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.
1
9
,
No
.
2
,
A
p
r
il
2
0
2
1
,
p
p
.
51
5
~
52
2
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
9
i
2
.
1
6
7
9
3
515
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
The co
nv
o
lutiona
l neural ne
tworks
f
o
r Amazig
h sp
e
ech
recog
nition sy
ste
m
M
er
y
a
m
T
elm
em
,
Yo
us
s
ef
G
ha
no
u
M
o
u
la
y
Ism
a
il
Un
i
v
e
rsity
,
M
e
k
n
e
s,
M
o
r
o
c
c
o
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Ma
y
2
6
,
2
0
2
0
R
ev
is
ed
Sep
1
3
,
2
0
2
0
Acc
ep
ted
Sep
2
4
,
2
0
2
0
In
t
h
is p
a
p
e
r,
we
p
re
se
n
t
a
n
a
p
p
ro
a
c
h
b
a
se
d
o
n
c
o
n
v
o
lu
ti
o
n
a
l
n
e
u
ra
l
n
e
two
r
k
s
to
b
u
il
d
a
n
a
u
t
o
m
a
ti
c
sp
e
e
c
h
re
c
o
g
n
it
io
n
s
y
ste
m
fo
r
th
e
Am
a
z
ig
h
lan
g
u
a
g
e
.
Th
is
s
y
ste
m
is
b
u
il
t
wit
h
Ten
s
o
rF
lo
w
a
n
d
u
se
s
m
e
l
fre
q
u
e
n
c
y
c
e
p
stra
l
c
o
e
fficie
n
t
(
M
F
CC
)
to
e
x
trac
t
f
e
a
tu
r
e
s.
In
o
rd
e
r
to
tes
t
th
e
e
ffe
c
t
o
f
th
e
sp
e
a
k
e
r'
s
g
e
n
d
e
r
a
n
d
a
g
e
o
n
t
h
e
a
c
c
u
ra
c
y
o
f
th
e
m
o
d
e
l
,
t
h
e
s
y
ste
m
wa
s
train
e
d
a
n
d
tes
ted
o
n
se
v
e
ra
l
d
a
tas
e
ts.
Th
e
first
e
x
p
e
rime
n
t
th
e
d
a
tas
e
t
c
o
n
sists
o
f
9
2
4
0
a
u
d
io
fi
les
.
Th
e
se
c
o
n
d
e
x
p
e
rime
n
t
th
e
d
a
tas
e
t
c
o
n
sists
o
f
9
2
4
0
a
u
d
io
fi
les
d
istri
b
u
ted
b
e
twe
e
n
fe
m
a
les
a
n
d
m
a
les
’
sp
e
a
k
e
rs.
Th
e
las
t
e
x
p
e
ri
m
e
n
t
3
th
e
d
a
tas
e
t
c
o
n
sists
o
f
1
3
8
6
0
a
u
d
i
o
fi
les
d
istri
b
u
te
d
b
e
twe
e
n
a
g
e
9
-
1
5
,
a
g
e
1
6
-
3
0
,
a
n
d
a
g
e
3
0
+
.
Th
e
re
su
l
t
s
h
o
ws
t
h
a
t
t
h
e
m
o
d
e
l
trai
n
e
d
o
n
a
d
a
tas
e
t
o
f
a
d
u
l
t
sp
e
a
k
e
r’
s a
g
e
+
3
0
c
a
teg
o
r
ies
g
e
n
e
ra
tes
th
e
b
e
st ac
c
u
ra
c
y
with
9
3
.
9
%
.
K
ey
w
o
r
d
s
:
Am
az
ig
h
lan
g
u
ag
e
C
o
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
Dee
p
lear
n
in
g
Me
l f
r
eq
u
en
c
y
ce
p
s
tr
al
co
ef
f
icien
t
Sp
ec
tr
o
g
r
am
Sp
ee
ch
r
ec
o
g
n
itio
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
:
Me
r
y
am
T
elm
em
T
ea
m
T
I
M,
Hig
h
Sch
o
o
l o
f
T
e
ch
n
o
lo
g
y
Mo
u
lay
I
s
m
ail
Un
iv
er
s
ity
Me
k
n
es,
Mo
r
o
cc
o
E
m
ail:
m
er
y
am
telm
em
@
g
m
ai
l.c
o
m
1.
I
NT
RO
D
UCT
I
O
N
Dee
p
lear
n
in
g
is
a
b
r
an
ch
o
f
m
ac
h
in
e
lear
n
in
g
.
I
t
co
n
s
is
ts
o
f
lear
n
in
g
h
ig
h
-
le
v
el
r
ep
r
ese
n
tatio
n
s
o
f
d
ata
u
s
in
g
d
ee
p
n
eu
r
al
n
etwo
r
k
s
.
W
ith
tech
n
o
lo
g
ical
an
d
s
cien
tific
ad
v
an
ce
s
,
d
ee
p
lear
n
i
n
g
h
as
m
ad
e
a
p
lace
in
m
an
y
ar
ea
s
esp
ec
ially
in
th
e
f
ield
o
f
au
to
m
atic
s
p
ee
ch
r
ec
o
g
n
itio
n
.
Au
t
o
m
atic
s
p
ee
ch
r
ec
o
g
n
itio
n
is
a
co
m
p
u
ter
tech
n
iq
u
e
in
ten
d
ed
t
o
tr
an
s
cr
ib
e
a
s
p
ee
ch
s
ig
n
al
in
to
tex
t
[
1
]
.
Sin
ce
a
lo
n
g
tim
e,
th
e
h
id
d
e
n
m
ar
k
o
v
m
o
d
els
[
2
,
3
]
it
was
a
p
er
f
ec
t
s
o
lu
tio
n
to
th
e
p
r
o
b
lem
s
o
f
s
p
ee
ch
r
ec
o
g
n
itio
n
.
B
u
t,
in
2
0
1
2
,
d
ee
p
lear
n
i
n
g
[
4
]
h
as
a
r
ev
o
lu
tio
n
with
th
e
a
p
p
e
ar
an
ce
o
f
co
n
v
o
lu
ti
o
n
al
n
e
u
r
a
l
n
etwo
r
k
(
C
NN
)
[
5
]
.
I
t
is
ar
g
u
ab
ly
th
at
th
e
m
o
s
t
p
o
p
u
lar
ar
c
h
itectu
r
e,
th
ey
h
a
v
e
ap
p
licatio
n
s
in
im
ag
e
an
d
v
id
e
o
r
ec
o
g
n
itio
n
,
r
ec
o
m
m
e
n
d
er
s
y
s
tem
s
[
6
]
,
m
ed
ical
im
ag
e
an
d
a
u
d
io
an
aly
s
is
[
7
]
,
s
u
cc
ess
f
u
lly
ap
p
lied
in
s
p
ee
c
h
r
ec
o
g
n
itio
n
.
