T
E
L
K
O
M
N
I
K
A
T
elec
o
m
m
un
ica
t
io
n,
Co
m
pu
t
ing
,
E
lect
ro
nics
a
nd
Co
ntr
o
l
Vo
l.
18
,
No
.
5
,
Octo
b
er
2
0
2
0
,
p
p
.
2
5
8
7~
2
5
9
6
I
SS
N:
1
6
9
3
-
6
9
3
0
,
ac
cr
ed
ited
First Gr
ad
e
b
y
Kem
en
r
is
tek
d
i
k
ti,
Dec
r
ee
No
: 2
1
/E/KPT
/2
0
1
8
DOI
:
1
0
.
1
2
9
2
8
/TE
L
KOM
NI
K
A.
v
1
8
i5
.
1
4
2
4
6
2587
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//jo
u
r
n
a
l.u
a
d
.
a
c.
id
/in
d
ex
.
p
h
p
/TELK
OM
N
I
K
A
So
und event
det
e
ction using
de
ep
neura
l net
wo
rks
Suk
-
H
wa
n J
un
g
,
Yo
ng
-
J
o
o
Chun
g
De
p
a
rtme
n
t
o
f
El
e
c
tro
n
ics
,
Ke
imy
u
n
g
U
n
iv
e
rsit
y
,
Ko
re
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Oct
1
,
2
0
1
9
R
ev
is
ed
Ap
r
2
3
,
2
0
2
0
Acc
ep
ted
Ma
y
8
,
2
0
2
0
We
a
p
p
li
e
d
v
a
ri
o
u
s
a
rc
h
it
e
c
tu
re
s
o
f
d
e
e
p
n
e
u
ra
l
n
e
two
r
k
s
fo
r
so
u
n
d
e
v
e
n
t
d
e
tec
ti
o
n
a
n
d
c
o
m
p
a
re
d
t
h
e
ir
p
e
rfo
rm
a
n
c
e
u
sin
g
two
d
iffere
n
t
d
a
tas
e
ts.
F
e
e
d
fo
rwa
rd
n
e
u
ra
l
n
e
two
r
k
(F
NN
),
c
o
n
v
o
l
u
ti
o
n
a
l
n
e
u
ra
l
n
e
tw
o
r
k
(CNN
),
re
c
u
rre
n
t
n
e
u
ra
l
n
e
two
rk
(RNN
)
a
n
d
c
o
n
v
o
l
u
ti
o
n
a
l
re
c
u
rre
n
t
n
e
u
r
a
l
n
e
two
r
k
(CRNN
)
we
re
imp
lem
e
n
ted
u
si
n
g
h
y
p
e
r
-
p
a
ra
m
e
ters
o
p
ti
m
ize
d
fo
r
e
a
c
h
a
rc
h
it
e
c
tu
re
a
n
d
d
a
tas
e
t.
T
h
e
re
su
l
ts
sh
o
w
th
a
t
t
h
e
p
e
rf
o
rm
a
n
c
e
o
f
d
e
e
p
n
e
u
ra
l
n
e
two
rk
s
v
a
rie
d
sig
n
ifi
c
a
n
tl
y
d
e
p
e
n
d
i
n
g
o
n
t
h
e
lea
rn
in
g
ra
te,
wh
ich
c
a
n
b
e
o
p
ti
m
ize
d
b
y
c
o
n
d
u
c
ti
n
g
a
se
ries
o
f
e
x
p
e
rime
n
ts
o
n
t
h
e
v
a
li
d
a
ti
o
n
d
a
ta
o
v
e
r
p
re
d
e
term
in
e
d
ra
n
g
e
s.
Am
o
n
g
t
h
e
imp
lem
e
n
ted
a
rc
h
it
e
c
tu
re
s,
th
e
CRNN
p
e
rfo
rm
e
d
b
e
st
u
n
d
e
r
a
ll
tes
ti
n
g
c
o
n
d
it
io
n
s,
fo
ll
o
we
d
b
y
CNN
.
Alth
o
u
g
h
RNN
wa
s
e
ffe
c
ti
v
e
in
trac
k
in
g
t
h
e
ti
m
e
-
c
o
rre
latio
n
i
n
fo
rm
a
ti
o
n
i
n
a
u
d
i
o
sig
n
a
ls,
it
e
x
h
ib
it
e
d
i
n
fe
rio
r
p
e
rfo
rm
a
n
c
e
c
o
m
p
a
re
d
to
th
e
CNN
a
n
d
t
h
e
CRNN
.
Ac
c
o
rd
in
g
ly
,
it
is
n
e
c
e
ss
a
ry
to
d
e
v
e
l
o
p
m
o
re
o
p
ti
m
iza
ti
o
n
str
a
teg
ies
fo
r
imp
lem
e
n
ti
n
g
RNN
in
s
o
u
n
d
e
v
e
n
t
d
e
tec
ti
o
n
.
K
ey
w
o
r
d
s
:
C
o
n
v
o
lu
t
io
n
al
n
eu
r
al
n
etwo
r
k
C
o
n
v
o
lu
tio
n
al
r
e
cu
r
r
e
n
t
n
eu
r
al
n
etwo
r
k
Dee
p
n
eu
r
al
n
etwo
r
k
s
Feed
f
o
r
war
d
n
eu
r
al
n
etwo
r
k
R
ec
u
r
en
t n
eu
r
al
n
etwo
r
k
So
u
n
d
ev
e
n
t 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
:
Yo
n
g
-
J
o
o
C
h
u
n
g
,
Dep
ar
tm
en
t o
f
E
lectr
o
n
ics,
Keim
y
u
n
g
Un
i
v
er
s
ity
,
Ko
r
ea
.
E
m
ail:
y
jju
n
g
@
k
m
u
.
ac
.
k
r
1.
I
NT
RO
D
UCT
I
O
N
Au
to
m
atic
s
o
u
n
d
ev
en
t
d
etec
tio
n
is
a
p
atter
n
r
ec
o
g
n
itio
n
tech
n
iq
u
e
th
at
au
to
m
atica
lly
id
en
tifie
s
v
ar
io
u
s
s
o
u
n
d
ev
e
n
ts
o
cc
u
r
r
in
g
d
aily
,
s
u
ch
as
g
lass
b
r
ea
k
in
g
,
b
ab
y
cr
y
in
g
,
p
eo
p
le
s
cr
ea
m
in
g
an
d
ca
r
h
o
r
n
in
g
.
I
n
ad
d
itio
n
to
i
d
en
tify
in
g
th
e
lab
el
o
f
s
o
u
n
d
e
v
en
ts
,
it
d
etec
ts
th
eir
o
n
s
et
an
d
o
f
f
s
et
tim
es
.
Au
to
m
atic
s
o
u
n
d
ev
en
t
d
etec
tio
n
h
as
r
ec
e
n
tly
g
ain
ed
p
o
p
u
lar
ity
o
win
g
to
its
n
u
m
er
o
u
s
p
o
s
s
ib
le
ap
p
l
icatio
n
s
,
in
clu
d
in
g
s
u
r
v
eillan
ce
,
u
r
b
an
s
o
u
n
d
an
aly
s
is
,
in
f
o
r
m
atio
n
r
etr
iev
al
f
r
o
m
m
u
ltime
d
ia
co
n
ten
t,
h
e
alth
ca
r
e,
b
ir
d
ca
ll
d
etec
tio
n
,
an
d
au
to
n
o
m
o
u
s
v
e
h
icles [
1
-
6
]
.
T
o
en
co
u
r
a
g
e
r
esear
ch
in
th
e
g
en
er
al
ar
ea
o
f
s
o
u
n
d
s
ig
n
al
class
if
icatio
n
in
clu
d
in
g
s
o
u
n
d
ev
en
t
d
etec
tio
n
,
t
h
e
“De
tectio
n
an
d
C
las
s
if
icatio
n
o
f
Aco
u
s
tic
Sce
n
es
an
d
E
v
e
n
ts
(
DC
ASE)
”
ch
allen
g
e
was
h
el
d
in
2
0
1
3
,
2
0
1
6
,
2
0
1
7
,
2
0
1
8
,
a
n
d
2
0
1
9
[
7
-
1
1
]
.
I
t
in
clu
d
es
two
d
if
f
er
en
t
ca
teg
o
r
ies:
ac
o
u
s
tic
s
ce
n
e
class
if
icatio
n
an
d
s
o
u
n
d
e
v
en
t
d
etec
tio
n
.
I
n
th
e
f
o
r
m
er
,
th
e
ty
p
e
o
f
ac
o
u
s
tic
en
v
ir
o
n
m
e
n
t
is
d
eter
m
in
e
d
u
s
in
g
a
lo
n
g
s
eg
m
en
t
o
f
au
d
io
s
ig
n
als,
wh
er
ea
s
,
in
th
e
latter
,
s
p
ec
if
ic
s
o
u
n
d
ev
en
t
s
o
cc
u
r
r
in
g
in
an
ac
o
u
s
tic
s
ce
n
e
ar
e
r
ec
o
g
n
ize
d
.
I
n
th
is
s
tu
d
y
,
we
o
n
ly
f
o
cu
s
o
n
s
o
u
n
d
ev
en
t
d
etec
tio
n
u
s
in
g
t
wo
p
u
b
lic
d
atab
ases
f
r
o
m
DC
ASE
2
0
1
6
an
d
[
1
2
]
.
B
ef
o
r
e
th
e
em
er
g
en
ce
o
f
m
eh
o
d
s
b
ased
o
n
d
ee
p
n
eu
r
al
n
etwo
r
k
s
,
Gau
s
s
ian
m
ix
tu
r
e
m
o
d
el
s
(
GM
Ms)
wer
e
wid
ely
u
s
ed
in
s
o
u
n
d
ev
en
t
d
etec
tio
n
.
I
n
f
ac
t,
a
GM
M
was
u
s
ed
as
a
b
aselin
e
r
ec
o
g
n
izer
in
th
e
DC
ASE
2
0
1
6
c
h
allen
g
e
f
o
r
T
ask
1
(
a
co
u
s
tic
s
ce
n
e
class
if
icatio
n
)
a
n
d
T
ask
3
(
m
o
n
o
p
h
o
n
ic
s
o
u
n
d
ev
en
t
d
etec
tio
n
)
.
T
h
e
s
im
p
le
GM
M
-
b
ased
b
ag
-
of
-
f
r
am
es
ap
p
r
o
ac
h
was
ad
o
p
te
d
in
th
e
b
aselin
e
s
y
s
tem
[
1
3
]
,
wh
er
e
th
e
m
el
-
f
r
e
q
u
en
c
y
ce
p
s
tr
al
co
ef
f
icien
ts
(
MFC
C
s
)
,
wh
ich
h
a
v
e
b
ee
n
wid
ely
em
p
lo
y
ed
in
s
p
ee
ch
r
ec
o
g
n
itio
n
,
wer
e
u
s
ed
as
ac
o
u
s
tic
f
ea
tu
r
es
f
o
r
th
e
GM
M.
I
n
ad
d
itio
n
to
GM
Ms,
tr
ad
itio
n
al
m
ac
h
in
e
lear
n
in
g
m
eth
o
d
s
,
s
u
ch
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
1
6
9
3
-
6
9
3
0
T
E
L
KOM
NI
KA
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l Co
n
tr
o
l
,
Vo
l.
18
,
No
.
5
,
Octo
b
e
r
2
0
2
0
:
2
5
8
7
-
259
6
2588
as
s
u
p
p
o
r
t
v
ec
to
r
m
ac
h
in
es
(
S
VM
s
)
[
1
4
]
an
d
n
o
n
-
n
eg
ativ
e
m
atr
ix
f
ac
t
o
r
izatio
n
(
NM
F)
[
1
5
]
,
wer
e
also
wid
el
y
u
s
ed
in
s
o
u
n
d
ev
e
n
t
d
etec
tio
n
b
ef
o
r
e
th
eir
in
f
er
i
o
r
ity
t
o
r
ec
e
n
t d
ee
p
lear
n
i
n
g
-
b
ased
m
eth
o
d
s
was d
em
o
n
s
tated
.
Du
r
in
g
th
e
last
d
ec
ad
e,
d
ee
p
n
eu
r
al
n
etw
o
r
k
s
h
a
v
e
ac
h
iev
e
d
g
r
ea
t
s
u
cc
ess
in
im
a
g
e
class
if
icatio
n
,
s
p
ee
ch
r
ec
o
g
n
itio
n
an
d
m
ac
h
i
n
e
tr
an
s
latio
n
[
1
6
-
2
0
]
.
C
u
r
r
e
n
tly
,
d
ee
p
n
eu
r
al
n
etwo
r
k
s
ex
h
ib
it
s
tate
-
of
-
th
e
-
ar
t
p
er
f
o
r
m
an
ce
in
all
th
ese
d
o
m
ai
n
s
.
I
n
s
o
u
n
d
ev
en
t
d
etec
tio
n
,
F
NNs
h
av
e
ac
h
iev
ed
b
etter
p
er
f
o
r
m
an
ce
c
o
m
p
ar
e
d
with
GM
Ms
an
d
SVMs
an
d
it
ap
p
ea
r
s
th
at
t
h
ey
h
av
e
r
ep
l
ac
ed
th
e
tr
a
d
itio
n
al
m
et
h
o
d
s
in
th
e
s
o
u
n
d
e
v
en
t
d
etec
tio
n
.
