
u
m
_
tim
es_
f
r
am
es))
,
m
o
d
e=
‘
co
n
s
tan
t’
10.
E
ls
e:
11.
F
i
=
F
i
12.
Up
d
ate
F'
with
F
i
'
13.
Per
f
o
r
m
an
al
y
s
is
o
n
F'
to
d
er
i
v
e
d
is
cr
im
in
ativ
e
ch
a
r
ac
ter
is
tics
an
d
s
im
ilar
ities
.
14.
L
et
S
=
{M
FC
C
S,
Sp
ec
tr
al
C
o
n
tr
ast,
T
o
n
n
etz}
b
e
th
e
s
et
o
f
s
elec
ted
f
ea
tu
r
es
15.
Ap
p
ly
PC
A
o
n
f
ea
tu
r
e
in
S
e
x
c
lu
d
in
g
MFC
C
to
r
etain
9
9
% o
f
th
e
ex
p
lain
e
d
v
a
r
ian
ce
f
ac
to
r
.
16.
Up
d
ate
th
e
s
et
S
with
th
e
f
ea
tu
r
es a
f
ter
d
im
en
s
io
n
ality
r
ed
u
ct
io
n
.
17.
R
etu
r
n
th
e
tr
an
s
f
o
r
m
ed
f
ea
t
u
r
e
m
atr
ix
F'
a
n
d
th
e
s
elec
ted
f
e
atu
r
es set
S
.
18.
C
o
n
s
tr
u
ct
a
f
ea
tu
r
e
m
atr
i
x
b
y
co
n
ca
ten
atin
g
ea
c
h
f
ea
tu
r
e
in
Set
S
E
n
d
(
a)
(
b
)
Fig
u
r
e
5
.
D
etailed
v
is
u
al
e
x
a
m
in
atio
n
o
f
(
a)
s
p
ec
tr
al
c
o
n
tr
a
s
t f
ea
tu
r
es
an
d
(
b
)
to
n
n
etz
f
ea
t
u
r
e
2
.
6
.
H
y
brid lea
rning
m
o
del
T
h
e
p
r
o
p
o
s
ed
h
y
b
r
id
lear
n
in
g
m
o
d
el
p
r
esen
ts
a
co
m
p
r
eh
en
s
iv
e
m
o
d
el
b
y
lev
er
a
g
in
g
t
h
e
d
is
tin
ct
s
tr
en
g
th
s
o
f
b
o
th
ar
ch
itectu
r
e
s
C
N
N
an
d
L
STM
n
etwo
r
k
s
to
ad
ep
tly
ca
p
tu
r
e
an
d
in
ter
p
r
et
th
e
s
p
ec
tr
al
an
d
tem
p
o
r
al
i
n
tr
icac
ies
with
in
au
d
io
f
ea
t
u
r
es.
T
h
e
p
r
o
p
o
s
ed
h
y
b
r
id
lear
n
in
g
ar
ch
itectu
r
e
c
o
n
s
is
ts
o
f
an
in
p
u
t
lay
e
r
d
esig
n
ed
ef
f
ec
tiv
ely
to
h
an
d
le
f
ea
tu
r
e
v
ec
to
r
s
.
T
h
is
lay
er
ac
ce
p
ts
d
im
en
s
io
n
s
o
f
s
h
ap
e
(
2
6
0
0
,
1
)
,
wh
e
r
e
2
6
0
0
s
y
m
b
o
lizes
th
e
c
u
m
u
lativ
e
len
g
th
o
f
t
h
e
MFC
C
s
,
s
p
ec
tr
al
co
n
tr
ast,
an
d
to
n
n
etz
f
ea
tu
r
es.
T
h
e
s
in
g
u
lar
ch
a
n
n
el
d
en
o
tes
its
u
n
id
im
en
s
io
n
al
n
atu
r
e,
u
n
d
er
s
co
r
in
g
th
e
i
m
p
o
r
tan
ce
o
f
ea
ch
f
ea
tu
r
e'
s
s
eq
u
en
tial
alig
n
m
e
n
t.