I
n
th
is
w
o
r
k
,
we
b
u
ilt
a
n
Am
az
i
g
h
s
p
ea
ch
r
ec
o
g
n
itio
n
s
y
s
tem
b
ased
o
n
C
NN
an
d
GPU
co
m
p
u
tatio
n
u
s
in
g
T
en
s
o
r
Flo
w
,
wh
ich
is
an
o
p
en
s
o
u
r
ce
lib
r
ar
y
wr
itten
in
p
y
th
o
n
an
d
C
++
with
a
m
o
d
el
an
d
r
o
b
u
s
t a
r
ch
itectu
r
e
th
at
ca
n
b
e
r
u
n
o
n
m
u
ltip
le
C
PUs
an
d
GPUs
.
T
h
is
p
ap
er
is
o
r
g
an
ized
as
f
o
llo
ws:
s
ec
tio
n
2
we
p
r
esen
t
th
e
r
elate
d
wo
r
k
,
s
ec
tio
n
3
we
d
escr
ib
e
th
e
p
r
in
cip
le
an
d
th
e
th
eo
r
y
o
f
s
p
ee
ch
r
ec
o
g
n
itio
n
,
s
ec
tio
n
4
we
d
escr
ib
e
t
h
e
C
NN,
s
ec
tio
n
5
we
p
r
esen
t
T
en
s
o
r
Flo
w
.
Fin
ally
,
th
e
s
ec
tio
n
6
illu
s
tr
ates
th
e
ex
p
e
r
im
en
t
al
r
esu
lts
f
o
llo
wed
b
y
c
o
n
clu
s
i
o
n
.
2.
RE
L
AT
E
D
WO
RK
In
o
u
r
p
r
ev
io
u
s
wo
r
k
[
3
]
we
h
av
e
d
ev
elo
p
ed
a
n
Am
az
ig
h
s
p
ee
ch
r
ec
o
g
n
itio
n
s
y
s
tem
b
ased
o
n
h
i
d
d
en
Ma
r
k
o
v
m
o
d
el
HM
Ms
u
s
in
g
an
o
p
en
s
o
u
r
ce
C
MU
Sp
h
in
x
-
4
.
T
h
e
c
o
r
p
u
s
co
n
s
is
ts
o
f
1
1
2
2
0
a
u
d
io
f
iles
.
T
h
e
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.
1
9
,
No
.
2
,
Ap
r
il
202
1
:
51
5
-
52
2
516
b
est
o
b
tain
ed
ac
cu
r
ac
y
was
9
0
%
wh
en
we
h
av
e
tr
ain
ed
o
u
r
m
o
d
el
b
y
u
s
in
g
1
2
8
Gau
s
s
ian
m
ix
tu
r
e
m
o
d
els,
an
d
5
n
u
m
b
er
o
f
HM
Ms
s
tates.
P
alo
H
K,
an
d
et
al
.
[
8
]
h
av
e
d
eter
m
in
ed
th
e
a
g
e
o
f
s
p
ea
k
er
b
ased
o
n
em
o
tio
n
al
s
p
ee
ch
p
r
o
s
o
d
y
a
n
d
clu
s
ter
in
g
th
em
u
s
in
g
f
u
zz
y
c
-
m
ea
n
s
alg
o
r
ith
m
.
T
h
is
r
ec
o
g
n
itio
n
o
f
s
p
ee
ch
em
o
tio
n
b
ased
o
n
s
u
itab
le
f
ea
tu
r
es
p
r
o
v
i
d
es
a
g
e
in
f
o
r
m
atio
n
t
h
at
h
elp
ed
th
e
s
o
ciety
in
d
if
f
e
r
en
t
way
s
.
T
h
e
y
h
av
e
u
s
ed
m
an
y
f
ea
tu
r
e
ex
tr
ac
tio
n
tech
n
i
q
u
es.
Am
o
n
g
th
e
ex
tr
ac
ted
f
ea
tu
r
es,
th
e
F0
,
en
er
g
y
o
r
am
p
litu
d
e,
a
n
d
s
p
ee
ch
r
ate.
Z
h
an
g
H
.
,
a
n
d
et
al
.
[
9
]
h
a
v
e
s
tu
d
ied
a
s
er
ies
o
f
n
eu
r
al
n
et
wo
r
k
s
b
ased
ac
o
u
s
tic
m
o
d
els;
tim
e
d
elay
n
eu
r
al
n
etwo
r
k
(
T
DNN
)
,
C
NNs,
an
d
th
e
lo
n
g
s
h
o
r
t
-
ter
m
m
em
o
r
y
(
L
STM
)
,
ap
p
lied
t
h
e
m
in
th
e
Mo
n
g
o
lian
s
p
ee
ch
r
ec
o
g
n
itio
n
s
y
s
tem
s
,
an
d
co
m
p
a
r
ed
th
eir
p
er
f
o
r
m
an
ce
.
T
h
e
r
esu
lt
s
h
o
ws
th
at
th
e
L
STM
is
th
e
m
o
s
t
ac
cu
r
ate
m
o
d
el
with
8
,
1
2
%
W
E
R
.
Sa
to
r
i
H
.
,
an
d
et
al
.
[
1
0
]
h
av
e
d
ev
el
o
p
ed
a
n
d
Am
az
ig
h
ASR
b
ased
o
n
th
e
C
MU
-
Sp
h
in
x
.
T
h
e
s
y
s
tem
g
en
er
ated
9
2
.
8
9
%
o
f
ac
c
u
r
ac
y
.
T
h
e
tr
ain
in
g
was
p
er
f
o
r
m
ed
b
y
u
s
in
g
u
s
in
g
1
6
Gau
s
s
ian
m
ix
tu
r
e
m
o
d
els.
Ku
m
ar
K
.
,
an
d
Ag
g
a
r
wal
R
.
[
1
1
]
h
av
e
b
u
ilt a
Hin
d
i r
ec
o
g
n
itio
n
s
y
s
tem
u
s
in
g
HT
K
b
ased
o
n
th
e
h
id
d
en
Ma
r
k
o
v
m
o
d
els
HM
Ms.
T
h
e
co
r
p
u
s
o
f
tr
ai
n
in
g
co
n
s
is
ts
o
f
1
0
2
wo
r
d
s
.
T
h
e
s
y
s
tem
p
r
o
d
u
ce
d
8
7
.
0
1
% o
f
ac
cu
r
ac
y
.
3.
AUTOM
AT
I
C
SP
E
E
CH
R
E
CO
G
NIT
I
O
N
SYS
T
E
M
T
h
e
p
r
o
b
lem
o
f
s
p
ee
ch
r
ec
o
g
n
itio
n
aim
s
to
co
n
v
e
r
t
th
e
s
p
ee
c
h
s
ig
n
al
to
s
eq
u
en
c
e
o
f
o
b
s
er
v
atio
n
s
X,
in
a
p
r
o
ce
s
s
ca
lled
f
ea
tu
r
e
e
x
tr
ac
tio
n
.
T
h
e
d
ec
o
d
er
lo
o
k
s
f
o
r
th
e
s
eq
u
e
n
ce
o
f
wo
r
d
s
W
*
m
ax
im
izin
g
th
e
f
o
llo
win
g
eq
u
atio
n
:
∗
=
(
|
)
(
1
)
Af
ter
ap
p
ly
in
g
th
e
B
ay
es th
eo
r
em
,
th
is
eq
u
atio
n
b
ec
o
m
es:
∗
=
(
|
)
(
)
(
)
(
2
)
P (
X)
is
co
n
s
id
er
ed
co
n
s
tan
t a
n
d
r
em
o
v
ed
f
r
o
m
(
2
)
.