Owin
g
to
th
eir
s
i
m
p
le
ar
ch
itectu
r
e,
FNNs
h
av
e
ad
v
an
tag
es
o
v
er
o
th
er
d
ee
p
n
eu
r
al
n
etwo
r
k
s
.
Sp
ec
if
ically
,
f
ewe
r
p
ar
a
m
eter
s
an
d
l
ess
co
m
p
u
tatio
n
al
tim
e
a
r
e
r
eq
u
ir
e
d
.
Sev
er
al
f
r
am
es
o
f
n
eig
h
b
o
r
in
g
a
u
d
io
f
ea
tu
r
es
(
u
s
u
ally
,
lo
g
-
m
el
f
ilter
b
an
k
(
L
MFB
)
en
er
g
ies)
ar
e
c
o
n
ca
ten
ated
in
th
e
tim
e
-
d
o
m
a
in
s
o
th
at
th
ey
ca
n
b
e
u
s
ed
as
in
p
u
t
to
th
e
n
etw
o
r
k
.
Su
b
s
eq
u
en
tly
,
th
ey
a
r
e
m
u
ltip
lied
b
y
weig
h
t
m
atr
ice
s
an
d
p
ass
th
r
o
u
g
h
n
o
n
lin
ea
r
f
u
n
ctio
n
s
,
a
n
d
,
h
en
ce
th
ey
ar
e
f
o
r
war
d
p
r
o
p
ag
a
ted
.
Ho
wev
er
,
th
e
s
tr
u
ctu
r
e
o
f
an
FNN
ca
n
n
o
t
ef
f
ec
tiv
ely
co
m
p
en
s
ate
f
o
r
t
h
e
tr
an
s
latio
n
al
v
ar
ian
ce
s
o
cc
u
r
r
i
n
g
in
im
ag
e
s
ig
n
als
o
win
g
to
th
e
f
ix
ed
co
n
n
ec
tio
n
s
b
etwe
en
th
e
in
p
u
t
an
d
h
id
d
en
u
n
its
.
Similar
p
r
o
b
lem
s
o
cc
u
r
in
s
o
u
n
d
ev
e
n
t
d
etec
tio
n
b
e
ca
u
s
e
th
e
v
ar
iatio
n
s
in
th
e
tim
e
-
f
r
eq
u
e
n
cy
d
o
m
ai
n
o
f
t
h
e
au
d
io
s
ig
n
al
m
ay
n
o
t
b
e
well
m
o
d
eled
b
y
th
e
FNN.
An
o
th
er
p
r
o
b
lem
is
th
at
th
e
tem
p
o
r
al
c
o
n
tex
t
is
r
es
tr
icted
to
s
h
o
r
t
-
tim
e
win
d
o
ws
o
f
th
e
in
p
u
t
au
d
i
o
;
th
er
ef
o
r
e,
it
is
d
if
f
icu
lt
to
m
o
d
el
lo
n
g
-
ter
m
co
r
r
elatio
n
s
in
th
e
a
u
d
io
s
ig
n
als.
C
o
m
p
ar
ed
with
FNNs
,
C
NNs
ca
n
ad
d
r
ess
th
e
p
r
o
b
lem
o
f
t
im
e
-
f
r
eq
u
e
n
cy
d
o
m
ain
v
ar
iatio
n
s
m
o
r
e
ef
f
icien
tly
.
Ho
wev
er
,
C
NNs
ca
n
n
o
t
ef
f
ec
tiv
el
y
m
o
d
el
lo
n
g
-
ter
m
co
n
te
x
t
co
r
r
elatio
n
s
i
n
th
e
tim
e
-
d
o
m
ai
n
.
R
ec
u
r
r
en
t
n
eu
r
al
n
etwo
r
k
s
(
R
NNs)
h
av
e
b
ee
n
q
u
ite
s
u
cc
ess
f
u
l
in
m
o
d
elin
g
tem
p
o
r
al
co
n
t
ex
t
in
f
o
r
m
atio
n
in
s
p
ee
ch
r
ec
o
g
n
itio
n
.
Ho
wev
e
r
,
o
win
g
to
th
eir
s
h
o
r
tco
m
in
g
s
in
ca
p
tu
r
in
g
th
e
in
v
ar
ia
n
ce
in
th
e
tim
e
-
f
r
e
q
u
en
c
y
do
m
ain
,
R
NNs
ar
e
u
n
a
b
le
to
o
u
tp
er
f
o
r
m
C
NNs
in
s
o
u
n
d
ev
en
t
d
etec
tio
n
.
Sev
er
al
a
p
p
r
o
ac
h
es
h
av
e
b
ee
n
p
r
o
p
o
s
ed
f
o
r
co
m
b
i
n
in
g
C
N
Ns
an
d
R
NNs
to
tak
e
ad
v
an
tag
e
o
f
b
o
th
n
etwo
r
k
s
.
R
ec
en
tly
,
co
n
v
o
lu
tio
n
al
r
ec
u
r
r
en
t
n
eu
r
al
n
etwo
r
k
s
(
C
R
NNs),
a
co
m
b
in
atio
n
o
f
C
NNs
an
d
R
N
Ns
in
a
s
in
g
le
n
etwo
r
k
,
h
av
e
b
ee
n
p
r
o
p
o
s
ed
f
o
r
s
o
u
n
d
ev
en
t
d
ete
ctio
n
,
s
p
ee
ch
r
ec
o
g
n
itio
n
an
d
m
u
s
ic
class
if
icatio
n
[
1
2
,
2
1
-
2
4
]
.
I
n
th
is
p
ap
er
,
we
p
r
o
p
o
s
e
th
e
u
s
e
o
f
a
C
R
NN
in
p
o
ly
p
h
o
n
ic
an
d
s
ce
n
e
-
in
d
ep
en
d
en
t
s
o
u
n
d
ev
e
n
t
d
etec
tio
n
an
d
s
u
g
g
est
o
p
tim
al
h
y
p
er
-
p
ar
am
eter
s
an
d
tr
ain
in
g
s
tr
ateg
ies.
T
h
u
s
,
th
e
ad
v
a
n
tag
e
o
f
C
R
NNs
o
v
er
C
NNs
an
d
R
NNs
i
s
ex
p
ec
ted
to
b
e
m
ax
im
ized
.
W
e
ev
alu
ated
th
e
p
er
f
o
r
m
an
ce
o
f
th
e
C
R
NN
o
n
r
ec
en
t
d
atasets
,
in
clu
d
in
g
f
r
o
m
th
e
DC
ASE
2
0
1
6
ch
allen
g
e.
W
e
also
co
m
p
ar
ed
th
e
p
er
f
o
r
m
a
n
ce
o
f
th
e
C
R
NN
wi
th
a
C
NN
,
an
FNN
an
d
a
n
R
NN
s
o
th
at
th
e
ad
v
an
at
g
es
o
f
t
h
e
C
R
NN
m
ay
b
e
b
etter
u
n
d
er
s
to
o
d
.
T
h
e
r
em
ain
d
e
r
o
f
th
is
p
ap
er
is
o
r
g
an
ized
as
f
o
llo
ws
;
i
n
s
ec
tio
n
2
,
we
p
r
esen
t
f
ea
tu
r
e
ex
tr
ac
tio
n
m
eth
o
d
a
n
d
d
ee
p
n
eu
r
al
ar
ch
itectu
r
es
u
s
ed
in
t
h
is
s
tu
d
y
.
I
n
s
ec
t
io
n
3
,
we
p
r
esen
t
an
d
d
is
cu
s
s
th
e
r
e
s
u
lts
o
f
v
ar
io
u
s
ex
p
er
im
en
ts
i
n
v
o
lv
in
g
th
e
FNN,
C
NN,
R
N
N
an
d
C
R
NN.
Fin
all
y
,
s
ec
tio
n
4
c
o
n
clu
d
es th
e
p
ap
er
.
2.
F
E
AT
U
RE
E
XT
RAC
T
I
O
N
AND
DE
E
P
NE
U
RAL A
RC
H
I
T
UR
E
S
2
.
1
.
F
ea
t
ure
ex
t
ra
ct
io
n
I
n
th
is
s
tu
d
y
,
we
u
s
e
L
MFB
o
u
tp
u
ts
as
f
ea
tu
r
es
f
o
r
d
e
ep
n
eu
r
al
n
etwo
r
k
s
.
W
e
f
ir
s
t
co
m
p
u
te
th
e
s
h
o
r
t
-
tim
e
Fo
u
r
ier
tr
an
s
f
o
r
m
(
STFT
)
o
f
4
0
-
m
s
au
d
i
o
s
ig
n
als
th
at
wer
e
s
am
p
led
at
4
4
.
1
k
Hz.
T
h
e
STFT
is
co
m
p
u
ted
ev
er
y
2
0
m
s
with
5
0
%
o
v
er
lap
.
A
to
tal
o
f
4
0
b
a
n
d
s
o
f
m
el
-
f
ilter
b
a
n
k
a
r
e
ex
tr
ac
ted
f
r
o
m
th
e
STFT
with
th
e
r
a
n
g
e
o
f
0
~2
2
,
0
5
0
Hz
an
d
ar
e
l
o
g
-
tr
a
n
s
f
o
r
m
ed
to
o
b
tain
a
4
0
-
d
im
e
n
s
io
n
al
L
MFB
f
o
r
ea
ch
2
0
m
s
tim
e
f
r
am
e.
Af
ter
c
o
m
p
u
tin
g
th
e
L
MFB
s
,
we
n
o
r
m
alize
th
em
b
y
s
u
b
tr
ac
tin
g
th
e
m
ea
n
an
d
d
iv
i
d
in
g
b
y
th
e
s
tan
d
ar
d
d
ev
iatio
n
co
m
p
u
ted
f
r
o
m
th
e
t
r
ain
in
g
d
ata.
2
.
2
.
F
NN
T
h
e
af
o
e
r
m
en
tio
n
ed
4
0
-
d
im
e
n
tio
n
al
L
MFB
s
ar
e
u
s
ed
as
f
ea
tu
r
es.
Fiv
e
s
u
cc
ess
iv
e
tim
e
f
r
am
es
ar
e
co
n
ca
ten
ated
to
f
o
r
m
1
0
0
-
d
im
en
s
io
n
al
f
ea
tu
r
e
v
ec
t
o
r
s
as
th
e
in
p
u
t
to
th
e
FNN.
E
ac
h
o
f
th
e
t
wo
2
h
id
d
e
n
lay
er
s
h
as
1
6
0
0
h
id
d
en
u
n
its
with
R
eL
U
ac
tiv
atio
n
.
On
e
o
u
tp
u
t
lay
er
with
s
ig
m
o
i
d
ac
tiv
atio
n
h
as
K
u
n
its
wh
er
e
K
is
th
e
n
u
m
b
er
o
f
s
o
u
n
d
ev
e
n
t
class
es
to
b
e
r
ec
o
g
n
ized
.
T
h
e
o
u
tp
u
ts
o
f
th
e
s
ig
m
o
id
ac
tiv
atio
n
ar
e
tak
en
as
th
e
p
o
s
ter
io
r
p
r
o
b
a
b
ilit
ies
f
o
r
ea
ch
o
f
th
e
class
es,
an
d
th
e
b
in
ar
ized
o
u
tp
u
ts
ar
e
co
m
p
ar
e
d
w
ith
th
e
g
r
o
u
n
d
tr
u
th
tab
le
to
d
eter
m
in
e
th
e
ac
cu
r
ac
y
o
f
th
e
FNN.
2
.
3
.
C
NN
T
h
e
in
p
u
t
to
th
e
C
NN
is
T
×
40
L
MFB
f
ea
tu
r
es,
an
d
th
e
o
v
er
all
s
tr
u
ctu
r
e
o
f
t
h
e
n
etwo
r
k
is
s
h
o
wn
in
Fig
u
r
e
1
.
W
e
u
s
e
d
if
f
er
en
t
s
tr
u
ctu
r
es
f
o
r
ea
ch
o
f
th
e
two
s
elec
ted
d
atasets
.
T
h
e
s
tr
u
ctu
r
e
in
th
e
f
ig
u
r
e
is
u
s
ed
f
o
r
t
h
e
T
UT
s
o
u
n
d
ev
e
n
ts
2
0
1
6
d
ataset.
T
h
e
T
f
r
am
es
o
f
th
e
4
0
-
d
im
en
s
io
n
al
L
M
FB
s
ar
e
in
p
u
t
to
th
e
co
n
v
o
lu
tio
n
al
lay
er
with
2
5
6
f
ea
tu
r
e
m
a
p
s
,
an
d
ea
ch
f
ea
tu
r
e
m
ap
h
as
a
two
-
d
im
en
s
io
n
al
5
×
5
co
n
v
o
l
u
tio
n
al
f
ilter
with
R
eL
U
ac
tiv
atio
n
.
T
h
e
o
u
tp
u
t
o
f
th
e
c
o
n
v
o
lu
tio
n
al
lay
e
r
p
ass
es
th
r
o
u
g
h
a
n
o
n
-
o
v
e
r
lap
p
in
g
m
a
x
p
o
o
lin
g
lay
er
to
r
ed
u
ce
th
e
d
im
en
s
i
o
n
ality
o
f
th
e
d
ata.