Su
cc
ee
d
in
g
th
e
in
p
u
t
is
a
co
n
v
o
lu
tio
n
al
f
r
am
ewo
r
k
co
n
s
is
tin
g
o
f
a
s
eq
u
e
n
ce
o
f
o
n
e
-
d
im
e
n
s
io
n
al
co
n
v
o
lu
tio
n
al
lay
er
s
.
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
Dis
crim
in
a
tive
d
ee
p
lea
r
n
in
g
b
a
s
ed
h
yb
r
id
s
p
ec
tr
o
-
temp
o
r
a
l fe
a
tu
r
es fo
r
…
(
P
r
a
n
ita
N
ir
a
j P
a
ls
a
p
u
r
e
)
137
W
ith
th
eir
weig
h
t
-
s
h
ar
in
g
m
ec
h
an
is
m
,
th
ese
lay
er
s
attem
p
t
to
d
etec
t
an
d
ex
t
r
ap
o
late
s
p
atial
r
elatio
n
s
h
ip
s
with
in
th
e
au
d
io
d
ata.
I
n
o
r
d
er
to
en
s
u
r
e
co
m
p
u
tatio
n
al
ef
f
icien
c
y
with
o
u
t
co
m
p
r
o
m
is
in
g
o
n
ess
en
tial
f
ea
tu
r
e
in
f
o
r
m
atio
n
,
ea
ch
c
o
n
v
o
lu
tio
n
al
lay
er
is
f
o
llo
wed
b
y
a
m
ax
-
p
o
o
lin
g
lay
er
,
ef
f
ec
tiv
el
y
d
o
wn
-
s
am
p
lin
g
th
e
f
ea
t
u
r
e
d
i
m
en
s
io
n
s
.
B
u
ild
in
g
o
n
th
e
s
p
a
tial
p
atter
n
s
ex
tr
ac
ted
b
y
th
e
c
o
n
v
o
lu
ti
o
n
al
lay
er
s
,
th
e
ar
ch
itectu
r
e
in
teg
r
ates
an
L
STM
lay
er
to
u
n
d
er
s
tan
d
s
eq
u
en
ce
s
.
T
h
e
L
STM
lay
er
o
f
f
er
s
th
e
m
o
d
el
a
s
o
p
h
is
ticated
m
em
o
r
y
m
ec
h
an
is
m
,
em
p
o
wer
in
g
it
to
lea
r
n
a
n
d
r
em
em
b
er
tem
p
o
r
al
p
atter
n
s
em
b
ed
d
e
d
with
i
n
th
e
au
d
io
f
ea
tu
r
es.
Po
s
t
th
e
L
STM
in
teg
r
atio
n
,
th
e
m
o
d
elli
n
g
p
h
ase
also
ad
d
s
two
f
u
lly
co
n
n
ec
ted
o
r
d
en
s
e
lay
er
s
.
T
h
ese
lay
er
s
,
lad
en
with
n
eu
r
o
n
s
ca
p
a
b
l
e
o
f
co
m
p
l
ex
n
o
n
-
lin
ea
r
tr
an
s
f
o
r
m
atio
n
s
,
f
u
r
t
h
er
r
e
f
in
e
t
h
e
f
ea
tu
r
es,
p
r
e
p
ar
in
g
th
em
f
o
r
th
e
f
in
al
class
if
icatio
n
task
.
T
h
e
s
tu
d
y
also
e
n
s
u
r
es
th
at
th
e
p
r
o
p
o
s
ed
lear
n
in
g
m
o
d
el
is
r
o
b
u
s
t
ag
ain
s
t
o
v
er
f
itti
n
g
is
s
u
es
in
th
e
tr
ain
in
g
p
r
o
ce
s
s
b
y
in
co
r
p
o
r
atin
g
L
2
r
e
g
u
lar
i
za
tio
n
an
d
d
r
o
p
o
u
t
tech
n
iq
u
es.