∗
=
(
|
)
(
)
(
3
)
3
.
1
.
P
re
-
pro
ce
s
s
ing
3
.
1
.
1
.
Audi
o
t
o
s
pect
rum
Sp
ee
ch
,
wh
atev
er
its
lan
g
u
ag
e
,
is
co
n
s
titu
te
o
f
a
f
in
ite
n
u
m
b
er
o
f
d
is
tin
ctiv
e
s
o
u
n
d
elem
e
n
ts
.
T
h
ese
elem
en
ts
f
o
r
m
elem
en
tar
y
lin
g
u
is
tic
u
n
its
an
d
h
av
e
t
h
e
p
r
o
p
er
ty
o
f
ch
an
g
in
g
th
e
m
ea
n
in
g
o
f
a
wo
r
d
.
T
h
ese
elem
en
tar
y
u
n
its
ar
e
ca
lled
p
h
o
n
em
es
[
3
]
.
T
h
e
Ph
o
n
em
es
ca
n
b
e
s
ee
n
as
th
e
b
asic
elem
en
ts
f
o
r
c
o
d
in
g
lin
g
u
is
tic
in
f
o
r
m
atio
n
.
T
h
e
Am
az
i
g
h
alp
h
ab
et
co
n
tain
s
3
3
p
h
o
n
em
es
[
1
0
]
as sh
o
wn
i
n
Fig
u
r
e
1.
Fig
u
r
e
1
.
Of
f
icial
tab
le
o
f
th
e
tifin
ag
h
e
alp
h
ab
et
as r
ec
o
m
m
e
n
d
ed
b
y
I
R
C
AM
[
1
1
]
h
as o
f
f
ic
ially
b
ee
n
th
e
o
n
ly
wr
itin
g
s
y
s
tem
f
o
r
tr
a
n
s
cr
ib
in
g
th
e
Am
az
ig
h
t la
n
g
u
a
g
e
in
M
o
r
o
cc
o
s
in
ce
2
0
0
3
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
Th
e
co
n
vo
lu
tio
n
a
l n
e
u
r
a
l n
etw
o
r
ks fo
r
A
ma
z
ig
h
s
p
ee
ch
r
ec
o
g
n
itio
n
s
ystem
(
Mer
ya
m
Telm
em
)
517
T
h
e
g
r
a
p
h
ic
s
y
s
tem
o
f
th
e
s
tan
d
ar
d
Am
az
ig
h
e
p
r
o
p
o
s
ed
b
y
t
h
e
I
R
C
AM
co
m
p
r
is
es [
1
1
]
-
2
7
co
n
s
o
n
an
ts
o
f
: la
b
els (
ⴼ
,
ⴱ
,
ⵎ
)
,
d
en
tal
(
ⵜ
,
ⴷ
,
ⵟ
,
ⴹ
,
ⵏ
,
ⵔ
,
ⵕ
,
ⵍ
).
-
T
h
e
alv
eo
lar
(
ⵙ
,
ⵣ
,
ⵚ
,
ⵥ
)
(
ⵛ
,
ⴳ
)
ⴽ
,
ⴳ
,
ⴽⵯ
,
ⴽⵯ
,
ⵀ
,
ⵃ
,
ⵀ
,
ⵃ
,
ⵀ
,
ⵀ
ⵀ
.
-
2
s
em
i
-
co
n
s
o
n
a
n
ts
:
ⵢ
an
d
ⵡ
-
v
o
wels: th
e
f
u
ll o
n
es (
ⴰ
,
ⵉ
,
ⵓ
)
,
n
eu
tr
al
(
ⵓ
).
C
NN
tak
es
in
p
u
t
an
im
ag
e,
s
o
t
o
b
e
ab
le
to
r
ec
o
g
n
ize
p
h
o
n
em
es
it
i
s
n
ec
es
s
ar
y
to
p
ass
o
n
s
p
e
ctr
u
m
to
tr
an
s
f
o
r
m
au
d
io
in
to
im
ag
e.
T
h
is
p
r
e
-
p
r
o
ce
s
s
in
g
p
h
ase
is
th
e
lo
n
g
est
a
n
d
m
o
s
t
im
p
o
r
tan
t
p
h
ase
t
o
b
u
ild
ASR
s
y
s
tem
.
I
n
s
p
ee
ch
r
ec
o
g
n
itio
n
s
y
s
tem
,
th
e
m
o
s
t
co
m
m
o
n
f
ea
tu
r
e
ex
tr
a
ctio
n
tech
n
i
q
u
es
ar
e
b
ased
o
n
s
p
ec
tr
u
m
:
PLP,
t
h
e
s
p
ec
tr
o
g
r
am
,
t
h
e
m
el
s
p
ec
tr
o
g
r
am
[1
2
]
,
m
el
f
r
eq
u
e
n
cy
ce
p
s
tr
al
co
ef
f
icien
t
(
MFC
C
)
.
I
n
th
is
wo
r
k
we
h
a
v
e
u
s
ed
MFC
C
tech
n
iq
u
e.
3
.
1
.
2
.
T
he
s
pect
ro
g
ra
m
T
h
e
s
p
ec
tr
o
g
r
am
is
a
r
ep
r
esen
tatio
n
o
f
an
a
u
d
io
f
ile
in
a
f
r
eq
u
en
cy
d
o
m
ain
.
I
n
o
r
d
er
to
c
o
n
v
er
t
r
aw
d
ata
to
s
p
ec
tr
o
g
r
am
we
ap
p
l
y
s
h
o
r
t
-
tim
e
f
o
u
r
ier
t
r
an
s
f
o
r
m
[1
3
]
.
T
h
e
p
r
o
d
u
ce
m
atr
ice
is
th
en
f
ed
in
t
o
a
m
u
lti
-
lay
er
C
NN
f
o
llo
wed
wi
th
a
f
u
lly
-
co
n
n
ec
ted
with
s
o
f
t
m
ax
ac
tiv
atio
n
wh
ich
g
en
e
r
ates
th
e
class
if
icatio
n
v
ec
to
r
.
T
h
e
f
o
llo
win
g
Fig
u
r
e
2
lis
ts
th
e
s
p
ec
tr
o
g
r
am
o
f
th
e
a
lp
h
ab
et
y
a,
y
ab
,
an
d
y
ad
:
Fig
u
r
e
2
.
Sp
ec
tr
o
g
r
am
alp
h
ab
et
y
a,
y
ab
,
an
d
y
ad
3
.
1
.
3
.