W
e
co
m
p
u
te
th
e
m
ax
p
o
o
lin
g
o
p
er
atio
n
o
n
ly
in
t
h
e
f
r
eq
u
en
c
y
d
o
m
ain
t
o
r
etain
th
e
tem
p
o
r
a
l
in
f
o
r
m
atio
n
i
n
th
e
L
MFB
s
.
T
h
is
i
s
in
co
n
tr
ast
to
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KOM
NI
KA
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l Co
n
tr
o
l
S
o
u
n
d
ev
en
t
d
etec
tio
n
u
s
in
g
d
ee
p
n
eu
r
a
l n
etw
o
r
ks
(
Suk
-
Hw
a
n
Ju
n
g
)
2589
C
NNs
u
s
ed
in
im
ag
e
clas
s
if
ic
atio
n
,
wh
er
e
th
e
m
ax
p
o
o
in
g
o
p
er
atio
n
is
p
er
f
o
r
m
e
d
in
b
o
th
d
im
en
s
io
n
s
.
Un
lik
e
im
ag
e
s
ig
n
als,
th
e
tim
e
r
eso
lu
t
io
n
in
f
o
r
m
atio
n
s
h
o
u
ld
b
e
m
ain
tain
ed
in
t
h
e
au
d
io
s
ig
n
als
to
d
eter
m
in
e
th
e
o
n
s
et
an
d
o
f
f
s
et
tim
es
in
t
h
e
s
o
u
n
d
ev
en
t
d
etec
tio
n
.
T
h
e
r
e
ar
e
th
r
ee
C
NN
lay
er
s
,
an
d
th
e
o
u
tp
u
t
h
as
a
d
im
en
s
io
n
o
f
T
×
1
×
256
,
wh
er
e
t
h
e
d
im
en
s
io
n
o
f
th
e
f
r
eq
u
en
cy
d
o
m
ain
is
r
ed
u
c
ed
to
1
,
wh
er
ea
s
t
h
e
d
i
m
en
s
io
n
o
f
th
e
tim
e
d
o
m
ain
is
u
n
ch
a
n
g
e
d
,
as
m
en
tio
n
ed
p
r
ev
i
o
u
s
ly
.
T
h
e
o
u
t
p
u
t
o
f
th
e
C
NN
lay
er
s
is
f
ed
in
to
a
s
in
g
le
f
ee
d
-
f
o
r
war
d
lay
er
th
at
h
as
2
5
6
u
n
its
with
R
eL
U
ac
tiv
atio
n
.
T
h
e
f
in
al
o
u
tp
u
t
lay
er
with
K(
=n
u
m
b
er
o
f
class
es)
u
n
its
o
f
s
ig
m
o
id
ac
tiv
atio
n
f
o
llo
ws
th
e
f
ee
d
-
f
o
r
war
d
lay
e
r
an
d
y
ield
s
th
e
s
o
u
n
d
ev
en
t
ac
tiv
ity
p
r
o
b
ab
ilit
ies
f
o
r
ea
ch
s
o
u
n
d
class
at
ea
ch
tim
e
f
r
am
e.
Fin
ally
,
th
e
p
r
o
b
ab
ilit
ies
ar
e
b
in
ar
ized
af
ter
th
r
esh
o
ld
in
g
o
v
er
a
co
n
s
tan
t
v
alu
e
(
0
.
5
)
,
an
d
th
e
ac
tiv
ity
o
f
a
class
at
a
tim
e
f
r
a
m
e
is
d
eter
m
in
ed
to
b
e
ac
tiv
e
o
r
in
ac
tiv
e
d
e
p
en
d
in
g
o
n
wh
eth
er
th
e
b
in
a
r
ized
p
r
o
b
ab
ilit
y
is
1
o
r
0
.
2
.
4
.
R
NN
T
h
e
ar
ch
itectu
r
e
o
f
th
e
R
NN
u
s
ed
in
th
is
s
tu
d
y
f
o
r
T
UT
s
o
u
n
d
ev
en
ts
2
0
1
6
is
s
h
o
wn
i
n
Fig
u
r
e
2
.
T
×
40
L
MFB
f
ea
tu
r
es
ar
e
u
s
ed
as
th
e
in
p
u
t
o
f
th
e
GR
U
in
th
e
R
N
N
a
r
ch
itectu
r
e.
W
e
u
s
e
th
r
ee
lay
er
s
o
f
GR
Us
with
2
5
6
u
n
its
,
f
o
llo
wed
b
y
f
o
u
r
f
ee
d
-
f
o
r
wa
r
d
lay
e
r
s
with
2
5
6
u
n
its
.
T
h
e
o
u
tp
u
t
lay
er
h
as
K
u
n
its
with
s
ig
m
o
id
ac
tiv
atio
n
.
B
y
u
s
in
g
m
u
ltip
le
f
ee
d
-
f
o
r
war
d
la
y
er
s
,
th
e
C
NN
an
d
R
NN
h
av
e
eq
u
ally
d
ee
p
l
ev
els,
th
u
s
allo
wi
n
g
th
eir
p
er
f
o
r
m
an
ce
co
m
p
a
r
is
o
n
.
Fig
u
r
e
1
.
C
NN
ar
ch
itectu
r
e
f
o
r
T
UT
s
o
u
n
d
ev
en
ts
2
0
1
6
Fig
u
r
e
2
.
R
NN
ar
ch
itectu
r
e
f
o
r
T
UT
s
o
u
n
d
ev
en
ts
2
0
1
6
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
1
6
9
3
-
6
9
3
0
T
E
L
KOM
NI
KA
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l Co
n
tr
o
l
,
Vo
l.
18
,
No
.
5
,
Octo
b
e
r
2
0
2
0
:
2
5
8
7
-
259
6
2590
2
.
5
.
CR
NN
T
h
e
ar
ch
itectu
r
e
o
f
th
e
C
R
NN
u
s
ed
in
th
is
s
tu
d
y
f
o
r
T
UT
s
o
u
n
d
e
v
en
ts
2
0
1
6
is
s
h
o
wn
i
n
Fig
u
r
e
3
.
I
t
co
n
s
is
ts
o
f
co
n
v
o
lu
tio
n
al
lay
er
s
in
a
ca
s
ca
d
e
with
r
ec
u
r
r
en
t
lay
e
r
s
f
o
llo
wed
b
y
an
o
u
tp
u
t
lay
e
r
.
T
h
e
co
n
v
o
lu
tio
n
lay
e
r
s
ac
t
as
a
r
o
b
u
s
t
(
tim
e
-
an
d
f
r
eq
u
en
cy
-
in
v
ar
ian
t)
f
ea
t
u
r
e
ex
tr
ac
to
r
.
T
h
e
r
ec
u
r
r
e
n
t
lay
er
s
p
r
o
v
id
e
co
n
tex
t
u
al
in
f
o
r
m
atio
n
in
th
e
tim
e
d
o
m
ain
,
wh
ich
i
s
h
ig
h
ly
im
p
o
r
tan
t
f
o
r
r
ec
o
g
n
i
zin
g
s
o
u
n
d
ev
en
ts
.
Fin
ally
,
th
e
o
u
tp
u
t
lay
e
r
g
en
e
r
ates
th
e
ac
tiv
ity
p
r
o
b
ab
ilit
ies
f
o
r
th
e
s
o
u
n
d
ev
e
n
t
class
es
f
o
r
a
g
iv
e
n
f
r
am
e
.
T
h
e
p
a
r
am
eter
s
o
f
th
e
co
n
v
o
lu
tio
n
al,
r
ec
u
r
r
en
t,
an
d
f
ee
d
f
o
r
w
ar
d
lay
e
r
s
ar
e
o
p
tim
iz
ed
th
r
o
u
g
h
b
ac
k
p
r
o
p
ag
atio
n
.
Fig
u
r
e
1
.
C
R
NN
ar
ch
itectu
r
e
f
o
r
T
UT
s
o
u
n
d
ev
e
n
ts
2
0
1
6
3.
E
XP
E
R
I
M
E
N
T
S
3
.
1
.
D
a
t
a
ba
s
es
W
e
ev
alu
ated
th
e
d
ee
p
n
eu
r
al
n
etwo
r
k
s
o
n
two
d
atasets
.
On
e
was
ar
tific
ial
ly
g
en
er
ated
,
(
T
UT
s
o
u
n
d
ev
en
ts
s
y
n
th
etic
2
0
1
6
ab
b
r
ev
iated
as
T
UT
-
SED
s
y
n
th
etic
)
,
an
d
th
e
o
th
er
(
T
UT
s
o
u
n
d
ev
en
ts
2
0
1
6
)
was
r
ec
o
r
d
e
d
in
r
ea
l
en
v
ir
o
n
m
e
n
t
s
.
T
h
e
f
o
r
m
er
was
s
elec
ted
s
i
n
c
e
th
e
an
n
o
tatio
n
s
in
r
ea
l
au
d
io
d
ata
ar
e
r
ath
er
s
u
b
jectiv
e;
th
er
ef
o
r
e,
th
e
g
r
o
u
n
d
tr
u
th
lab
elin
g
m
ay
d
ep
e
n
d
ex
ce
s
s
iv
ely
o
n
th
e
an
n
o
ta
to
r
s
,
p
ar
ticu
lar
ly
in
th
e
p
r
esen
ce
o
f
p
o
l
y
p
h
o
n
ic
s
o
u
n
d
ev
en
ts
.
T
UT
-
SED
Sy
n
th
etic
was
g
en
e
r
ated
b
y
m
ix
in
g
is
o
lated
s
o
u
n
d
e
v
en
ts
f
r
o
m
1
6
d
if
f
e
r
en
t
class
es.
A
to
tal
1
0
0
m
ix
tu
r
es
wer
e
cr
ea
ted
f
r
o
m
9
9
4
s
o
u
n
d
s
am
p
les
an
d
d
iv
i
d
ed
in
to
tr
ain
in
g
,
test
in
g
an
d
v
alid
atio
n
d
ata,
with
p
r
o
p
o
r
tio
n
s
6
0
%,
2
0
%,
an
d
2
0
%,
r
esp
ec
tiv
ely
.
T
h
e
to
tal
len
g
th
o
f
th
e
m
ix
tu
r
e
d
ata
was 5
6
6
m
in
.
Seg
m
en
ts
o
f
len
g
th
3
-
1
5
s
wer
e
s
elec
ted
f
r
o
m
s
o
u
n
d
ev
e
n
t
in
s
tan
ce
s
to
co
n
s
titu
te
a
m
ix
tu
r
e,
an
d
th
er
e
wer
e
n
o
co
m
m
o
n
s
o
u
n
d
ev
e
n
t in
s
tan
ce
s
b
etwe
en
tr
ain
in
g
,
test
in
g
,
an
d
v
alid
ati
o
n
d
ata.
T
UT
s
o
u
n
d
e
v
en
ts
2
0
1
6
c
o
n
s
is
ts
o
f
r
ec
o
r
d
in
g
s
in
two
r
ea
l
e
n
v
ir
o
n
m
en
ts
:
r
esid
en
tial
an
d
h
o
m
e.
E
ac
h
r
ec
o
r
d
in
g
was
o
b
tain
ed
f
r
o
m
d
if
f
er
en
t
lo
ca
tio
n
s
to
en
s
u
r
e
l
ar
g
e
v
ar
iab
ilit
y
.
Au
d
io
s
am
p
le
s
with
th
e
len
g
th
o
f
3
-
5
m
i
n
wer
e
r
ec
o
r
d
e
d
at
ea
ch
lo
ca
tio
n
,
an
d
t
h
e
to
tal
len
g
th
o
f
th
e
au
d
io
s
am
p
les
is
7
8
m
in
.
A
to
tal
o
f
7
m
an
u
ally
a
n
n
o
tate
d
class
es
co
r
r
esp
o
n
d
t
o
th
e
r
esid
en
tial
en
v
ir
o
n
m
en
t,
wh
er
ea
s
1
1
a
n
n
o
tated
s
o
u
n
d
ev
en
t
class
es
co
r
r
esp
o
n
d
to
th
e
h
o
m
e
en
v
ir
o
n
m
en
t.
T
h
e
f
o
u
r
-
f
o
ld
cr
o
s
s
-
v
alid
atio
n
a
p
p
r
o
ac
h
was
ad
o
p
ted
in
th
e
tr
ain
in
g
an
d
test
in
g
p
r
o
ce
d
u
r
e
to
co
m
p
en
s
ate
f
o
r
th
e
s
m
a
ll
am
o
u
n
t
o
f
d
ata
in
t
h
is
d
ataset.
T
wen
ty
p
er
ce
n
t
o
f
th
e
tr
ain
i
n
g
d
ata
wer
e
allo
c
ated
as
v
alid
atio
n
d
ata
i
n
th
e
t
r
ain
in
g
p
h
ase.
T
UT
s
o
u
n
d
ev
e
n
ts
2
0
1
6
was
u
s
ed
in
th
e
DC
ASE
2
0
1
6
ch
allen
g
e,
wh
er
e
th
e
two
en
v
ir
o
n
m
en
t
s
wer
e
s
ep
ar
ately
ev
alu
ate
d
f
o
r
s
ce
n
e
-
d
e
p
en
d
e
n
t
class
if
icatio
n
.