T
h
e
f
in
al
lay
er
o
f
th
e
m
o
d
el
is
th
e
o
u
tp
u
t
lay
e
r
,
wh
ich
ac
ts
as
th
e
f
in
al
d
ec
is
io
n
-
m
a
k
in
g
m
o
d
u
le
o
f
th
e
p
r
o
p
o
s
ed
s
y
s
tem
.
I
t
co
n
s
is
ts
o
f
a
s
in
g
le
n
eu
r
o
n
with
a
s
ig
m
o
id
ac
tiv
atio
n
f
u
n
ctio
n
,
an
d
it
o
u
tp
u
ts
th
e
p
r
o
b
ab
ilit
y
o
f
th
e
in
p
u
t
au
d
io
b
ein
g
g
en
u
i
n
e
o
r
s
p
o
o
f
ed
.
T
o
tr
ain
th
e
h
y
b
r
id
m
o
d
el
ef
f
ec
tiv
ely
,
th
e
s
tu
d
y
u
s
ed
th
e
Ad
am
o
p
tim
izer
,
k
n
o
wn
f
o
r
its
ad
ap
t
ab
ilit
y
,
with
a
lear
n
in
g
r
ate
in
itialized
at
0
.
0
0
1
.
T
h
e
s
tu
d
y
a
ls
o
u
s
ed
th
e
b
in
ar
y
cr
o
s
s
-
en
tr
o
p
y
lo
s
s
f
u
n
ctio
n
to
q
u
an
tify
th
e
m
o
d
el'
s
er
r
o
r
s
an
d
g
u
id
e
its
lear
n
in
g
,
as
it
is
we
ll
-
s
u
ited
f
o
r
b
in
ar
y
class
if
icatio
n
task
s
.
T
h
e
s
tu
d
y
u
s
ed
two
ca
llb
ac
k
s
to
en
s
u
r
e
c
o
n
s
is
ten
t a
n
d
ef
f
icien
t
tr
ain
in
g
,
i.e
.
,
T
en
s
o
r
B
o
ar
d
f
o
r
p
er
f
o
r
m
an
ce
m
o
n
ito
r
in
g
a
n
d
R
ed
u
ce
L
R
On
Plateau
f
o
r
d
y
n
am
ic
lear
n
in
g
r
ate
ad
ju
s
tm
e
n
ts
.
T
h
e
m
o
d
el
w
as
tr
ain
ed
o
v
er
1
0
0
ep
o
ch
s
,
p
r
o
c
ess
in
g
b
atch
es o
f
1
0
0
s
am
p
les ea
ch
.
Fig
u
r
e
6
p
r
esen
ts
th
e
en
tire
f
lo
w
c
h
ar
t w
o
r
k
f
o
r
d
etec
tin
g
b
o
n
a
f
id
e
a
n
d
v
o
i
ce
s
p
o
o
f
i
n
g
attac
k
s
u
s
in
g
t
h
e
p
r
o
p
o
s
ed
h
y
b
r
id
s
p
ec
tr
al
-
tem
p
o
r
al
lear
n
in
g
m
o
d
el.
Fig
u
r
e
6
.
F
lo
wch
ar
t
o
f
th
e
p
r
o
p
o
s
ed
s
y
s
tem
f
o
r
b
o
n
a
f
id
e
an
d
s
p
o
o
f
s
ig
n
al
d
etec
tio
n
3.
RE
SU
L
T
AND
DI
SCUS
SI
O
N
T
h
e
d
esig
n
an
d
d
ev
el
o
p
m
en
t
o
f
th
e
p
r
o
p
o
s
ed
s
y
s
tem
f
o
r
v
o
ic
e
s
p
o
o
f
in
g
attac
k
d
etec
tio
n
is
d
o
n
e
u
s
in
g
Py
th
o
n
ex
ec
u
ted
in
An
ac
o
n
d
a
d
is
tr
ib
u
tio
n
.