M
F
CC
I
n
tr
o
d
u
ce
d
b
y
Dav
is
an
d
Me
r
m
elst
r
ein
in
1
9
8
0
[
2
]
,
MFC
C
s
ar
e
ca
lcu
late
a
f
o
llo
w
[
1
4
]
:
−
Fra
m
e
th
e
s
ig
n
al
in
to
s
h
o
r
t f
r
a
m
es;
−
Ma
p
th
e
p
o
we
r
s
o
f
th
e
s
p
ec
tr
u
m
o
b
tain
ed
ab
o
v
e
o
n
t
o
th
e
m
el
s
ca
le,
u
s
in
g
tr
ian
g
u
lar
o
v
er
la
p
p
in
g
win
d
o
ws;
−
T
ak
e
th
e
lo
g
s
o
f
th
e
p
o
wer
s
at
ea
ch
o
f
t
h
e
m
el
f
r
e
q
u
en
cies;
−
T
ak
e
th
e
d
is
cr
ete
co
s
in
e
tr
a
n
s
f
o
r
m
o
f
th
e
lis
t o
f
m
el
lo
g
p
o
wer
s
,
as if
it we
r
e
a
s
ig
n
al;
−
T
h
e
MFC
C
s
ar
e
th
e
am
p
litu
d
es o
f
th
e
r
esu
ltin
g
s
p
ec
tr
u
m
.
T
h
e
im
ag
e
p
r
o
d
u
ce
d
b
y
th
ese
Pre
-
p
r
o
ce
s
s
in
g
s
tep
s
is
th
en
f
e
d
in
to
m
u
lti
-
lay
er
c
o
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
s
,
with
a
f
u
lly
-
co
n
n
ec
ted
la
y
er
f
o
llo
wed
b
y
a
s
o
f
tm
ax
at
t
h
e
en
d
4.
CO
NVO
L
U
T
I
O
NA
L
NE
UR
AL
NE
T
WO
RK
S
4
.
1
.
T
he
perc
ept
ro
n
Per
ce
p
tr
o
n
is
a
v
e
r
y
s
im
p
le
l
ea
r
n
in
g
m
ac
h
in
e
alg
o
r
ith
m
b
a
s
ed
o
n
a
m
o
d
el
o
f
b
io
lo
g
ical
n
eu
r
o
n
s
,
wh
ich
tak
es
an
in
p
u
t
v
ec
to
r
,
weig
h
m
atr
ice,
an
d
an
ac
tiv
ati
o
n
f
u
n
ctio
n
to
p
r
o
d
u
ce
th
e
d
esire
d
o
u
tp
u
t
[
1
5
,
16]
.
T
h
e
weig
h
ts
ar
e
th
e
p
r
o
p
er
ty
o
f
th
e
c
o
n
n
ec
tio
n
wh
i
ch
r
ep
r
esen
t
th
e
s
tr
en
g
th
o
f
th
e
co
n
n
ec
tio
n
.
E
ac
h
co
n
n
ec
tio
n
h
as a
d
if
f
er
e
n
t w
eig
h
t v
alu
e
wh
ile
b
ias is
th
e
p
r
o
p
e
r
ty
o
f
th
e
n
eu
r
o
n
as sh
o
wn
in
Fig
u
r
e
3.
4
.
2
.
T
he
m
ultila
y
er
perc
ept
r
o
ns
M
L
P
W
h
en
we
co
m
b
in
e
m
an
y
p
er
c
ep
tr
o
n
s
,
we
f
o
r
m
a
m
u
ltil
ay
er
p
er
ce
p
tr
o
n
o
r
m
o
r
e
p
r
ec
is
ely
a
n
ar
tific
ial
n
eu
r
al
n
etwo
r
k
[
1
5
]
.
T
h
e
f
ir
s
t
lay
er
is
th
e
in
p
u
t
lay
er
,
co
r
r
esp
o
n
d
in
g
to
th
e
d
ata
f
ea
t
u
r
es.
T
h
e
last
lay
er
is
th
e
o
u
tp
u
t la
y
e
r
,
wh
ich
p
r
o
v
id
es th
e
o
u
tp
u
t p
r
o
b
ab
ilit
ies o
f
class
es o
r
lab
els
as sh
o
wn
in
Fig
u
r
e
4.
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.
1
9
,
No
.
2
,
Ap
r
il
202
1
:
51
5
-
52
2
518
Fig
u
r
e
3
.
A
p
e
rc
e
p
tro
n
Fig
u
r
e
4
.
T
h
e
m
u
lt
il
a
y
e
r
p
e
rc
e
p
t
ro
n
s
4
.
3
.
CNN
T
h
e
C
NNs
o
r
C
o
n
v
Nets
th
es
e
ar
e
a
p
ar
ticu
lar
f
o
r
m
o
f
n
e
u
r
al
n
etwo
r
k
[
17
,
18
]
th
at
tak
e
s
an
in
p
u
t
im
ag
e
in
s
p
ir
ed
b
y
th
e
wo
r
k
o
f
Hu
b
el
a
n
d
W
iesel
o
n
th
e
p
r
im
ar
y
v
is
u
al
c
o
r
tex
o
f
th
e
ca
t
[
1
9
]
as
s
h
o
wn
i
n
Fig
u
r
e
5
.
T
h
e
C
NN
ar
ch
itectu
r
e
h
as
two
co
m
p
o
n
en
ts
:
th
e
co
n
v
o
lu
ti
v
e
p
ar
t
o
r
f
ea
tu
r
e
ex
tr
ac
tio
n
p
ar
t,
we
u
s
e
s
p
ec
tr
o
g
r
am
tech
n
iq
u
e
to
ex
tr
ac
t
th
e
f
ea
t
u
r
e.
A
n
d
t
h
e
class
if
icatio
n
p
ar
t,
th
e
v
ec
to
r
o
f
f
ea
t
u
r
e
ex
t
r
ac
ted
b
y
t
h
e
co
n
v
o
l
u
tiv
e
p
ar
t
is
f
ee
d
to
th
e
f
u
lly
co
n
n
ec
ted
lay
er
s
lead
in
g
in
to
th
e
o
u
t
p
u
t
lay
er
w
h
ich
r
ep
r
esen
ts
th
e
class
if
ier
.
T
h
e
co
n
v
o
l
u
tiv
e
p
ar
t c
o
n
s
is
ts
o
f
[
5
-
2
0
]
.
C
o
n
v
o
lu
tio
n
al
lay
er
:
c
o
n
v
o
lu
tio
n
is
o
n
e
o
f
th
e
m
ain
b
u
i
ld
in
g
b
lo
ck
s
o
f
a
C
NN.
b
ased
o
n
its
co
n
v
o
l
u
tio
n
al
m
ath
e
m
atica
l
p
r
in
cip
le
[
2
1
]
,
is
co
n
s
is
ts
o
f
a
s
et
o
f
lear
n
ab
le
f
ilter
s
,
o
r
k
er
n
els.
E
ac
h
f
ilter
is
ap
p
lied
b
y
i
n
d
ep
e
n
d
en
tly
s
tr
id
in
g
o
v
er
th
e
en
tire
in
p
u
t,
cr
ea
ti
n
g
an
o
u
tp
u
t
f
ea
t
u
r
e
m
ap
f
o
r
e
ac
h
f
ilter
.
At
ev
er
y
lo
ca
tio
n
,
a
m
atr
i
x
m
u
ltip
licatio
n
is
p
er
f
o
r
m
ed
an
d
s
u
m
s
th
e
r
esu
lt o
n
to
th
e
f
ea
tu
r
e
m
a
p
as
s
h
o
wn
in
Fig
u
r
e
6
.
Fig
u
r
e
5
.
I
ll
u
s
tr
atio
n
o
f
t
h
e
ar
c
h
itectu
r
e
u
s
ed
f
o
r
th
e
C
NN
with
m
an
y
la
y
er
s
Fig
u
r
e
6
.