I
n
th
is
s
tu
d
y
,
th
e
class
es
f
r
o
m
th
e
two
wer
e
n
o
t
d
is
tin
g
u
is
h
ed
,
r
esu
ltin
g
i
n
1
8
(
=7
+
1
1
)
s
o
u
n
d
ev
en
t
class
es
to
b
e
r
ec
o
g
n
ize
d
f
o
r
s
ce
n
e
-
in
d
ep
en
d
en
t
class
if
icatio
n
.
T
h
er
ef
o
r
e,
o
n
ly
o
n
e
cl
ass
if
ier
is
r
eq
u
ir
ed
,
r
ath
er
th
an
two
,
as wa
s
th
e
ca
s
e
in
th
e
D
C
ASE
2
0
1
6
c
h
allen
g
e.
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KOM
NI
KA
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l Co
n
tr
o
l
S
o
u
n
d
ev
en
t
d
etec
tio
n
u
s
in
g
d
ee
p
n
eu
r
a
l n
etw
o
r
ks
(
Suk
-
Hw
a
n
Ju
n
g
)
2591
3
.
2
.
E
v
a
lua
t
i
o
n
m
et
rics
E
v
alu
atio
n
m
eth
o
d
s
u
s
e
eith
er
s
eg
m
en
t
-
o
r
e
v
en
t
-
b
ased
m
et
r
ics
[
2
5
]
.
I
n
t
h
e
f
o
r
tm
er
,
th
e
e
v
alu
atio
n
o
f
a
d
ee
p
n
eu
r
al
n
etwo
r
k
f
o
r
s
o
u
n
d
ev
e
n
t
d
etec
tio
n
u
s
es
th
e
er
r
o
r
r
ate
an
d
F
-
s
co
r
e
in
a
f
ix
ed
tim
e
g
r
id
.
T
h
e
b
in
ar
ized
o
u
tp
u
ts
o
f
th
e
n
etwo
r
k
ar
e
co
m
p
ar
ed
with
th
e
g
r
o
u
n
d
tr
u
t
h
T
ab
le
in
1
s
s
eg
m
en
ts
.
A
s
o
u
n
d
ev
e
n
t
class
is
as
s
u
m
ed
to
b
e
d
etec
te
d
co
r
r
ec
tly
in
a
g
iv
e
n
s
eg
m
e
n
t
if
b
o
t
h
th
e
g
r
o
u
n
d
tr
u
th
tab
l
e
an
d
t
h
e
b
in
ar
ized
o
u
tp
u
t
co
r
r
esp
o
n
d
in
g
to
th
e
cl
ass
ar
e
ac
tiv
e
th
r
o
u
g
h
o
u
t
th
e
s
eg
m
en
t.
A
f
alse
p
o
s
itiv
e
im
p
lies
th
at
th
e
g
r
o
u
n
d
tr
u
th
tab
le
in
d
icate
s
th
at
a
s
o
u
n
d
ev
e
n
t
class
is
in
ac
tiv
e,
b
u
t
th
e
b
in
ar
ize
d
o
u
tp
u
t
is
ac
tiv
e
.
I
n
co
n
tr
ast,
a
f
alse
n
eg
ativ
e
im
p
lies
th
at
th
e
g
r
o
u
n
d
tr
u
t
h
tab
le
i
n
d
icate
s
th
at
th
e
clas
s
is
ac
tiv
e,
b
u
t
th
e
o
u
tp
u
t
is
in
ac
tiv
e.
Fin
ally
,
a
tr
u
e
p
o
s
itiv
e
im
p
lies
th
at
b
o
t
h
th
e
g
r
o
u
n
d
tr
u
th
ta
b
le
an
d
t
h
e
o
u
tp
u
t
in
d
icate
th
at
a
s
o
u
n
d
ev
en
t
class
is
ac
tiv
e.
F
-
s
co
r
e
is
ca
lcu
lated
as f
o
llo
ws
;
P
=
+
,
=
+
,
=
2
+
wh
er
e
T
P,
FP
an
d
FN
ar
e
co
u
n
ts
o
f
tr
u
e
p
o
s
itiv
es,
f
alse p
o
s
i
tiv
es,
an
d
f
alse n
eg
ativ
es,
r
esp
ec
tiv
ely
.
Fu
r
th
er
,
P
d
en
o
tes p
r
ec
is
io
n
,
a
n
d
R
is
r
ec
all.
An
o
th
er
m
etr
ic
is
t
h
e
er
r
o
r
r
ate
(
E
R
)
,
wh
ic
h
is
ca
lcu
late
d
in
te
r
m
s
o
f
in
s
er
tio
n
s
,
d
elet
io
n
s
,
an
d
s
u
b
s
titu
tio
n
s
.
A
s
u
b
s
titu
tio
n
e
r
r
o
r
o
cc
u
r
s
wh
en
th
e
b
in
ar
ize
d
o
u
tp
u
t
d
etec
ts
a
s
o
u
n
d
ev
en
t
class
in
a
s
eg
m
e
n
t,
b
u
t
th
e
lab
el
o
f
th
e
d
etec
ted
class
is
d
if
f
er
en
t
f
r
o
m
t
h
at
o
f
th
e
g
r
o
u
n
d
tr
u
th
tab
le.
A
s
u
b
s
titu
tio
n
er
r
o
r
im
p
lies
th
a
t
a
f
alse
p
o
s
itiv
e
an
d
a
f
alse
n
eg
ativ
e
o
cc
u
r
s
im
u
ltan
eo
u
s
ly
in
a
s
eg
m
en
t.
W
h
en
o
n
ly
f
a
ls
e
p
o
s
itiv
es
o
cc
u
r
in
a
s
eg
m
en
t,
th
ey
ar
e
co
u
n
ted
as
in
s
er
tio
n
s
,
an
d
wh
en
o
n
ly
f
a
ls
e
n
eg
ativ
es
o
cc
u
r
,
th
ey
ar
e
co
u
n
ted
as
d
eletio
n
s
.
T
h
e
E
R
is
ca
lcu
lated
as f
o
llo
ws
;
E
R
=
∑
(
)
=
1
+
∑
(
)
=
1
+
∑
(
)
=
1
∑
(
)
=
1
wh
er
e,
N
(
k
)
is
th
e
n
u
m
b
er
o
f
ac
tiv
e
g
r
o
u
n
d
tr
u
th
e
v
en
ts
in
a
s
eg
m
en
t
k
an
d
S
(
k
)
,
D
(
k
)
an
d
I
(
k
)
d
e
n
o
t
e
th
e
n
u
m
b
er
o
f
s
u
b
s
titu
tio
n
s
,
d
eletio
n
s
an
d
in
s
er
tio
n
s
,
r
esp
ec
t
iv
ely
.
K
is
th
e
to
tal
n
u
m
b
er
o
f
s
eg
m
en
t
s.
I
n
ev
en
t
-
b
ased
m
etr
ics,
a
s
o
u
n
d
e
v
en
t
is
ass
u
m
ed
to
b
e
co
r
r
ec
tly
d
etec
ted
if
t
h
e
b
in
ar
ize
d
o
u
tp
u
t
o
f
th
e
n
etwo
r
k
h
as
tim
e
-
in
ter
v
als
o
v
er
lap
p
in
g
with
th
o
s
e
o
f
th
e
co
r
r
ec
t
lab
el
in
th
e
g
r
o
u
n
d
tr
u
th
tab
le.
A
2
0
0
m
s
to
ler
an
ce
is
allo
wed
f
o
r
o
n
s
et
tim
e,
an
d
t
h
e
s
am
e
am
o
u
n
t
o
f
tim
e
(
2
0
0
m
s
)
o
r
5
0
%
o
f
th
e
d
u
r
atio
n
o
f
th
e
co
r
r
ec
t
lab
el
is
allo
wed
f
o
r
th
e
o
f
f
s
et
ti
m
e.
A
f
alse
p
o
s
itiv
e
o
cc
u
r
s
w
h
en
an
ac
tiv
e
b
in
ar
ized
o
u
tp
u
t
d
o
es
n
o
t
co
r
r
esp
o
n
d
to
th
e
co
r
r
ec
t
lab
el
in
th
e
g
r
o
u
n
d
tr
u
th
tab
le
with
in
th
e
allo
w
ed
to
ler
an
ce
.
I
f
a
s
o
u
n
d
ev
en
t in
th
e
g
r
o
u
n
d
tr
u
th
tab
le
d
o
es n
o
t c
o
r
r
esp
o
n
d
to
th
e
b
in
ar
ized
o
u
tp
u
t w
ith
th
e
s
am
e
lab
el,
a
f
alse n
eg
ativ
e
o
cc
u
r
s
.
3
.
3
.
Resul
t
s
W
e
ap
p
lied
b
atch
n
o
r
m
aliza
tio
n
af
ter
th
e
c
o
n
v
o
lu
tio
n
al
lay
e
r
s
an
d
a
d
r
o
p
o
u
t r
ate
o
f
0
.
2
5
was a
p
p
lied
to
th
e
co
n
v
o
lu
tio
n
al
an
d
r
ec
u
r
r
en
t
lay
e
r
s
.
W
e
tr
ain
e
d
th
e
n
et
wo
r
k
s
u
s
in
g
a
b
in
ar
y
c
r
o
s
s
-
en
tr
o
p
y
lo
s
s
f
u
n
ctio
n
with
th
e
Ad
am
o
p
tim
izer
.
E
ar
ly
s
to
p
p
in
g
was
u
s
ed
to
r
ed
u
ce
o
v
er
f
itti
n
g
.
T
h
e
tr
ain
in
g
was
s
to
p
p
ed
if
th
e
v
alu
e
o
f
th
e
lo
s
s
f
u
n
ctio
n
d
id
n
o
t
i
m
p
r
o
v
e
f
o
r
m
o
r
e
t
h
an
1
0
0
e
p
o
ch
s
.
As
th
e
p
er
f
o
r
m
an
ce
o
f
d
ee
p
n
eu
r
al
n
etwo
r
k
s
v
ar
ies
with
th
e
lear
n
in
g
r
ate,
we
attem
p
ted
to
s
elec
t
th
e
o
p
tim
al
lear
n
in
g
r
ate
f
o
r
all
n
etwo
r
k
s
b
y
test
in
g
th
eir
p
er
f
o
r
m
an
ce
o
n
th
e
v
alid
atio
n
d
ata.
T
h
e
p
e
r
f
o
r
m
an
ce
o
f
th
e
C
R
N
N
o
n
th
e
T
UT
-
SED
Sy
n
t
h
etic
as
th
e
lear
n
in
g
r
ate
ch
an
g
es is
s
h
o
wn
in
T
ab
l
e
1
.
T
ab
le
1
.
Per
f
o
r
m
an
ce
o
f
C
R
NN
o
n
T
UT
-
SED
Sy
n
th
etic
as l
ea
r
n
in
g
r
ate
ch
an
g
es
(
b
o
ld
f
a
ce
n
u
m
b
er
s
r
ep
r
esen
t th
e
b
est r
esu
lts
)
Le
a
r
n
i
n
g
r
a
t
e
V
a
l
i
d
a
t
i
o
n
d
a
t
a
Te
st
i
n
g
d
a
t
a
Ep
o
c
h
S
e
g
m
e
n
t
-
b
a
se
d
(F
-
sco
r
e
/
ER
)
Ev
e
n
t
-
b
a
s
e
d
(F
-
c
o
r
e
/
ER
)
S
e
g
m
e
n
t
-
b
a
se
d
(F
-
sco
r
e
/
ER
)
Ev
e
n
t
-
b
a
s
e
d
(F
-
sco
r
e
/
ER
)
10
−
3
6
1
.
6
9
%
/
0
.
5
2
3
7
.
6
9
%
/
0
.
9
6
6
0
.
6
1
%
/
0
.
5
3
3
7
.
0
5
%
/
0
.
9
7
16
10
−
4
6
8
.
7
5
%
/
0
.
4
5
4
3
.
4
9
%
/
0
.
8
8
6
4
.
2
1
%
/
0
.
5
0
4
0
.
5
0
%
/
0
.
9
6
33
10
−
5
6
6
.
4
4
%
/
0
.
4
9
3
9
.
1
0
%
/
0
.
9
6
6
3
.
7
6
%
/
0
.
5
2
3
6
.
4
8
%
/
1
.
0
4
1
5
7
10
−
6
4
4
.
1
6
%
/
0
.
6
9
9
.
8
3
%/
1
.
2
4
4
3
.
3
8
%
/
0
.
7
1
1
0
.
8
2
%
/
1
.
2
7
1
9
1
As
s
h
o
wn
in
T
ab
le
1
,
th
e
b
est
p
er
f
o
r
m
an
ce
is
o
b
tain
e
d
wh
en
th
e
lear
n
in
g
r
ate
is
10
−
4
f
o
r
all
co
n
d
itio
n
s
.
T
h
e
o
p
tim
al
lear
n
in
g
r
ate
f
o
r
th
e
v
alid
atio
n
d
ata
is
al
s
o
o
p
tim
al
f
o
r
th
e
test
in
g
d
ata.