T
h
is
s
ec
tio
n
p
r
es
en
ts
th
e
ex
p
er
im
e
n
tal
o
u
tc
o
m
e
an
d
d
is
cu
s
s
es
th
e
p
er
f
o
r
m
an
ce
c
o
n
ce
r
n
in
g
t
r
ain
in
g
ac
cu
r
ac
y
,
v
alid
atio
n
ac
cu
r
ac
y
,
co
n
f
u
s
io
n
m
atr
ix
,
an
d
o
th
er
class
if
icatio
n
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,
a
n
d
E
E
R
.
Fig
u
r
e
7
d
is
p
lay
s
th
e
m
o
d
el'
s
tr
ain
in
g
p
er
f
o
r
m
an
ce
tr
en
d
s
,
with
Fig
u
r
e
7
(
a)
d
ep
ictin
g
tr
ai
n
in
g
an
d
v
alid
atio
n
ac
cu
r
ac
y
an
d
Fig
u
r
e
7
(
b
)
d
etailin
g
tr
ain
in
g
an
d
v
alid
atio
n
lo
s
s
.
An
aly
s
is
o
f
th
e
g
r
a
p
h
t
r
en
d
s
r
ev
ea
ls
a
s
u
s
tain
ed
h
ig
h
tr
ain
in
g
ac
cu
r
ac
y
u
p
to
1
0
0
ep
o
c
h
s
,
r
ea
ch
in
g
9
9
.
9
8
%
an
d
9
9
.
9
6
%
f
o
r
tr
ain
in
g
an
d
v
alid
atio
n
,
r
esp
ec
tiv
ely
.
T
h
e
s
tab
le
tr
ain
in
g
lo
s
s
im
p
lies
m
in
im
al
er
r
o
r
s
i
n
d
is
tin
g
u
is
h
in
g
b
etwe
en
g
e
n
u
in
e
an
d
s
p
o
o
f
ed
s
ig
n
als.
T
h
e
f
r
eq
u
e
n
t
in
v
o
ca
tio
n
o
f
th
e
R
ed
u
ce
L
R
On
Plateau
ca
llb
ac
k
,
wh
ich
ad
ju
s
ts
th
e
le
ar
n
in
g
r
ate
wh
en
v
alid
atio
n
p
er
f
o
r
m
a
n
c
e
p
latea
u
s
,
in
d
icate
s
th
at
th
e
m
o
d
el
was
p
r
o
g
r
ess
iv
ely
r
e
f
in
in
g
to
war
d
s
o
p
tim
al
p
er
f
o
r
m
an
ce
.
I
n
Fig
u
r
e
8
,
th
e
co
n
f
u
s
io
n
m
atr
ix
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
.
1
4
,
No
.
1
,
Feb
r
u
ar
y
2
0
2
5
:
130
-
1
4
1
138
p
r
esen
ts
th
e
clas
s
if
icatio
n
p
er
f
o
r
m
an
ce
o
f
th
e
p
r
o
p
o
s
ed
s
p
ec
tr
al
-
tem
p
o
r
al
lear
n
in
g
m
o
d
el
f
o
r
b
o
n
af
id
e
an
d
s
p
o
o
f
class
es.
(
a)
(
b)
Fig
u
r
e
7
.
P
er
f
o
r
m
an
c
e
tr
en
d
o
f
th
e
m
o
d
el'
s
tr
ain
in
g
an
d
v
ali
d
atio
n
p
er
f
o
r
m
a
n
ce
o
n
(
a)
ac
c
u
r
ac
y
an
d
(
b
)
lo
s
s
Fig
u
r
e
8
.