C
o
n
v
o
lu
tio
n
al
m
ath
e
m
atica
l p
r
in
cip
le
[
2
0
]
Po
o
lin
g
lay
er
:
r
ed
u
ce
s
th
e
s
ize
o
f
th
e
im
ag
e.
I
t
is
an
ess
en
tial
lay
er
o
f
ten
p
lace
d
b
etwe
en
two
lay
er
s
o
f
co
n
v
o
lu
tio
n
.
T
h
e
R
eL
U
co
r
r
ec
tio
n
lay
er
t
h
er
ef
o
r
e
r
ep
lac
es
all
n
eg
ativ
e
v
alu
es
r
ec
eiv
e
d
.
T
h
e
clas
s
if
icatio
n
p
ar
t
co
n
s
is
ts
o
f
f
u
lly
co
n
n
ec
te
d
lay
er
s
.
Fu
lly
-
co
n
n
ec
te
d
lay
e
r
s
ar
e
u
s
ed
af
ter
th
e
f
in
al
co
n
v
o
lu
tio
n
al
lay
er
in
o
r
d
er
to
m
atch
th
e
o
u
tp
u
t
s
ize
o
f
th
e
n
e
u
r
al
n
etwo
r
k
to
th
e
d
e
s
ir
ed
o
u
tp
u
t
s
ize.
T
h
e
o
u
tp
u
t
is
th
en
p
ass
ed
th
r
o
u
g
h
a
s
o
f
tm
ax
f
u
n
ctio
n
i
n
o
r
d
er
t
o
cr
ea
te
a
p
r
o
b
a
b
ilit
y
r
ep
r
esen
tatio
n
f
o
r
t
h
e
p
r
ed
ictio
n
s
f
o
r
ea
ch
class
in
th
e
s
u
p
er
v
is
ed
lear
n
in
g
s
ettin
g
.
T
o
u
s
e
th
e
f
u
lly
-
c
o
n
n
ec
te
d
lay
e
r
s
,
th
e
o
u
tp
u
t f
r
o
m
t
h
e
f
in
al
co
n
v
o
lu
tio
n
al
lay
er
is
co
m
m
o
n
l
y
f
latten
ed
o
u
t,
o
r
th
e
f
ea
tu
r
e
m
a
p
s
ar
e
s
u
b
s
am
p
led
to
a
s
ize
o
f
1
.
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
Th
e
co
n
vo
lu
tio
n
a
l n
e
u
r
a
l n
etw
o
r
ks fo
r
A
ma
z
ig
h
s
p
ee
ch
r
ec
o
g
n
itio
n
s
ystem
(
Mer
ya
m
Telm
em
)
519
C
NN
Ar
ch
itectu
r
e:
I
n
th
is
p
ap
er
th
e
C
NN
u
s
es to
w
co
n
v
o
lu
ti
o
n
al
lay
er
s
:
−
C
o
n
v
o
lu
tio
n
al
lay
e
r
with
6
4
-
8
an
d
2
0
f
ilter
s
−
R
elu
lay
er
−
Ma
x
p
o
o
lin
g
lay
er
%with
2
×2
f
ilter
−
C
o
n
v
o
lu
tio
n
al
lay
e
r
with
6
4
-
4
an
d
1
0
f
ilter
s
−
Ma
x
p
o
o
lin
g
lay
er
%with
2
×2
f
ilter
−
R
elu
lay
er
5.
T
E
N
SO
RF
L
O
W
T
en
s
o
r
Flo
w
is
an
o
p
e
n
s
o
u
r
ce
lib
r
ar
y
d
e
v
elo
p
e
d
b
y
Go
o
g
le
’
s
AI
o
r
g
a
n
izatio
n
,
as
a
m
i
d
d
lew
ea
r
lib
r
ar
y
th
at
ca
n
b
e
u
s
ed
to
b
u
ild
d
ee
p
l
ea
r
n
in
g
n
eu
r
al
n
etwo
r
k
s
,
T
en
s
o
r
Flo
w
is
wr
itten
in
p
y
th
o
n
a
n
d
c+
+
with
a
m
o
d
el
an
d
r
o
b
u
s
t
ar
c
h
etec
tu
r
e
th
at
c
an
b
e
r
u
n
o
n
m
u
ltip
le
C
PUs
an
d
GPUs
[
2
2
]
as
s
h
o
wn
in
Fig
u
r
e
7
.
Sp
ea
k
e
r
s
an
d
th
e
test
h
as
w
ith
T
en
s
o
r
Flo
w,
m
ac
h
in
e
lea
r
n
in
g
alg
o
r
ith
m
s
ar
e
b
ased
o
n
th
e
co
n
ce
p
t o
f
th
e
d
ata
f
lo
w
g
r
ap
h
o
r
co
m
p
u
tio
n
al
g
r
a
p
h
[
2
3
]
.
T
h
e
n
o
d
es
o
f
th
is
g
r
a
p
h
r
ep
r
esen
t
m
ath
em
atica
l
o
p
er
atio
n
s
.
T
h
e
e
d
g
es
ar
e
te
n
s
o
r
s
.
I
n
ter
m
s
o
f
T
en
s
o
r
Flo
w,
a
ten
s
o
r
is
ju
s
t
a
m
u
lti
-
d
im
en
s
io
n
al
ar
r
a
y
.
E
ac
h
d
ata
f
lo
w
g
r
a
p
h
co
m
p
u
tatio
n
r
u
n
s
with
in
a
s
ess
io
n
o
n
o
n
e
o
r
m
o
r
e
C
PUs
o
r
o
n
e
o
r
m
o
r
e
GPUs
.
A
co
m
p
u
tatio
n
al
g
r
ap
h
in
T
e
n
s
o
r
Flo
w
co
n
s
is
ts
o
f
s
ev
er
al
p
ar
ts
:
−
T
en
s
o
r
:
a
m
u
lti
-
d
im
e
n
s
io
n
al
a
r
r
ay
.
−
Gr
ap
h
:
a
ce
n
tr
al
h
u
b
th
at
co
n
n
ec
ts
all
th
e
v
ar
iab
les,
p
lace
h
o
l
d
er
s
,
co
n
s
tan
ts
to
o
p
e
r
atio
n
s
.
−
C
o
n
s
tan
ts
: a
r
e
f
ix
ed
v
alu
e
ten
s
o
r
s
-
n
o
t tr
ain
a
b
le.
−
Var
iab
les ar
e
ten
s
o
r
s
in
itialized
in
a
s
ess
io
n
-
tr
ain
ab
le.
−
Placeh
o
ld
er
s
:
ar
e
te
n
s
o
r
s
o
f
v
alu
es
th
at
ar
e
u
n
k
n
o
wn
d
u
r
in
g
th
e
g
r
ap
h
c
o
n
s
tr
u
ctio
n
,
b
u
t
p
ass
ed
as
in
p
u
t
d
u
r
in
g
a
s
ess
io
n
.
−
Op
er
atio
n
s
: a
r
e
f
u
n
ctio
n
s
o
n
ten
s
o
r
s
.