Acc
o
r
in
g
ly
,
th
e
s
elec
tio
n
o
f
th
e
lea
r
n
in
g
r
at
e
b
ased
o
n
t
h
e
v
a
lid
atio
n
d
ata
i
s
q
u
ite
r
ea
s
o
n
ab
le.
Similar
p
e
r
f
o
r
m
an
ce
v
ar
iatio
n
s
with
th
e
lear
n
in
g
r
ate
co
u
ld
also
b
e
o
b
s
er
v
ed
f
o
r
th
e
FNN,
C
NN,
an
d
R
NN.
T
h
e
tab
le
s
h
o
ws
th
at
as
th
e
lear
n
in
g
r
ate
d
ec
r
ea
s
es,
th
e
n
u
m
b
er
o
f
ep
o
ch
s
f
o
r
wh
ich
we
o
b
tain
th
e
b
est
r
esu
lts
in
cr
ea
s
es.
T
h
is
is
d
u
e
to
th
e
s
lo
w
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
1
6
9
3
-
6
9
3
0
T
E
L
KOM
NI
KA
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l Co
n
tr
o
l
,
Vo
l.
18
,
No
.
5
,
Octo
b
e
r
2
0
2
0
:
2
5
8
7
-
259
6
2592
co
n
v
er
g
en
ce
o
f
th
e
weig
h
t
p
ar
am
eter
s
d
u
r
in
g
tr
ain
in
g
.
W
h
en
th
e
lear
n
in
g
r
ate
is
10
−
4
,
th
e
e
p
o
ch
n
u
m
b
er
is
3
3
,
wh
er
ea
s
it
is
1
9
1
wh
en
th
e
lear
n
in
g
r
ate
is
10
−
7
.
T
h
e
s
lo
w
co
n
v
er
g
en
ce
also
r
esu
lts
in
p
o
o
r
p
er
f
o
r
m
an
ce
,
wh
ich
is
r
el
ated
to
u
n
d
e
r
f
itti
n
g
.
T
h
e
v
ar
iatio
n
o
f
th
e
l
o
s
s
f
u
n
ctio
n
an
d
ac
cu
r
ac
y
at
th
e
o
u
tp
u
t
o
f
th
e
C
R
NN
d
u
r
in
g
tr
ai
n
in
g
wh
e
n
th
e
lear
n
in
g
r
ate
v
ar
ies
f
r
o
m
10
−
4
to
10
−
7
is
s
h
o
wn
in
Fig
u
r
e
4
.
W
h
en
th
e
lear
n
in
g
r
ate
is
10
−
4
,
th
e
lo
s
s
f
u
n
ctio
n
o
n
th
e
v
alid
atio
n
d
at
a
r
ea
ch
es
its
m
in
im
u
m
at
ap
p
r
o
x
im
ately
3
0
ep
o
c
h
s
(
ex
ac
t
ly
3
3
)
;
th
er
af
te
r
,
it
f
lu
ctu
ates
b
u
t
n
ev
er
d
r
o
p
s
b
el
o
w
th
e
m
in
im
u
m
.
Ho
wev
er
,
o
n
th
e
tr
ain
in
g
d
ata,
th
e
lo
s
s
f
u
n
ctio
n
co
n
tin
u
es
to
d
ec
r
ea
s
e
th
r
o
u
g
h
o
u
t
th
e
d
u
r
ati
o
n
o
f
t
h
e
tr
ain
in
g
(
we
s
et
th
e
m
ax
im
u
m
n
u
m
b
er
o
f
e
p
o
ch
s
to
2
0
0
)
.
As
o
v
er
f
itti
n
g
s
h
o
u
ld
b
e
av
o
i
d
ed
,
we
s
to
p
t
h
e
iter
atio
n
at
3
3
ep
o
c
h
s
u
s
in
g
th
e
af
o
r
e
m
en
tio
n
e
d
ea
r
ly
s
to
p
p
in
g
alg
o
r
ith
m
.
Me
an
wh
ile,
we
ca
n
o
b
s
er
v
e
q
u
ite
d
if
f
er
e
n
t
ch
ar
ac
ter
is
tics
wh
en
th
e
lea
r
n
in
g
r
ate
is
10
−
5
.
T
h
e
l
o
s
s
f
u
n
ctio
n
o
n
th
e
v
alid
atio
n
d
ata
d
ec
r
ea
s
es
f
o
r
a
s
ig
n
if
ican
tly
lo
n
g
er
p
er
i
o
d
a
n
d
r
ea
ch
es
its
m
in
im
u
m
at
1
5
7
.
T
h
e
lo
n
g
er
iter
atio
n
s
ca
u
s
e
p
e
r
f
o
r
m
an
ce
d
eg
r
ad
atio
n
o
n
b
o
th
th
e
v
alid
atio
n
an
d
test
in
g
d
ata
o
win
g
t
o
u
n
d
er
f
itti
n
g
.
T
h
is
p
h
en
o
m
en
o
n
b
ec
o
m
es m
o
r
e
m
an
if
est as we
f
u
r
th
er
d
ec
r
ea
s
e
th
e
lear
n
in
g
r
ate.
W
h
en
th
e
lea
r
n
in
g
r
ate
is
10
−
7
,
th
e
lo
s
s
f
u
n
ctio
n
d
o
es
n
o
t
r
ea
c
h
its
m
in
im
u
m
u
n
til
th
e
en
d
o
f
th
e
t
r
ain
in
g
.
A
s
im
ilar
tr
en
d
is
o
b
s
er
v
e
d
w
h
en
we
m
o
n
ito
r
th
e
ac
cu
r
ac
y
in
s
tead
o
f
th
e
lo
s
s
f
u
n
ctio
n
.
Fig
u
r
e
4
.
Var
iatio
n
o
f
l
o
s
s
f
u
n
ctio
n
an
d
ac
c
u
r
ac
y
with
lea
r
n
i
n
g
r
ate
(
C
R
NN
o
n
T
UT
-
SED
Sy
n
th
etic
d
ataset)
T
h
e
p
er
f
o
r
m
an
ce
co
m
p
ar
is
o
n
b
etwe
en
th
e
FNN,
C
N
N,
R
N
N
,
an
d
C
R
NN
o
n
th
e
T
UT
-
SED
Sy
n
th
etic
d
ataset
is
s
h
o
wn
in
T
ab
le
2
.
T
h
e
lear
n
in
g
r
ate
is
s
et
to
10
−
4
in
all
n
etwo
r
k
s
ac
co
r
d
in
g
t
o
th
e
p
r
e
v
io
u
s
ex
p
er
im
en
ts
.
T
h
e
e
v
alu
taio
n
i
s
r
ep
r
esen
ted
in
b
o
th
th
e
s
eg
m
en
t
-
an
d
e
v
en
t
-
b
ased
m
eth
o
d
s
.
As
p
r
esen
ted
in
T
ab
le
2
,
th
e
C
R
NN
p
er
f
o
r
m
s
b
est
u
n
d
er
all
test
i
n
g
c
o
n
d
itio
n
s
.
Alth
o
u
g
h
th
e
C
NN
h
as
ex
h
ib
te
d
q
u
ite
s
atis
f
ac
to
r
y
p
er
f
o
r
m
a
n
ce
in
im
ag
e
class
if
icatio
n
,
it
i
s
in
f
er
io
r
to
th
e
C
R
NN
in
s
o
u
n
d
ev
en
t
d
etec
tio
n
b
ec
au
s
e
C
NNs
lack
th
e
ab
ilit
y
to
m
o
d
el
th
e
tim
e
c
o
r
r
elatio
n
in
f
o
r
m
atio
n
(
p
a
r
ticu
lar
ly
l
o
n
g
-
ter
m
c
o
r
r
ela
tio
n
)
in
a
u
d
i
o
s
ig
n
als.
Ho
wev
er
,
it
ca
n
b
e
o
b
s
er
v
ed
th
at
u
s
in
g
t
h
e
R
NN
alo
n
e
co
u
ld
n
o
t
r
esu
lt
in
im
p
r
o
v
ed
p
er
f
o
r
m
an
ce
s
co
m
p
ar
ed
with
t
h
e
C
NN.
T
h
is
im
p
lies
th
at
t
h
e
tim
e
-
f
r
eq
u
e
n
cy
in
v
a
r
ian
t
f
ea
tu
r
e
ex
tr
ac
tio
n
b
y
C
NNs
is
h
ig
h
ly
im
p
o
r
tan
t in
s
o
u
n
d
ev
en
t
d
etec
tio
n
as is
e
ca
s
e
with
in
v
ar
ian
t f
ea
tu
r
es in
im
ag
e
class
if
icatio
n
.
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KOM
NI
KA
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l Co
n
tr
o
l
S
o
u
n
d
ev
en
t
d
etec
tio
n
u
s
in
g
d
ee
p
n
eu
r
a
l n
etw
o
r
ks
(
Suk
-
Hw
a
n
Ju
n
g
)
2593
T
ab
le
2
.
Per
f
o
r
m
an
ce
co
m
p
a
r
is
o
n
b
etwe
en
FNN,
C
NN,
R
N
N
an
d
C
R
NN
o
n
T
UT
-
SED
s
y
n
th
etic
S
e
g
m
e
n
t
-
b
a
se
d
Ev
e
n
t
-
b
a
s
e
d
F
-
sco
r
e
ER
F
-
sco
r
e
ER
F
N
N
5
4
.
5
7
%
0
.
8
2
1
.
4
5
%
3
.
5
1
C
N
N
6
0
.
3
8
%
0
.
6
6
3
1
.
3
1
%
1
.
8
7
R
N
N
4
7
.
2
8
%
0
.
6
6
2
8
.
9
7
%
1
.
3
2
C
R
N
N
6
4
.
2
1
%
0
.
5
4
0
.
5
0
%
0
.
9
6
I
n
ad
d
i
t
io
n
to
T
U
T
-
S
E
D
Sy
n
t
h
e
t
i
c
,
w
e
in
v
e
s
t
ig
a
t
ed
t
h
e
p
er
f
o
r
m
an
c
e
o
f
t
h
e
n
e
t
wo
r
k
s
u
s
i
n
g
T
U
T
s
o
u
n
d
e
v
e
n
t
s
2
0
1
6
.
T
h
e
p
e
r
f
o
r
m
a
n
c
e
o
f
t
h
e
C
R
N
N
o
n
T
U
T
s
o
u
n
d
e
v
e
n
t
s
2
0
1
6
f
o
r
v
a
r
y
in
g
l
e
ar
n
in
g
r
a
t
e
s
i
s
s
h
o
w
n
i
n
T
a
b
le
3
.
C
o
n
t
r
ar
y
t
o
t
h
e
r
e
s
u
l
t
o
n
th
e
T
UT
-
S
E
D
S
y
n
t
h
e
t
i
c
d
a
t
a
s
e
t
in
T
a
b
l
e
1
,
we
c
a
n
c
o
n
c
lu
d
e
t
h
a
t
t
h
e
b
e
s
t
p
e
r
f
o
r
m
a
n
c
e
i
s
n
o
w
o
b
t
a
i
n
ed
w
i
t
h
a
le
a
r
n
i
n
g
r
a
t
e
o
f
10
−
3
e
x
c
e
p
t
wh
en
t
h
e
ev
en
t
-
b
a
s
e
d
p
e
r
f
o
r
m
a
n
c
e
i
s
m
e
a
s
u
r
e
d
o
n
t
h
e
t
e
s
t
i
n
g
d
a
ta
.
T
h
i
s
im
p
l
i
e
s
t
h
a
t
th
e
o
p
t
im
a
l
l
e
ar
n
in
g
r
a
te
v
a
r
i
e
s
d
e
p
en
d
i
n
g
o
n
th
e
t
r
a
in
i
n
g
d
a
t
a
s
e
t
.
M
o
r
e
o
v
er
,
t
h
e
p
e
r
f
o
r
m
a
n
c
e
o
f
t
h
e
C
R
N
N
o
n
th
e
T
U
T
s
o
u
n
d
ev
en
t
s
2
0
1
6
d
a
ta
s
e
t w
a
s
n
o
t a
s
g
o
o
d
a
s
t
h
a
t
o
n
T
U
T
-
S
E
D
S
y
n
t
h
e
t
i
c.
T
h
i
s
m
ay
b
e
d
u
e
t
o
t
h
e
s
m
a
l
l
am
o
u
n
t
o
f
tr
a
in
i
n
g
d
a
t
a
an
d
t
h
e
s
c
e
n
e
-
i
n
d
e
p
en
d
e
n
t
c
l
a
s
s
i
f
i
c
a
t
i
o
n
in
T
UT
s
o
u
n
d
ev
e
n
t
s
2
0
1
6
.
A
l
th
o
u
g
h
th
e
n
u
m
b
e
r
o
f
w
e
ig
h
t
p
a
r
a
m
e
t
e
r
s
o
f
t
h
e
C
R
N
N
o
n
T
UT
s
o
u
n
d
ev
en
t
s
2
0
1
6
w
a
s
r
ed
u
c
e
d
b
y
a
p
p
r
o
x
i
m
a
te
l
y
2
0
%
c
o
m
p
ar
e
d
w
i
t
h
t
h
a
t
o
n
T
UT
-
S
E
D
Sy
n
th
e
t
i
c
,
t
h
e
p
e
r
f
o
r
m
a
n
c
e
d
e
g
r
a
d
a
t
i
o
n
c
o
u
ld
n
o
t
b
e
a
l
l
e
v
i
a
te
d
.