T
h
e
co
n
f
u
s
io
n
m
atr
ix
f
o
r
t
h
e
class
if
icatio
n
m
o
d
el
T
h
e
in
ter
p
r
etatio
n
o
f
th
is
co
n
f
u
s
io
n
m
atr
ix
is
p
r
esen
ted
in
T
ab
le
1
.
A
clo
s
er
an
aly
s
is
o
f
th
e
o
u
tco
m
e
s
tatis
t
ics
s
h
o
ws
th
at
th
e
m
o
d
e
l
ac
cu
r
ately
id
en
tifie
d
2
,
4
4
3
in
s
tan
ce
s
as
b
o
n
af
id
e
o
u
t
o
f
2
,
5
4
8
in
s
tan
ce
s
in
t
h
e
test
d
ataset,
s
h
o
wca
s
in
g
its
ab
ilit
y
to
d
i
s
ce
r
n
g
en
u
in
e
in
s
t
an
ce
s
.
Ho
wev
er
,
th
er
e
wer
e
1
0
5
i
n
s
tan
ce
s
wh
er
e
b
o
n
af
id
e
(
g
en
u
in
e)
was
m
is
tak
en
as
s
p
o
o
f
ed
.
I
n
th
e
co
n
tex
t
o
f
s
p
o
o
f
in
s
tan
ce
s
,
th
e
p
r
o
p
o
s
ed
class
if
icatio
n
m
o
d
el
co
r
r
ec
tly
p
r
ed
icted
2
2
,
2
6
1
s
am
p
les
as
s
p
o
o
f
s
ig
n
als
o
u
t
o
f
2
2
,
2
9
6
test
s
am
p
les
,
it
also
f
alsely
ca
teg
o
r
ized
3
5
s
p
o
o
f
in
s
tan
ce
s
as b
o
n
af
id
e
.
T
ab
le
1
.
Su
m
m
a
r
y
o
f
co
n
f
u
s
io
n
p
lo
t w
ith
class
if
icatio
n
s
tatis
tics
f
o
r
s
p
o
o
f
a
n
d
b
o
n
af
i
d
e
cla
s
s
if
icatio
n
C
l
a
s
s l
a
b
e
l
To
t
a
l
sam
p
l
e
s
Tr
u
e
p
o
si
t
i
v
e
F
a
l
se
n
e
g
a
t
i
v
e
B
o
n
a
f
i
d
e
2
5
4
8
2
4
4
3
1
0
5
S
p
o
o
f
2
2
2
9
6
2
2
2
6
1
35
T
h
e
o
u
tco
m
e
s
tatis
tics
f
r
o
m
T
ab
le
1
d
em
o
n
s
tr
ate
th
e
co
n
s
id
er
ab
le
d
is
ce
r
n
m
en
t
ca
p
ab
ili
ties
o
f
th
e
p
r
o
p
o
s
ed
class
if
icatio
n
m
o
d
el
in
d
is
tin
g
u
is
h
in
g
b
etwe
en
b
o
n
af
i
d
e
an
d
s
p
o
o
f
ed
s
p
e
ak
er
s
.
T
h
e
m
o
d
el
d
em
o
n
s
tr
ated
h
ig
h
p
r
ec
is
io
n
,
ac
cu
r
ac
y
,
r
ec
all,
an
d
F1
s
co
r
e
s
f
o
r
s
p
ee
ch
i
n
s
tan
ce
s
in
th
e
te
s
t
d
ataset,
as
s
h
o
wn
in
T
ab
le
2
,
co
r
r
ec
tly
id
e
n
tify
i
n
g
m
o
s
t
in
s
tan
ce
s
o
f
b
o
n
af
id
e
an
d
s
p
o
o
f
.
I
n
t
h
e
b
o
n
af
id
e
cla
s
s
,
th
e
m
o
d
el
y
ield
s
im
p
r
ess
iv
e
r
esu
lts
,
with
a
p
r
ec
is
io
n
o
f
0
.