−
Ses
s
io
n
: A
s
e
s
s
io
n
cr
ea
tes a
r
u
n
tim
e
in
wh
ich
o
p
er
atio
n
s
ar
e
ex
ec
u
ted
an
d
T
en
s
o
r
s
ar
e
ev
al
u
ated
.
W
e
o
p
ted
f
o
r
T
en
s
o
r
Flo
w
f
o
r
th
e
f
o
llo
win
g
r
ea
s
o
n
s
:
T
en
s
o
r
Flo
w
co
m
es
with
a
co
m
p
lete
s
et
o
f
v
is
u
aliza
tio
n
to
o
ls
th
at
m
ak
e
it
ea
s
y
to
u
n
d
er
s
tan
d
,
d
eb
u
g
,
an
d
o
p
tim
ize
ap
p
licatio
n
s
.
T
en
s
o
r
Flo
w
also
h
as
a
lar
g
e
co
m
m
u
n
ity
o
f
u
s
er
s
an
d
l
o
ts
o
f
d
o
cu
m
en
ta
tio
n
.
Fig
u
r
e
7
.
T
e
n
s
o
r
Flo
w
ar
ch
itec
tu
r
e
r
elea
s
ed
b
y
Go
o
g
le
f
o
r
im
p
lem
en
tin
g
th
e
m
ac
h
in
e
lea
r
n
i
n
g
m
o
d
els
6.
E
XP
E
R
I
M
E
N
T
S AN
D
RE
S
UL
T
S
6
.
1
.
E
nv
iro
nm
ent
W
e
ch
o
o
s
e
to
in
s
tall
T
en
s
o
r
Flo
w
with
GPU,
u
s
e
v
ir
tu
alen
v
in
s
tallatio
n
o
n
a
wo
r
k
s
tatio
n
h
p
Z
6
4
0
I
n
tel
4
co
r
e
,
an
d
we
u
s
e
L
in
u
x
Ub
u
n
tu
1
6
.
0
4
,
to
av
o
id
t
h
e
p
r
o
b
lem
[
2
3
]
.
T
h
e
f
o
llo
win
g
s
o
f
twar
e
n
ee
d
s
to
b
e
in
s
talled
p
r
o
p
er
ly
[
2
4
]
:
−
p
ip
an
d
Vir
tu
alen
v
;
−
C
UDA
T
o
o
lk
it 9
.
0
;
−
GPU
ca
r
d
with
C
o
m
p
u
te
C
ap
a
b
ilit
y
(
C
UDA
)
3
.
0
;
−
GPU
d
r
iv
er
s
;
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.
1
9
,
No
.
2
,
Ap
r
il
202
1
:
51
5
-
52
2
520
−
cu
DNN
SDK
v
7
.
−
Prio
r
e
to
in
s
tallin
g
T
en
s
o
r
Flo
w
with
GPU
s
u
p
p
o
r
t,
e
n
s
u
r
e
th
at
th
e
s
y
s
tem
s
u
p
p
o
r
t
all
N
VI
DI
A
s
o
f
twar
e
r
eq
u
ir
em
e
n
ts
.
6
.
2
.
Co
rpus
T
o
tr
ain
o
u
r
m
o
d
el
we
u
s
e
th
e
d
ataset
co
llected
b
y
Sato
r
i
H,
an
d
al
[
1
0
]
.
T
h
e
s
ig
n
als
wer
e
r
ec
o
r
d
ed
i
n
a
n
o
n
-
n
o
is
y
s
p
ac
e
b
y
th
e
s
am
e
m
icr
o
p
h
o
n
e;
th
e
r
ec
o
r
d
in
g
f
il
es a
r
e
in
MS
W
AV
f
o
r
m
at
with
a
s
p
ec
if
ic
s
am
p
le
r
ate
–
1
6
k
Hz,
16
bi
t
m
o
n
o
.
E
ac
h
s
p
ea
k
er
was
in
v
ited
t
o
p
r
o
n
o
n
ce
3
3
Am
az
i
g
h
letter
s
1
0
tim
es.
Du
r
in
g
tr
ain
in
g
,
th
e
co
r
p
u
s
is
s
ep
ar
ated
in
to
:
−
T
r
ain
in
g
d
ata:
8
0
% o
f
th
e
d
ata
;
−
Vali
d
atio
n
d
ata:1
0
% o
f
th
e
d
at
a
is
r
eser
v
ed
f
o
r
th
e
ev
al
u
atio
n
o
f
th
e
p
r
ec
is
io
n
d
u
r
in
g
th
e
tr
ain
in
g
;
−
Data
test
s
: 1
0
% o
f
th
e
d
ata
is
u
s
ed
to
ev
alu
ate
ac
c
u
r
ac
y
o
n
c
e
th
e
tr
ain
in
g
is
co
m
p
lete.
I
n
th
e
f
o
llo
win
g
T
ab
le
1
.
W
e
d
ef
in
e
h
o
w
m
an
y
au
d
io
f
iles
u
s
ed
in
tr
ain
i
n
g
,
v
alid
atio
n
,
an
d
test
d
ata
f
o
r
3
s
ep
ar
ated
ex
p
er
im
e
n
ts
d
escr
ib
ed
in
th
is
p
ap
e
r
.
T
ab
le
1
.
T
r
ai
n
in
g
,
v
alid
atio
n
,
an
d
test
d
ata
f
o
r
th
e
3
E
x
p
er
i
m
en
ts
Co
C
o
r
p
u
s
(
a
u
d
i
o
f
i
l
e
)
Tr
a
i
n
i
n
g
V
a
l
i
d
a
t
i
o
n
Te
st
Ex
p
e
r
i
me
n
t
1
9
2
4
0
7
3
9
2
9
2
4
9
2
4
Ex
p
e
r
i
m
e
n
t
2
:
F
e
ma
l
e
s
9
2
4
0
7
3
9
2
9
2
4
9
2
4
Ex
p
e
r
i
m
e
n
t
2
:
M
a
l
s
9
2
4
0
7
3
9
2
9
2
4
9
2
4
Ex
p
e
r
i
m
e
n
t
3
:
a
g
e
9
-
1
5
4
6
2
0
3
6
9
6
4
6
2
4
6
2
Ex
p
e
r
i
m
e
n
t
3
:
a
g
e
1
6
-
30
4
6
2
0
3
6
9
6
4
6
2
4
6
2
Ex
p
e
r
i
m
e
n
t
3
:
a
g
e
+
3
0
4
6
2
0
3
6
9
6
4
6
2
4
6
2
6
.
3
.
T
ra
in CNN
wit
h
T
ens
o
r
F
lo
w
B
asically
,
th
er
e
ar
e
3
s
tep
s
to
b
u
ild
a
C
NN
m
o
d
el
in
T
en
s
o
r
f
lo
w:
−
Pre
p
r
o
ce
s
s
in
g
th
e
d
ata;
−
B
u
ild
th
e
m
o
d
el;
b
u
ild
th
e
n
o
d
es a
n
d
o
p
er
atio
n
s
an
d
h
o
w
th
e
y
ar
e
co
n
n
ec
ted
to
ea
ch
o
t
h
er
;
−
T
r
ain
an
d
esti
m
ate
th
e
m
o
d
el
o
n
s
o
m
e
d
ata;
W
e
f
ee
d
o
u
r
C
NN
b
y
s
p
ec
to
r
g
r
am
r
es
u
lts
f
r
o
m
th
e
p
r
e
p
r
o
ce
s
s
in
g
p
h
ase
to
tr
ain
an
d
p
r
ed
ict
th
e
lab
els.