T
ab
le
3
.
Per
f
o
r
m
an
ce
o
f
C
R
NN
o
n
T
UT
s
o
u
n
d
ev
e
n
ts
2
0
1
6
as lea
r
n
in
g
r
ate
v
a
r
ies
(
b
o
ld
f
ac
e
n
u
m
b
er
s
r
ep
r
esen
t th
e
b
est r
esu
lts
)
L
e
a
n
i
n
g
r
a
t
e
V
a
l
i
d
a
t
i
o
n
d
a
t
a
Te
st
i
n
g
d
a
t
a
e
p
o
c
h
S
e
g
m
e
n
t
-
b
a
se
d
(F
-
sco
r
e
/
ER
)
Ev
e
n
t
-
b
a
s
e
d
(F
-
sco
r
e
/
ER
)
S
e
g
m
e
n
t
-
b
a
se
d
(F
-
sco
r
e
/
ER
)
Ev
e
n
t
-
b
a
s
e
d
(F
-
sco
r
e
/
ER
)
10
−
3
5
8
.
6
2
%
/
0
.
7
0
5
.
5
4
%/
4
.
9
9
3
7
.
1
8
%
/
0
.
9
0
6
.
5
8
%/
3
.
0
8
2
5
.
5
10
−
4
5
0
.
7
5
%
/
0
.
7
3
5
.
3
2
%/
4
.
1
1
3
6
.
4
1
%
/
0
.
8
8
7
.
8
1
%/
3
.
0
6
8
7
.
5
10
−
5
1
.
4
4
%
/
0
.
9
9
0
.
0
0
%/
1
.
0
5
0
.
2
0
%
/
1
.
0
0
0
.
0
0
%/
1
.
0
0
1
8
7
.
8
10
−
6
9
.
5
9
%
/
1
0
.
3
4
1
.
1
3
%/
8
2
.
5
5
5
.
8
8
%
/
1
.
0
8
1
.
1
5
%/
1
.
8
5
1
9
9
.
8
T
h
e
v
a
r
ia
t
i
o
n
o
f
th
e
lo
s
s
f
u
n
c
t
i
o
n
a
n
d
a
cc
u
r
a
c
y
a
t
th
e
o
u
tp
u
t
o
f
t
h
e
C
R
N
N
d
u
r
i
n
g
t
r
a
in
i
n
g
o
n
th
e
T
U
T
s
o
u
n
d
e
v
e
n
t
s
2
0
1
6
d
a
t
as
e
t
(
a
s
t
h
e
l
e
a
r
n
i
n
g
r
a
t
e
v
a
r
i
es
f
r
o
m
10
−
3
t
o
10
−
6
)
i
s
s
h
o
wn
in
F
i
g
u
r
e
5
.
A
s
i
m
i
l
a
r
t
r
en
d
t
o
th
a
t
in
F
i
g
u
r
e
4
c
an
b
e
o
b
s
e
r
v
e
d
.
H
o
w
ev
e
r
,
w
i
t
h
t
h
e
s
a
m
e
l
ea
r
n
in
g
r
a
t
e,
th
e
l
o
s
s
f
u
n
c
t
i
o
n
o
n
t
r
a
i
n
in
g
d
a
t
a
c
o
n
v
er
g
e
s
f
a
s
t
er
o
n
th
e
s
o
u
n
d
ev
e
n
t
s
2
0
1
6
d
a
t
a
s
e
t
th
a
n
o
n
t
h
e
T
U
T
-
S
E
D
S
y
n
th
e
t
i
c
d
a
t
a
s
e
t
.
T
h
i
s
i
m
p
l
i
e
s
t
h
a
t
a
s
m
a
l
l
e
r
lea
r
n
in
g
r
a
t
e
i
s
d
e
s
i
r
ab
l
e
f
o
r
th
e
f
o
r
m
er
to
p
r
e
v
e
n
t
o
v
e
r
f
i
t
t
in
g
.
T
h
i
s
i
s
r
e
f
l
e
c
t
ed
i
n
t
h
e
p
er
f
o
r
m
an
c
e
s
co
r
e
s
in
T
a
b
l
e
3
,
wh
er
e
t
h
e
b
e
s
t
s
co
r
e
s
a
r
e
o
b
ta
i
n
ed
w
h
e
n
th
e
l
ea
r
n
in
g
r
a
t
e
i
s
10
−
3
e
x
c
ep
t
f
o
r
o
n
e
c
a
s
e.
T
h
e
p
e
r
f
o
r
m
a
n
c
e
c
o
m
p
ar
i
s
o
n
b
e
t
w
e
en
F
N
N
,
C
N
N
,
R
N
N
,
a
n
d
C
R
N
N
o
n
th
e
s
o
u
n
d
ev
e
n
t
s
2
0
1
6
d
a
t
a
s
e
t
i
s
s
h
o
wn
in
T
a
b
le
4
.
T
h
e
l
e
ar
n
in
g
r
a
te
i
s
s
e
t
t
o
10
−
3
i
n
a
l
l
n
e
t
wo
r
k
s
ac
c
o
r
d
in
g
t
o
t
h
e
r
e
s
u
l
t
s
i
n
T
a
b
l
e
3
.
T
ab
l
e
4
d
em
o
n
s
t
r
a
t
e
t
h
a
t
t
h
e
C
R
N
N
p
er
f
o
r
m
s
b
e
s
t
i
n
t
e
r
m
s
o
f
th
e
s
e
g
m
e
n
t
-
b
a
s
ed
F
-
s
c
o
r
e
an
d
e
r
r
o
r
r
a
t
e
.
T
h
e
C
R
N
N
i
s
f
o
l
lo
w
e
d
b
y
C
N
N
,
an
d
th
e
R
N
N
i
s
t
h
e
w
o
r
s
t
.
T
h
i
s
i
s
i
n
a
c
co
r
d
an
c
e
w
i
t
h
t
h
e
r
e
s
u
l
t
s
o
n
t
h
e
T
UT
-
S
E
D
S
y
n
th
e
t
i
c
d
a
ta
s
e
t
s
h
o
wn
in
T
a
b
l
e
2
.
Ho
w
e
v
e
r
,
r
eg
a
r
d
i
n
g
th
e
e
v
en
t
-
b
a
s
e
d
m
e
t
r
i
c
s
,
u
n
ex
p
e
c
t
ed
r
e
s
u
l
t
s
c
an
b
e
o
b
s
e
r
v
ed
.
Ne
v
er
t
h
e
l
e
s
s
,
th
e
l
o
w
F
-
s
c
o
r
e
an
d
e
r
r
o
r
r
a
t
e
in
T
a
b
l
e
4
i
m
p
ly
th
a
t
t
h
e
s
e
r
e
s
u
l
t
s
a
r
e
n
o
t
c
r
ed
i
b
l
e
a
n
d
an
d
m
ay
t
h
e
r
e
f
o
r
e
b
e
ig
n
o
r
ed
.
T
ab
le
4
.
Per
f
o
r
m
an
ce
co
m
p
a
r
is
o
n
b
etwe
en
FNN,
C
NN,
R
N
N
an
d
C
R
NN
on
th
e
s
o
u
n
d
e
v
en
ts
2
0
1
6
d
ata
s
et
S
e
g
m
e
n
t
-
b
a
se
d
Ev
e
n
t
-
b
a
s
e
d
F
-
sco
r
e
ER
F
-
sco
r
e
ER
F
N
N
2
5
.
1
1
%
1
.
3
2
2
.
4
2
%
9
.
8
1
C
N
N
3
5
.
2
8
%
0
.
9
8
7
.
5
4
%
4
.
4
9
R
N
N
2
4
.
3
4
%
1
.
0
2
4
.
2
8
%
2
.
4
8
C
R
N
N
3
7
.
1
8
%
0
.
9
0
6
.
5
8
%
3
.
0
8
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
1
6
9
3
-
6
9
3
0
T
E
L
KOM
NI
KA
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l Co
n
tr
o
l
,
Vo
l.
18
,
No
.
5
,
Octo
b
e
r
2
0
2
0
:
2
5
8
7
-
259
6
2594
Fig
u
r
e
5
.
Var
iatio
n
o
f
l
o
s
s
f
u
n
ctio
n
an
d
ac
c
u
r
ac
y
as lea
r
n
in
g
r
ate
ch
an
g
es
(
C
R
NN
o
n
s
o
u
n
d
ev
en
ts
2
0
1
6
)
4
.
CO
NCLUS
I
O
N
S
D
e
e
p
n
e
u
r
a
l
n
e
t
w
o
r
k
s
h
av
e
b
e
e
n
w
i
d
e
ly
u
s
e
d
in
v
a
r
io
u
s
a
r
e
a
s
o
f
p
a
t
t
er
n
r
e
co
g
n
i
t
i
o
n
.
R
e
c
e
n
t
l
y
,
i
n
s
o
u
n
d
ev
e
n
t
d
e
t
e
c
t
io
n
,
n
u
m
e
r
o
u
s
a
p
p
r
o
ac
h
e
s
b
a
s
e
d
o
n
d
ee
p
n
eu
r
a
l
n
e
t
w
o
r
k
s
h
av
e
b
e
e
n
p
r
o
p
o
s
ed
a
n
d
h
av
e
e
x
h
ib
t
e
d
s
u
p
er
i
o
r
p
e
r
f
o
r
m
an
c
e
t
o
a
m
o
n
g
o
t
h
e
r
co
n
v
e
n
t
io
n
a
l
m
e
th
o
d
s
,
s
u
c
h
a
s
G
M
M
s
a
n
d
S
V
M
s
.
I
n
t
h
i
s
p
a
p
er
,
we
p
r
o
p
o
s
ed
th
e
ap
p
l
ica
t
i
o
n
o
f
r
ep
r
e
s
e
n
t
a
t
iv
e
d
e
ep
n
e
u
r
a
l n
e
t
wo
r
k
s
t
o
t
h
e
s
o
u
n
d
ev
e
n
t d
e
t
ec
t
i
o
n
.
W
e
a
p
p
l
i
ed
an
F
N
N
,
a
C
N
N
,
an
R
N
N
,
a
n
d
a
C
R
N
N
t
o
t
wo
in
d
e
p
en
d
e
n
t
d
a
t
a
s
e
t
s
f
o
r
s
o
u
n
d
e
v
e
n
t
d
e
te
c
t
i
o
n
.
T
h
e
r
e
s
u
l
t
d
e
m
o
n
s
t
r
a
t
e
d
t
h
a
t
t
h
e
p
e
r
f
o
r
m
an
c
e
s
o
f
t
h
e
s
e
n
e
t
wo
r
k
s
v
a
r
i
ed
s
ig
n
i
f
i
c
a
n
t
ly
d
e
p
e
n
d
in
g
o
n
t
h
e
l
e
ar
n
in
g
r
a
t
e
.
T
h
e
o
p
t
im
a
l
le
a
r
n
i
n
g
r
at
e
w
a
s
s
e
l
ec
t
e
d
b
a
s
ed
o
n
th
e
l
o
s
s
f
u
n
c
t
i
o
n
o
n
t
h
e
v
a
l
i
d
a
ti
o
n
d
a
t
a
;
th
i
s
w
a
s
c
o
n
f
i
r
m
e
d
t
o
b
e
q
u
i
t
e
r
e
a
s
o
n
a
b
l
e
b
a
s
ed
o
n
ex
t
e
n
s
iv
e
e
x
p
e
r
i
m
en
t
a
l
r
e
s
u
l
t
s
o
n
th
e
t
e
s
t
in
g
d
a
ta
.
A
s
m
a
l
l
l
e
a
r
n
in
g
r
a
t
e
te
n
d
s
to
u
n
d
e
r
f
i
t
t
h
e
n
e
t
w
o
r
k
s
t
o
t
h
e
tr
a
i
n
in
g
d
a
t
a
,
w
h
e
r
e
a
s
a
n
ex
c
e
s
s
i
v
e
ly
l
a
r
g
e
l
e
ar
n
in
g
r
a
t
e
r
e
s
u
l
t
s
i
n
o
v
e
r
f
i
t
t
i
n
g
.
I
t
w
a
s
a
l
s
o
d
em
o
n
s
t
r
a
t
ed
t
h
at
t
h
e
a
m
o
u
n
t
o
f
tr
a
in
i
n
g
d
a
ta
a
n
d
th
e
t
y
p
e
o
f
c
l
a
s
s
e
s
c
o
n
s
i
d
e
r
a
b
ly
a
f
f
e
c
te
d
t
h
e
p
er
f
o
r
m
a
n
ce
o
f
th
e
n
e
t
w
o
r
k
s
.
T
h
e
p
e
r
f
o
r
m
an
ce
o
n
T
UT
-
S
E
D
S
y
n
t
h
e
t
i
c
w
a
s
s
i
g
n
if
i
c
an
t
l
y
b
e
t
t
er
t
h
a
n
t
h
a
t
o
n
s
o
u
n
d
e
v
e
n
t
s
2
0
1
6
,
t
h
e
s
i
z
e
o
f
w
h
i
ch
i
s
ap
p
r
o
x
im
a
t
e
l
y
o
n
e
s
e
v
e
n
th
th
a
t o
f
T
U
T
-
S
E
D
Sy
n
th
e
t
i
c,
w
h
i
c
h
c
o
n
t
a
i
n
s
a
u
d
i
o
c
l
a
s
s
e
s
t
h
a
t
ar
e
d
i
f
f
i
c
u
l
t
t
o
d
i
s
t
in
g
u
i
s
h
.