9
9
,
r
ec
all
o
f
0
.
9
6
an
d
F1
-
s
co
r
e
o
f
0
.
9
7
.
T
h
is
s
u
g
g
est
s
th
at
th
e
m
o
d
el
h
as
a
v
er
y
lo
w
f
alse
p
o
s
itiv
e
r
ate,
an
d
m
o
s
t
o
f
th
e
s
am
p
les
class
if
ied
as
b
o
n
af
i
d
e
ar
e
tr
u
e
p
o
s
itiv
es.
W
ith
a
lar
g
er
n
u
m
b
er
o
f
s
am
p
les,
th
e
s
p
o
o
f
class
witn
ess
es
a
r
em
ar
k
ab
le
p
er
f
o
r
m
an
ce
,
with
n
o
f
alse
n
eg
ativ
es
o
r
p
o
s
itiv
es.
W
h
en
co
n
s
id
er
in
g
th
e
o
v
e
r
all
o
u
tco
m
e,
th
e
m
o
d
el's
p
r
ec
is
io
n
is
9
9
.
5
6
%,
an
d
r
ec
all
is
h
ig
h
er
at
9
9
.
8
4
%,
d
en
o
tin
g
th
at
th
e
m
o
d
el
s
u
cc
ess
f
u
lly
id
en
tifie
s
alm
o
s
t
all
tr
u
e
in
s
tan
ce
s
o
f
ea
ch
c
lass
in
th
e
d
ataset.
Su
b
s
eq
u
en
tly
,
th
e
F1
-
s
co
r
e
is
also
h
ig
h
at
9
9
.
6
9
%,
s
ig
n
if
y
in
g
an
ex
ce
p
tio
n
al
b
alan
ce
b
et
wee
n
p
r
ec
is
io
n
an
d
r
ec
all.
Ad
d
itio
n
ally
,
E
E
R
is
0
.
0
1
4
,
wh
ic
h
is
q
u
ite
lo
w
v
s
th
e
ep
o
ch
s
p
r
esen
ted
in
Fig
u
r
e
9
.
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
Dis
crim
in
a
tive
d
ee
p
lea
r
n
in
g
b
a
s
ed
h
yb
r
id
s
p
ec
tr
o
-
temp
o
r
a
l fe
a
tu
r
es fo
r
…
(
P
r
a
n
ita
N
ir
a
j P
a
ls
a
p
u
r
e
)
139
T
ab
le
2
.
Pre
s
en
ts
a
q
u
a
n
titativ
e
ev
alu
atio
n
o
f
class
if
icatio
n
m
o
d
el
p
er
f
o
r
m
a
n
ce
m
etr
ics
To
t
a
l
s
a
m
p
l
e
s
C
l
a
s
s
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
F1
-
sc
o
r
e
2
5
4
8
B
o
n
a
f
i
d
e
0
.
9
9
0
.
9
6
0
.
9
7
2
2
2
9
6
S
p
o
o
f
1
.
0
0
1
.
0
0
1
.
0
0
O
v
e
r
a
l
l
o
u
t
c
o
m
e
A
c
c
u
r
a
c
y
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
F1
-
sc
o
r
e
EER
9
9
.
4
4
9
9
.
5
6
9
9
.
8
4
9
9
.
6
9
0
.
0
1
4
Fig
u
r
e
9
.
An
al
y
s
is
o
f
E
E
R
v
s
e
p
o
ch
s
Fro
m
Fig
u
r
e
9
,
it c
an
b
e
o
b
s
er
v
ed
th
at
th
e
E
E
R
s
tar
t
s
h
ig
h
at
0
.
5
8
1
in
ep
o
c
h
1
,
m
o
d
er
ately
d
ec
r
ea
s
es
to
0
.
5
7
4
in
ep
o
c
h
2
,
a
n
d
f
u
r
t
h
er
to
0
.