T
h
e
lab
els
u
s
ed
i
n
th
is
p
ap
e
r
ar
e
«
s
ilen
t",
"y
a"
,
"y
ab
"
,
"y
a
d
".
.
.
E
ac
h
co
lu
m
n
r
e
p
r
esen
ts
a
s
et
o
f
s
am
p
les
th
at
was
p
r
ed
icted
to
b
e
ea
ch
,
s
o
th
e
f
ir
s
t
co
lu
m
n
r
ep
r
esen
ts
all
th
e
clip
s
th
at
we
r
e
in
ten
d
ed
to
b
e
s
ilen
ce
.
T
h
e
s
ec
o
n
d
co
lu
m
n
r
e
p
r
esen
ts
all
th
o
s
e
th
at
wer
e
p
r
ed
icted
to
b
e
y
a
wo
r
d
,
an
d
th
e
th
ir
d
"
y
ab
"
[
1
4
]
.
At
th
e
en
d
o
f
th
e
tr
ain
in
g
.
A
f
in
al
co
n
f
u
s
io
n
m
a
tr
ix
will
g
en
e
r
ate.
T
h
e
c
o
lu
m
n
s
o
f
th
is
m
atr
ix
r
ep
r
esen
t
th
e
p
r
ed
ictio
n
la
b
els
an
d
th
e
lin
es r
ep
r
esen
t th
e
ac
tu
al
l
ab
els.
So
,
s
h
e
g
iv
es a
g
o
o
d
s
u
m
m
ar
y
o
f
tr
ain
in
g
er
r
o
r
s
.
6
.
3
.
1
.
E
x
perim
ent
1
T
h
e
co
r
p
u
s
co
n
s
is
ts
o
f
9
2
4
0
au
d
io
f
iles
,
2
8
s
p
ea
k
er
s
was
i
n
v
ited
to
p
r
o
n
o
u
n
ce
3
3
letter
s
Am
az
ig
h
1
0
tim
es.
T
h
e
co
r
p
u
s
is
d
iv
id
e
d
in
to
7
3
9
2
tr
ain
,
9
2
4
test
,
an
d
9
2
4
v
alid
atio
n
au
d
io
f
iles
.
T
h
e
r
esu
lts
s
h
o
w
th
at
th
e
s
y
s
tem
p
r
o
d
u
ce
s
8
9
.
8
%
o
f
ac
cu
r
ac
y
as sh
o
wn
i
n
Fi
g
u
r
e
8.
Fig
u
r
e
8
.
A
g
r
ap
h
s
h
o
win
g
t
h
e
tr
ain
in
g
C
NN
m
o
d
els p
r
o
g
r
es
s
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
Th
e
co
n
vo
lu
tio
n
a
l n
e
u
r
a
l n
etw
o
r
ks fo
r
A
ma
z
ig
h
s
p
ee
ch
r
ec
o
g
n
itio
n
s
ystem
(
Mer
ya
m
Telm
em
)
521
6
.
3
.
2
.
E
x
perim
ent
2
I
n
o
r
d
er
to
test
th
e
ef
f
ec
t
o
f
t
h
e
g
en
d
er
o
n
th
e
q
u
ality
o
f
t
h
e
ac
o
u
s
tic
m
o
d
el,
th
e
co
r
p
u
s
co
n
s
is
ts
o
f
9
2
4
0
au
d
io
f
iles
:
1
4
f
em
ales
a
n
d
1
4
fe
m
ales’
s
p
ea
k
e
r
s
wer
e
in
v
ited
to
p
r
o
n
o
u
n
ce
3
3
letter
s
Am
az
ig
h
1
0
tim
es.
I
n
th
e
f
o
ll
o
win
g
T
ab
le
2
we
d
ef
in
ed
h
o
w
m
an
y
a
u
d
io
s
f
iles
tr
ain
in
g
,
v
alid
atio
n
,
an
d
test
d
ata,
an
d
r
esu
lt.
T
h
e
r
esu
lts
s
h
o
w
th
at
th
e
b
est r
esu
lts
wer
e
r
ec
o
r
d
ed
f
o
r
m
ales w
ith
9
3
.
8
% o
f
ac
cu
r
ac
y
.
T
ab
le
2
.
R
ec
o
g
n
itio
n
ac
cu
r
ac
y
f
o
r
ex
p
er
im
en
t
2
: c
o
r
p
u
s
co
n
s
i
s
ts
o
f
9
2
4
0
a
u
d
io
f
iles
C
o
r
p
u
s
Tr
a
i
n
i
n
g
V
a
l
i
d
a
t
i
o
n
Te
st
A
c
c
u
r
a
c
y
%
F
e
mal
e
s ‘C
o
r
p
u
s
4
6
2
0
3
6
9
6
4
6
2
4
6
2
9
3
.
3
M
a
l
e
s ‘
C
o
r
p
u
s
4
6
2
0
3
6
9
6
4
6
2
4
6
2
9
3
.
8
6
.
3
.
3
.
E
x
perim
ent
3
I
n
th
is
ex
p
er
im
en
t,
we
ev
alu
ate
th
e
p
er
f
o
r
m
an
ce
o
f
a
s
y
s
tem
,
wh
ich
was
tr
ain
ed
an
d
test
ed
f
o
r
d
if
f
er
e
n
t
ag
e.
T
h
e
c
o
r
p
u
s
c
o
n
s
is
ts
o
f
1
3
8
6
0
a
u
d
io
f
iles
:
4
2
s
p
ea
k
er
s
was
in
v
ited
to
p
r
o
n
o
u
n
ce
3
3
l
etter
s
Am
az
ig
h
1
0
tim
es,
we
h
av
e
clas
s
if
ied
th
e
s
p
ea
k
er
s
‘
ag
es
in
to
th
r
ee
ca
teg
o
r
ies:
ag
e
9
-
15,
ag
e
1
6
-
3
0
,
a
n
d
ag
e
+3
0
.
I
n
th
e
f
o
llo
win
g
T
a
b
le
3
we
d
ef
in
ed
h
o
w
m
an
y
a
u
d
io
s
f
iles
tr
ain
i
n
g
,
v
alid
atio
n
,
a
n
d
test
d
ata,
an
d
r
esu
lt.
T
h
e
r
esu
lts
s
h
o
w
th
at
th
e
b
est
r
esu
lts
wer
e
r
ec
o
r
d
e
d
f
o
r
+3
0
ag
e
ca
teg
o
r
y
.
T
o
test
th
e
ef
f
ec
t o
f
s
ex
o
r
ag
e
v
ar
iatio
n
o
n
th
e
q
u
ality
o
f
th
e
s
y
s
tem
,
it
h
as
b
e
en
tr
ain
ed
an
d
test
ed
with
d
if
f
er
en
t
co
r
p
u
s
es
o
u
r
r
esu
lts
ar
e
alr
ea
d
y
en
c
o
u
r
ag
e
d
,
th
e
b
est r
esu
lts
p
r
o
d
u
ce
9
3
.