A
l
th
o
u
g
h
th
e
n
u
m
b
er
o
f
we
i
g
h
t
p
ar
a
m
e
te
r
s
o
f
t
h
e
n
e
t
wo
r
k
s
o
n
s
o
u
n
d
e
v
en
ts
2
0
1
6
w
a
s
r
e
d
u
c
ed
b
y
2
0
%
t
o
c
o
m
p
en
s
a
t
e
f
o
r
t
h
e
s
m
a
ll
a
m
o
u
n
t
o
f
t
r
a
in
i
n
g
d
a
t
a
,
t
h
e
p
er
f
o
r
m
an
c
e
g
ap
w
as
q
u
i
t
e
l
a
r
g
e
.
F
i
n
a
l
l
y
,
th
e
C
R
N
N
o
u
tp
e
r
f
o
r
m
e
d
th
e
o
th
e
r
n
e
t
w
o
r
k
s
,
a
m
o
n
g
w
h
ic
h
th
e
C
N
N
w
a
s
t
h
e
s
ec
o
n
d
m
o
s
t
e
f
f
e
c
t
iv
e
.
T
h
e
F
N
N
an
d
R
N
N
p
e
r
f
o
r
m
e
d
w
o
r
s
e
t
h
a
n
t
h
e
C
R
N
N
a
n
d
C
N
N
.
T
h
e
p
o
o
r
p
er
f
o
r
m
a
n
c
e
o
f
th
e
R
N
N
i
m
p
l
i
e
s
t
h
a
t
t
i
m
e
-
f
r
eq
u
en
c
y
in
v
ar
i
a
n
t
f
e
a
t
u
r
e
s
f
r
o
m
t
h
e
C
NN
a
r
e
h
i
g
h
l
y
im
p
o
r
t
an
t
in
s
o
u
n
d
e
v
e
n
t
d
e
t
e
c
t
io
n
.
I
n
f
u
t
u
r
e
wo
r
k
,
w
e
w
i
l
l
s
tu
d
y
a
v
a
r
ia
n
t
o
f
th
e
C
R
N
N
a
r
c
h
i
tec
t
u
r
e
th
a
t
ca
n
u
s
e
t
h
e
c
h
ar
a
c
t
er
i
s
t
i
c
s
o
f
th
e
R
N
N
m
o
r
e
ef
f
ec
t
i
v
e
ly
b
y
c
o
n
s
i
d
er
in
g
d
if
f
er
e
n
t
m
e
t
h
o
d
s
o
f
c
o
u
p
l
in
g
w
i
t
h
th
e
C
N
N
.
I
n
ad
d
i
t
io
n
,
t
h
e
s
e
g
m
en
t
l
en
g
th
i
n
t
h
e
C
R
N
N
s
h
o
u
l
d
b
e
o
p
t
im
i
z
e
d
to
a
ch
i
ev
e
th
e
ap
p
r
o
p
r
i
a
te
m
e
m
o
r
y
l
en
g
th
in
t
h
e
G
R
Us
.
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KOM
NI
KA
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l Co
n
tr
o
l
S
o
u
n
d
ev
en
t
d
etec
tio
n
u
s
in
g
d
ee
p
n
eu
r
a
l n
etw
o
r
ks
(
Suk
-
Hw
a
n
Ju
n
g
)
2595
ACK
NO
WL
E
DG
M
E
N
T
S
T
h
is
r
es
ea
r
ch
was
s
u
p
p
o
r
ted
b
y
B
asic
Scien
ce
R
esear
ch
Pro
g
r
am
th
r
o
u
g
h
th
e
Natio
n
al
R
esear
ch
Fo
u
n
d
atio
n
o
f
Ko
r
ea
(
NR
F)
f
u
n
d
ed
b
y
Min
is
tr
y
o
f
E
d
u
ca
tio
n
(
No
.
2
0
1
8
R
1
A2
B
6
0
0
9
3
2
8
)
.
RE
F
E
R
E
NC
E
S
[
1
]
M
.
K.
Na
n
d
wa
n
a
,
A.
Zi
a
e
i,
a
n
d
J.
H.
L.
Ha
n
se
n
,
“
Ro
b
u
st
U
n
su
p
e
rv
ise
d
De
tec
ti
o
n
o
f
Hu
m
a
n
S
c
r
e
a
m
s
In
No
isy
Ac
o
u
stic E
n
v
ir
o
n
m
e
n
ts
,”
Pro
c
e
e
d
in
g
s
o
f
th
e
I
n
ter
n
a
ti
o
n
a
l
C
o
n
fer
e
n
c
e
o
n
Aco
u
stics
,
S
p
e
e
c
h
a
n
d
S
i
g
n
a
l
Pr
o
c
e
ss
in
g
.
Brisb
a
n
e
,
p
p
.
1
6
1
-
1
6
5
,
2
0
1
5
.
[2
]
M
.
Cro
c
c
o
,
M
.
C
h
ristan
i,
A.
Tru
c
c
o
,
a
n
d
V.
M
u
r
in
o
,
“
Au
d
io
S
u
r
v
e
il
lan
c
e
:
A
S
y
ste
m
a
ti
c
Re
v
iew
,
”
ACM
Co
mp
u
ti
n
g
S
u
rv
e
y
s,
v
o
l
.
4
8
,
n
o
.
4
,
p
p
.
1
-
4
6
,
2
0
1
6
.
[3
]
J.
S
a
lam
o
n
a
n
d
J.
P
.
Be
ll
o
,
“
F
e
a
tu
re
Lea
rn
in
g
wit
h
De
e
p
S
c
a
tt
e
rin
g
fo
r
Urb
a
n
S
o
u
n
d
An
a
l
y
sis,”
23
rd
Eu
ro
p
e
a
n
S
i
g
n
a
l
Pro
c
e
ss
in
g
Co
n
fer
e
n
c
e
(EUS
IPC
O)
,
p
p
.
7
2
4
-
7
2
8
,
2
0
1
5
.
[4
]
Y.
Wan
g
,
L
.
Ne
v
e
s,
a
n
d
F
.
M
e
tze
,
“
Au
d
io
-
b
a
se
d
M
u
lt
ime
d
ia
Ev
e
n
t
De
tec
ti
o
n
Us
i
n
g
De
e
p
R
e
c
u
rre
n
t
Ne
u
ra
l
Ne
two
rk
s,”
IEE
E
I
n
ter
n
a
ti
o
n
a
l
Co
n
fer
e
n
c
e
o
n
Aco
u
stics
,
S
p
e
e
c
h
a
n
d
S
i
g
n
a
l
Pro
c
e
ss
in
g
(ICA
S
S
P).
S
h
a
n
g
h
a
i,
p
p
.
2
7
4
2
-
2
7
4
6
,
2
0
1
6
.
[5
]
F
.
Brig
g
s,
B.
Lak
sh
m
i
n
a
ra
y
a
n
a
n
,
L.
Ne
a
l,
X.
Z.
F
e
rn
,
R.
Ra
ich
,
S
.
J.
K.
Ha
d
ley
,
A.
S
.
Ha
d
le
y
,
a
n
d
M
.
G
.
Be
tt
s,
“
Ac
o
u
stic
Clas
sifica
ti
o
n
o
f
M
u
lt
ip
le
S
im
u
lt
a
n
e
o
u
s
Bir
d
S
p
e
c
i
e
s:
A
M
u
lt
i
-
i
n
sta
n
c
e
M
u
lt
i
-
Lab
e
l
Ap
p
ro
a
c
h
,
”
T
h
e
J
o
u
r
n
a
l
o
f
th
e
Ac
o
u
stica
l
S
o
c
iety
o
f
Ame
ric
a
,
v
o
l
.
1
3
1
,
n
o
.
6
,
p
p
.
4
6
4
0
-
4
6
4
0
,
2
0
1
2
.
[6
]
S
.
Nta
lam
p
iras
,
e
t
a
l.
,
"
On
Ac
o
u
stic
S
u
rv
e
i
ll
a
n
c
e
o
f
Ha
z
a
rd
o
u
s
S
it
u
a
ti
o
n
s
,
"
IE
EE
In
ter
n
a
ti
o
n
a
l
Co
n
fer
e
n
c
e
o
n
Aco
u
stics
,
S
p
e
e
c
h
a
n
d
S
i
g
n
a
l
Pr
o
c
e
ss
in
g
(ICAS
S
P),
p
p
.
1
6
5
-
1
6
8
,
2
0
0
9
.
[7
]
A.
M
e
sa
ro
s,
T.
He
it
to
la,
a
n
d
T
.
Virtan
e
n
,
“
TUT
Da
tab
a
se
fo
r
Ac
o
u
stic
S
c
e
n
e
Clas
sifica
ti
o
n
a
n
d
S
o
u
n
d
Ev
e
n
t
De
tec
ti
o
n
,”
24
th
Eu
r
o
p
e
a
n
S
ig
n
a
l
Pro
c
e
ss
in
g
C
o
n
fer
e
n
c
e
(EUS
IPC
O).
Bu
d
a
p
e
st,
p
p
.
1
1
2
8
-
1
1
3
2
,
2
0
1
6
.
[8
]
D.
S
to
we
ll
,
D.
G
ian
n
o
u
li
s,
E.
Be
n
e
to
s,
M
.
Lag
ra
n
g
e
,
a
n
d
M
.
D.
P
lu
m
b
ley
,
“
De
tec
ti
o
n
a
n
d
Clas
sifica
ti
o
n
o
f
Ac
o
u
stic
S
c
e
n
e
s a
n
d
E
v
e
n
ts,”
IEE
E
T
ra
n
s.
On
M
u
lt
ime
d
i
a
,
vol
.
1
7
,
n
o
.
1
0
,
p
p
.
1
7
3
3
-
1
7
4
6
,
2
0
1
5
.
[9
]
A.
M
e
sa
ro
s,
T.
He
it
to
la,
A.
Dim
e
n
t,
B
.
E
li
z
a
ld
e
,
A.
S
h
a
h
,
E.
Vi
n
c
e
n
t,
B.
Ra
j
,
a
n
d
T.
Virtan
e
n
,
“
DCA
S
E
2
0
1
7
Ch
a
ll
e
n
e
g
S
e
tu
p
;
Tas
k
s,
Da
tas
e
ts
a
n
d
Ba
se
li
n
e
S
y
ste
m
,
”
Pro
c
e
e
d
in
g
s
o
f
t
h
e
De
tec
ti
o
n
a
n
d
Cla
ss
if
ic
a
ti
o
n
o
f
Aco
u
stic
S
c
e
n
e
s a
n
d
Eve
n
ts
2
0
1
7
W
o
rk
sh
o
p
(DCAS
E
2
0
1
7
).
M
u
n
ich
,
p
p
.
1
1
2
3
-
1
1
2
7
,
2
0
1
7
.
[1
0
]
G
.
De
k
k
e
rs,
e
t
a
l.
,
“
DCA
S
E
2
0
1
8
c
h
a
ll
e
n
g
e
-
Tas
k
5
:
M
o
n
it
o
ri
n
g
o
f
Do
m
e
stic
Ac
ti
v
it
ies
b
a
se
d
o
n
M
u
lt
i
-
c
h
a
n
n
e
l
Ac
o
u
stics
,
”
KU
L
e
u
v
e
n
,
T
e
c
h
.
Re
p
.
,
J
u
ly
2
0
1
8
.
[1
1
]
N.
Tu
rp
a
u
lt
,
R.
S
e
rize
l,
A.
S
h
a
h
a
n
d
J.
S
a
lam
o
n
,
“
S
o
u
n
d
Ev
e
n
t
De
tec
ti
o
n
in
D
o
m
e
stic
En
v
ir
o
n
e
m
e
n
t
s
with
Wea
k
ly
Lab
e
led
Da
ta
a
n
d
S
o
u
n
d
sc
a
p
e
S
y
n
th
e
sis,”
W
o
rk
sh
o
p
o
n
De
tec
ti
o
n
a
n
d
Cl
a
ss
if
ica
ti
o
n
o
f
Ac
o
u
stic
S
c
e
n
e
s
a
n
d
Eve
n
ts
,
Ne
w Yo
rk
Cit
y
,
Un
it
e
d
S
tate
s,
2
0
1
9
.
[1
2
]
E.
Ca
k
ir,
G
.
P
a
ra
sc
a
n
d
o
l
o
,
T
.
He
it
to
la,
H.
H
u
tt
u
n
e
n
,
a
n
d
T.
Virtan
e
n
,
“
Co
n
v
o
l
u
ti
o
n
a
l
Re
c
u
rre
n
t
Ne
u
ra
l
Ne
two
rk
s
fo
r
P
o
ly
p
h
o
n
ic
S
o
u
n
d
Ev
e
n
t
De
tec
ti
o
n
,
”
IEE
E/
ACM
T
r
a
n
s.
On
Au
d
i
o
S
p
e
e
c
h
a
n
d
L
a
n
g
u
a
g
e
Pro
c
e
ss
in
g
,
v
o
l.