5
5
5
in
e
p
o
ch
3
.
T
h
is
in
d
icate
s
th
at
th
e
m
o
d
el
is
in
th
e
ea
r
ly
s
tag
es
o
f
lear
n
in
g
,
m
a
k
in
g
s
u
b
s
tan
tial
e
r
r
o
r
s
in
d
is
tin
g
u
is
h
in
g
b
etwe
e
n
class
es.
Ho
wev
er
,
th
er
e
is
an
im
p
r
o
v
em
en
t
in
ep
o
ch
s
4
to
7
,
with
E
E
R
v
alu
es
s
h
ar
p
ly
d
ec
lin
i
n
g
f
r
o
m
0
.
1
3
6
to
0
.
1
3
1
.
I
t
s
u
g
g
ests
th
at
t
h
e
m
o
d
el
h
as
m
a
d
e
s
ig
n
if
ican
t
lear
n
in
g
s
an
d
o
p
ti
m
izatio
n
s
,
im
p
r
o
v
in
g
its
o
v
er
all
clas
s
if
icatio
n
ac
cu
r
ac
y
.
I
n
th
e
last
th
r
ee
ep
o
ch
s
,
E
E
R
r
ed
u
ce
s
d
r
a
m
atica
lly
to
0
.
0
2
4
in
e
p
o
ch
8
an
d
r
em
ain
s
co
n
s
tan
t
in
ep
o
c
h
9
b
ef
o
r
e
a
ch
iev
in
g
t
h
e
lo
west
v
alu
e
o
f
0
.
0
1
4
in
ep
o
c
h
1
0
.
T
h
e
co
n
s
is
ten
t
r
ed
u
ctio
n
in
E
E
R
v
alu
es
f
r
o
m
ep
o
ch
1
to
ep
o
ch
1
0
d
em
o
n
s
tr
ates
th
e
ef
f
ec
tiv
e
n
ess
o
f
th
e
tr
ai
n
in
g
p
r
o
ce
s
s
,
in
d
icatin
g
a
co
n
s
is
ten
t
lear
n
in
g
an
d
ad
ap
tat
io
n
b
y
t
h
e
m
o
d
el,
o
p
tim
izin
g
its
class
if
icatio
n
ca
p
ab
ilit
ies
o
v
er
tim
e.
T
ab
le
3
p
r
o
v
id
es
a
c
o
m
p
ar
ativ
e
a
n
aly
s
is
in
ter
m
s
o
f
E
E
R
ag
ain
s
t two
s
im
ilar
ex
is
tin
g
m
eth
o
d
o
lo
g
ies.
T
ab
le
3
.
Pre
s
en
ts
co
m
p
a
r
ativ
e
an
aly
s
is
in
ter
m
s
o
f
E
E
R
Ex
i
s
t
i
n
g
1
[
1
5
]
Ex
i
s
t
i
n
g
2
[
2
7
]
P
r
o
p
o
se
d
2
.
1
9
0
.
2
3
0
.
0
1
4
T
h
e
co
m
p
ar
ativ
e
a
n
aly
s
is
p
r
esen
ted
in
T
ab
le
3
,
in
ter
m
s
o
f
E
E
R
,
clea
r
ly
d
em
o
n
s
tr
ates th
e
s
u
p
er
io
r
ity
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el
o
v
er
th
e
ex
is
tin
g
m
o
d
els [
1
5
]
,
[
2
7
]
.
Mittal
an
d
Du
a
[
1
5
]
d
em
o
n
s
tr
ated
an
E
E
R
o
f
2
.
1
9
,
wh
ich
is
co
n
s
id
er
ab
ly
h
ig
h
e
r
th
an
th
e
ex
is
tin
g
m
o
d
el
[
2
7
]
a
n
d
p
r
o
p
o
s
ed
s
y
s
tem
.
I
n
Mittal
an
d
Du
a
[
1
5
]
o
n
e
-
class
lear
n
in
g
m
eth
o
d
is
im
p
lem
en
ted
to
id
en
tif
y
g
en
u
in
e
s
p
ee
ch
r
ep
r
esen
tatio
n
s
an
d
in
tr
o
d
u
ce
an
a
n
g
u
lar
m
ar
g
in
with
o
u
t
d
ata
au
g
m
en
tatio
n
.
Ho
wev
er
,
th
is
ap
p
r
o
a
ch
h
as
lim
itatio
n
s
in
ter
m
s
o
f
s
ca
lab
i
lity
an
d
ad
ap
tab
ilit
y
to
d
iv
er
s
e
s
p
o
o
f
i
n
g
s
ce
n
ar
io
s
.
I
n
c
o
n
tr
ast,
Ad
i
b
an
et
a
l.
[
2
7
]
a
d
o
p
te
d
au
to
e
n
co
d
er
an
d
Siam
ese
n
etwo
r
k
s
,
wh
ich
r
esu
lted
E
E
R
o
f
0
.
2
3
%,
d
e
m
o
n
s
tr
atin
g
en
h
a
n
ce
d
p
er
f
o
r
m
an
ce
co
m
p
ar
ed
t
o
[
1
5
]
in
d
if
f
er
en
tiatin
g
leg
itima
te
f
r
o
m
s
p
o
o
f
ed
v
o
ices.
I
n
th
e
ca
s
e
o
f
th
e
p
r
o
p
o
s
ed
s
y
s
tem
h
as
ac
h
iev
ed
a
q
u
ite
lo
w
E
E
R
o
f
0
.
0
1
4
,
in
d
icatin
g
a
s
u
p
er
io
r
lev
el
o
f
ac
cu
r
ac
y
an
d
r
eliab
ilit
y
in
p
r
ev
en
tin
g
v
o
ic
e
s
p
o
o
f
in
g
attac
k
s
,
s
u
g
g
esti
n
g
th
e
m
o
d
el
h
as
ac
h
iev
ed
a
h
ig
h
ly
o
p
tim
ized
an
d
ac
cu
r
ate
s
tate,
m
a
k
in
g
it
r
eli
ab
le
f
o
r
class
if
y
in
g
b
o
n
af
id
e
an
d
s
p
o
o
f
class
es.
T
h
e
p
r
o
p
o
s
ed
s
y
s
tem
ac
h
iev
e
s
g
o
o
d
p
er
f
o
r
m
an
ce
d
u
e
to
t
h
e
u
s
ag
e
o
f
h
y
b
r
i
d
s
p
ec
tr
o
-
tem
p
o
r
al
f
ea
tu
r
es c
o
m
b
in
ed
with
a
s
o
p
h
is
ticated
lear
n
in
g
m
o
d
el.
4.
CO
NCLU
SI
O
N
T
h
is
s
tu
d
y
p
r
o
p
o
s
es
a
n
o
v
el
le
ar
n
in
g
m
eth
o
d
m
o
d
el
to
s
o
lv
e
th
e
p
r
o
b
lem
o
f
lo
g
ical
s
p
ee
ch
s
p
o
o
f
in
g
attac
k
s
o
n
ASV
s
y
s
tem
s
.
T
h
e
m
o
d
el
c
o
m
b
in
es
t
h
e
lear
n
in
g
ca
p
ab
ilit
ies
o
f
C
NN
an
d
L
ST
M
n
etwo
r
k
s
,
wh
ich
ar
e
g
o
o
d
at
lear
n
in
g
f
r
o
m
s
p
e
ctr
al
an
d
tem
p
o
r
al
d
ata.
T
h
e
p
r
o
p
o
s
ed
m
eth
o
d
s
h
o
ws
r
em
ar
k
ab
le
ad
a
p
tab
ilit
y
in
Evaluation Warning : The document was created with Spire.PDF for Python.