2
6
,
n
o
.
6,
p
p
.
1
2
9
1
-
1
3
0
3
,
2
0
1
7
.
[1
3
]
J.
J.
Au
c
o
u
t
u
rier,
B
.
De
fre
v
il
le,
a
n
d
F
.
P
a
c
h
e
t,
“
T
h
e
Ba
g
-
of
-
F
r
a
m
e
s
Ap
p
ro
a
c
h
t
o
A
u
d
i
o
P
a
t
tern
Re
c
o
g
n
it
i
o
n
:
A
S
u
fficie
n
t
M
o
d
e
l
f
o
r
Urb
a
n
S
o
u
n
d
sc
a
p
e
s
b
u
t
N
o
t
fo
r
P
o
l
y
p
h
o
n
ic
m
u
sic
,
”
J
o
u
rn
a
l
o
f
t
h
e
Aco
u
stica
l
S
o
c
iety
o
f
Ame
ric
a
,
v
o
l.
1
2
2
,
n
o
.
2
,
p
p
.
8
8
1
-
8
9
1
,
2
0
0
7
.
[1
4
]
C.
C.
Ch
a
n
g
,
C.
J
.
Li
n
,
“
LIBS
V
M
:
A
Li
b
ra
r
y
f
o
r
S
u
p
p
o
rt
Ve
c
to
r
M
a
c
h
i
n
e
s,”
ACM
T
ra
n
sa
c
ti
o
n
s
o
n
I
n
telli
g
e
n
t
S
y
ste
ms
a
n
d
T
e
c
h
n
o
lo
g
y
,
v
o
l.
2
,
n
o
.
3
,
p
p
.
1
-
2
7
,
2
0
1
1
.
[1
5
]
D.
D.
LE
E,
a
n
d
H.
S
.
S
e
u
n
g
,
“
Lea
rn
in
g
th
e
P
a
rts
o
f
O
b
jec
ts
b
y
No
n
-
n
e
g
a
ti
v
e
M
a
tri
x
F
a
c
to
riza
ti
o
n
,
”
N
a
tu
re
,
v
o
l.
4
0
1
,
p
p
.
7
8
8
-
7
9
1
,
1
9
9
5
.
[1
6
]
A.
Kriz
h
e
v
sk
y
,
I.
S
u
tsk
e
v
e
r,
a
n
d
G
.
E.
Hin
to
n
,
“
Im
a
g
e
n
e
t
Clas
sifica
ti
o
n
wit
h
De
e
p
Co
n
v
o
lu
t
io
n
a
l
Ne
u
ra
l
Ne
two
rk
s,”
Ad
v
a
n
c
e
s i
n
Ne
u
ra
l
In
f
o
rm
a
ti
o
n
Pro
c
e
ss
in
g
S
y
ste
ms
,
p
p
.
1
0
9
7
-
1
1
0
5
,
2
0
1
2
.
[1
7
]
G
ra
v
e
s,
A.
M
o
h
a
m
e
d
,
a
n
d
G
.
E.
Hin
to
n
,
“
S
p
e
e
c
h
Re
c
o
g
n
it
io
n
wit
h
De
e
p
Re
c
u
rre
n
t
Ne
u
ra
l
Ne
two
r
k
s
,”
Pro
c
e
e
d
i
n
g
s
o
f
t
h
e
IEE
E
I
n
t.
Co
n
f.
o
n
Aco
u
sti
c
s S
p
e
e
c
h
a
n
d
S
i
g
n
a
l
Pro
c
e
ss
in
g
(ICAS
S
P)
,
p
p
.
6
6
4
5
-
6
6
4
9
,
2
0
1
3
.
[1
8
]
K.
Ch
o
,
B.
Va
n
M
e
rrien
b
o
e
r,
C
.
G
u
lce
h
re
,
D.
Ba
h
d
a
n
a
u
,
F
.
B
o
u
g
a
re
s,
H.
S
c
h
we
n
k
,
a
n
d
Y.
Be
n
g
i
o
,
“
Lea
rn
in
g
P
h
ra
se
Re
p
re
se
n
tatio
n
s
Us
in
g
R
n
n
En
c
o
d
e
r
-
De
c
o
d
e
r
fo
r
S
tatisti
c
a
l
M
a
c
h
i
n
e
Tran
sla
ti
o
n
,”
Pro
c
e
e
d
in
g
s
o
f
t
h
e
2
0
1
4
Co
n
f.
o
n
Em
p
irica
l
M
e
th
o
d
s
in
N
a
tu
r
a
l
L
a
n
g
u
a
g
e
Pr
o
c
e
ss
in
g
(E
M
NL
P)
,
p
p
.
1
7
2
4
-
1
7
3
4
,
2
0
1
4
.
[1
9
]
D.
Ba
h
d
a
n
a
u
,
J.
Ch
o
ro
ws
k
i,
D.
S
e
rd
y
u
k
,
P
.
Bra
k
e
l
a
n
d
Y.
Be
n
g
io
,
"
En
d
-
to
-
e
n
d
Atten
ti
o
n
-
b
a
se
d
Lar
g
e
Vo
c
a
b
u
lary
S
p
e
e
c
h
Re
c
o
g
n
it
io
n
,
"
IEE
E
In
te
rn
a
ti
o
n
a
l
Co
n
fer
e
n
c
e
o
n
Aco
u
stics
,
S
p
e
e
c
h
a
n
d
S
i
g
n
a
l
Pr
o
c
e
ss
in
g
(IC
AS
S
P)
,
p
p
.
4
9
4
5
-
4
9
4
9
,
2
0
1
6
.
[2
0
]
V.
M
n
ih
,
N.
He
e
ss
,
A.
G
ra
v
e
s
,
e
t
a
l.
,
“
Re
c
u
rre
n
t
M
o
d
e
ls
o
f
Vis
u
a
l
Atten
ti
o
n
,
”
Ad
v
a
n
c
e
s
in
Ne
u
ra
l
In
f
o
rm
a
ti
o
n
Pro
c
e
ss
in
g
S
y
ste
ms
,
p
p
.
2
2
0
4
-
2
2
1
2
,
2
0
1
4
.
[2
1
]
T.
N.
S
a
i
n
a
th
,
O.
Vin
y
a
ls,
A.
S
e
n
io
r,
a
n
d
H.
S
a
k
,
“
Co
n
v
o
l
u
ti
o
n
a
l,
Lo
n
g
S
h
o
rt
-
term
M
e
m
o
ry
,
F
u
l
ly
Co
n
n
e
c
ted
De
e
p
Ne
u
ra
l
Ne
two
rk
s
,
”
Pr
o
c
e
e
d
in
g
s
o
f
th
e
2
0
1
5
I
EE
E
I
n
t.
C
o
n
f.
o
n
Ac
o
u
stics
,
S
p
e
e
c
h
a
n
d
S
i
g
n
a
l
Pro
c
e
ss
in
g
(ICA
S
S
P)
,
Brisb
a
n
e
,
p
p
.
4
5
8
0
-
4
5
8
4
,
2
0
1
5
.
[
2
2
]
K
.
C
h
o
i
,
G
.
F
a
z
e
k
a
s
,
M
.
S
a
n
d
l
e
r
,
K
.
C
h
o
,
“
C
o
n
v
o
l
u
t
i
o
n
a
l
R
e
c
u
r
r
e
n
t
N
e
u
r
a
l
N
e
t
w
o
r
k
s
f
o
r
M
u
s
i
c
C
l
a
s
s
i
f
i
c
a
t
i
o
n
,
”
P
r
o
c
e
e
d
i
n
g
s
o
f
t
h
e
2
0
1
7
I
E
E
E
I
n
t
.
C
o
n
f
.
o
n
A
c
o
u
s
t
i
c
s
,
S
p
e
e
c
h
a
n
d
S
i
g
n
a
l
P
r
o
c
e
s
s
i
n
g
(
I
C
A
S
S
P
)
,
p
p
.
2
3
9
2
-
2
3
9
6
,
2
0
1
7
.
[2
3
]
S
.
H.
J
u
n
g
,
Y.
J.
C
h
u
n
g
,
“
Au
d
io
Ev
e
n
t
De
tec
ti
o
n
Us
in
g
CNN
a
n
d
CRNN
,”
Pro
c
e
e
d
in
g
s
o
f
t
h
e
7
th
In
ter
n
a
ti
o
n
a
l
Co
n
fer
e
n
c
e
o
n
Ne
x
t
Ge
n
e
ra
ti
o
n
Co
mp
u
ter
a
n
d
In
fo
rm
a
t
io
n
T
e
c
h
n
o
lo
g
y
,
Ho
k
k
a
i
d
o
,
p
p
.
1
3
4
-
1
3
7
,
2
0
1
8
.
[2
4
]
R.
Ha
rb
,
F
.
P
e
r
n
k
o
p
f,
“
S
o
u
n
d
E
v
e
n
t
De
tec
ti
o
n
Us
in
g
Wea
k
ly
Lab
e
led
S
e
m
i
-
su
p
e
rv
ise
d
Da
ta
with
G
CRN
Ns
,
VA
T
a
n
d
S
e
lf
-
a
d
a
p
t
iv
e
Lab
e
l
Re
fin
e
m
e
n
t,
”
W
o
rk
sh
o
p
o
n
De
tec
ti
o
n
a
n
d
Cla
ss
if
ica
ti
o
n
o
f
Ac
o
u
stic
S
c
e
n
e
s
a
n
d
Eve
n
ts
,
S
u
rre
y
,
UK
,
Oc
t.
2
0
1
8
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
1
6
9
3
-
6
9
3
0
T
E
L
KOM
NI
KA
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l Co
n
tr
o
l
,
Vo
l.
18
,
No
.
5
,
Octo
b
e
r
2
0
2
0
:
2
5
8
7
-
259
6
2596
[2
5
]
A.
M
e
sa
ro
s,
T.
He
it
to
la,
a
n
d
T.
Vi
rtan
e
n
,
“
M
e
tri
c
s
fo
r
P
o
ly
p
h
o
n
ic
S
o
u
n
d
Ev
e
n
t
De
tec
ti
o
n
,
”
Ap
p
li
e
d
S
c
ien
c
e
s
,
v
o
l
.
6
,
n
o
.
6
,
p
p
.
1
6
2
-
1
7
8
,
2
0
1
6
.
B
I
O
G
RAP
H
I
E
S O
F
AUTH
O
RS
S
u
k
wha
n
J
u
n
g
re
c
e
iv
e
d
h
is
B.
S
c
.
a
n
d
M
.
S
c
.
De
g
re
e
in
El
e
c
tro
n
ics
En
g
i
n
e
e
rin
g
fr
o
m
Ke
imy
u
n
g
Un
iv
e
rsity
,
Da
e
g
u
,
S
o
u
th
K
o
re
a
in
2
0
1
6
a
n
d
2
0
1
8
,
re
sp
e
c
ti
v
e
l
y
.
He
h
a
s
b
e
e
n
wit
h
S
a
m
j
u
El
e
c
tro
in
c
s
Co
.
si
n
c
e
M
a
rc
h
2
0
1
8
.
His
m
a
in
re
se
a
rc
h
in
tere
sts
a
r
e
a
u
d
io
e
v
e
n
t
d
e
tec
ti
o
n
u
n
d
e
r
n
o
isy
e
n
v
iro
n
m
e
n
ts an
d
d
e
e
p
lea
r
n
in
g
fo
r
a
rti
ficia
l
i
n
telli
g
e
n
c
e
.
Yo
n
g
j
o
o
Chu
n
g
re
c
e
iv
e
d
h
is
B
.
S
c
.
d
e
g
re
e
in
El
e
c
tro
n
ics
E
n
g
i
n
e
e
rin
g
fro
m
S
e
o
u
l
Na
ti
o
n
a
l
Un
iv
e
rsity
,
S
e
o
u
l,
S
o
u
th
K
o
re
a
in
1
9
8
8
.
He
e
a
rn
e
d
h
is
M
.
S
c
.
a
n
d
P
h
D
d
e
g
re
e
in
El
e
c
tri
c
a
l
a
n
d
El
e
c
tro
n
ics
En
g
in
e
e
rin
g
fro
m
Ko
re
a
Ad
v
a
n
c
e
d
S
c
ien
c
e
a
n
d
Tec
h
n
o
lo
g
y
,
Da
e
jo
n
,
S
o
u
t
h
Ko
re
a
n
in
1
9
9
5
.
He
is
c
u
rre
n
tl
y
a
P
ro
fe
ss
o
r
with
th
e
De
p
a
rtme
n
t
o
f
E
lec
tro
n
ics
En
g
i
n
e
e
rin
g
a
t
Ke
imy
u
n
g
Un
iv
e
rsity
,
Da
e
g
u
,
S
.
Ko
re
a
.
His
re
se
a
rc
h
in
tere
sts
a
re
in
t
h
e
a
re
a
s
o
f
sp
e
e
c
h
re
c
o
g
n
i
ti
o
n
,
a
u
d
i
o
e
v
e
n
t
d
e
tec
ti
o
n
,
m
a
c
h
i
n
e
lea
rn
in
g
a
n
d
p
a
tt
e
rn
re
c
o
g
n
it
io
n
.
Evaluation Warning : The document was created with Spire.PDF for Python.