9
% o
f
ac
cu
r
ac
y
.
T
ab
le
3
.
R
ec
o
g
n
itio
n
ac
cu
r
ac
y
f
o
r
ex
p
er
im
en
t
3
: c
o
r
p
u
s
co
n
s
i
s
ts
o
f
1
3
8
6
0
au
d
i
o
f
iles
A
g
e
C
o
r
p
u
s
Tr
a
i
n
i
n
g
V
a
l
i
d
a
t
i
o
n
t
e
st
A
c
c
u
r
a
c
y
%
9
-
15
4
6
2
0
3
6
9
6
4
6
2
4
6
2
9
1
.
4
6
-
30
4
6
2
0
3
6
9
6
4
6
2
4
6
2
9
2
.
4
+
3
0
4
6
2
0
3
6
9
6
4
6
2
4
6
2
9
3
.
9
6
.
3
.
4
.
Co
m
pa
ra
t
iv
e
a
na
ly
s
is
T
h
e
p
r
esen
ted
w
o
r
k
h
as
b
ee
n
co
m
p
ar
e
d
with
th
e
e
x
is
tin
g
s
im
ilar
task
r
ec
o
g
n
itio
n
,
esp
e
cially
,
th
e
em
o
tio
n
r
ec
o
g
n
itio
n
s
y
s
tem
SER an
d
s
o
u
n
d
ev
en
t
r
ec
o
g
n
iti
o
n
.
T
h
e
f
o
llo
win
g
T
ab
le
4
lis
ts
a
n
u
m
b
er
o
f
r
esu
lts
f
r
o
m
o
u
r
p
r
e
v
io
u
s
wo
r
k
[
3
]
,
Z
h
en
g
W
.
Q
.
,
et
al
.
[
2
5
]
,
Z
h
an
g
H
.
,
et
al
.
[
2
6
]
,
a
n
d
o
u
r
Pro
p
o
s
ed
.
I
n
o
u
r
p
r
e
v
io
u
s
wo
r
k
[
3
]
we
h
av
e
d
ev
el
o
p
ed
t
h
e
Am
az
ig
h
s
p
ee
ch
r
ec
o
g
n
itio
n
s
y
s
tem
b
ased
o
n
h
id
d
e
n
Ma
r
k
o
v
m
o
d
el
HM
Ms
u
s
in
g
th
e
o
p
en
s
o
u
r
ce
C
MU
Sp
h
in
x
-
4
.
T
h
e
co
r
p
u
s
co
n
s
is
t
s
o
f
1
1
2
2
0
au
d
i
o
f
iles
.
T
h
e
s
y
s
tem
o
b
tain
ed
b
est
p
er
f
o
r
m
an
ce
o
f
9
0
%
wh
en
t
r
ain
ed
u
s
in
g
1
2
8
Gau
s
s
ian
m
ix
tu
r
e
m
o
d
els,
an
d
5
n
u
m
b
er
o
f
HM
Ms
s
tates.
Z
h
en
g
W
.
Q
.
,
et
a
l.
[
2
4
]
,
h
av
e
d
ev
elo
p
e
d
th
e
em
o
tio
n
r
ec
o
g
n
i
tio
n
s
y
s
tem
b
ased
o
n
co
n
v
o
lu
ti
o
n
n
eu
r
al
n
etwo
r
k
s
with
2
co
n
v
o
lu
tio
n
s
+2
p
o
o
lin
g
lay
er
s
,
an
d
u
s
in
g
lab
elled
tr
ain
in
g
au
d
i
o
d
ata
an
d
u
s
ed
th
e
lo
g
-
s
p
ec
tr
o
g
r
am
to
ex
tr
ac
t
f
ea
tu
r
e,
co
m
p
o
n
e
n
t
a
n
aly
s
is
P
C
A
to
r
ed
u
ce
th
e
d
im
en
s
io
n
ality
.
T
h
e
s
y
s
tem
ac
h
iev
ed
ab
o
u
t
4
0
%
ac
cu
r
ac
y
.
Z
h
a
n
g
H
.
,
et
al
.
[
2
5
]
,
h
av
e
p
r
o
p
o
s
ed
th
e
s
o
u
n
d
ev
e
n
t
d
etec
tio
n
b
ased
o
n
co
n
v
o
l
u
tio
n
n
eu
r
al
n
etwo
r
k
s
with
2
co
n
v
o
lu
tio
n
+2
p
o
o
lin
g
lay
er
s
,
a
n
d
s
p
ec
tr
o
g
r
am
to
ex
tr
ac
t
f
ea
tu
r
e
.
T
h
e
s
y
s
tem
th
e
s
y
s
tem
ac
h
iev
ed
ab
o
u
t
9
4
.
0
7
%
o
f
ac
cu
r
ac
y
.
I
n
o
u
r
p
r
o
p
o
s
ed
wo
r
k
;
th
e
s
y
s
tem
o
b
tain
ed
t
h
e
b
est
p
er
f
o
r
m
an
ce
o
f
9
3
.
9
%
o
f
ac
cu
r
ac
y
wh
e
n
tr
ain
ed
u
s
in
g
+3
0
ag
e
ca
te
g
o
r
y
.
R
esu
lts
ar
e
v
er
y
s
atis
f
ac
to
r
y
if
co
m
p
ar
e
d
with
th
e
e
x
is
tin
g
s
im
ilar
wo
r
k
s
.
T
ab
le
4
.
T
a
b
u
lar
co
m
p
ar
is
o
n
o
f
th
e
r
ec
o
g
n
itio
n
ac
cu
r
ac
y
o
f
th
e
p
r
o
p
o
s
ed
s
y
s
tem
s
A
p
p
r
o
c
h
C
o
r
p
u
s
A
c
c
u
r
a
c
y
%
M
F
C
C
-
H
M
M
1
1
2
2
0
90
S
p
e
c
t
r
+
C
N
N
(
2
C
o
n
v
+
2
p
o
o
l
)
5
2
8
8
33
S
p
e
c
t
r
+
P
C
A
+
C
N
N
(
2
C
o
n
v
+
2
p
o
o
l
)
5
2
8
8
40
M
F
C
C
+
C
N
N
(
2
C
o
n
v
+
2
p
o
o
l
)
4
0
0
0
9
4
.
0
7
Ex
p
:
1
M
F
C
C
+
C
N
N
(
2
C
o
n
v
+
2
p
o
o
l
)
9
2
4
0
8
9
.
8
Ex
p
:
2
F
e
m
a
l
e
s
M
F
C
C
+
C
N
N
(
2
C
o
n
v
+
2
p
o
o
l
)
4
6
2
0
9
3
.
3
Ex
p
:
2
M
a
l
s
M
F
C
C
+
C
N
N
(
2
C
o
n
v
+
2
p
o
o
l
)
4
6
2
0
9
3
.
8
Ex
p
:
3
A
g
e
9
-
15
M
F
C
C
+
C
N
N
(
2
C
o
n
v
+
2
p
o
o
l
)
4
6
2
0
9
1
.
4
Ex
p
:
3
A
g
e
1
6
-
30
M
F
C
C
+
C
N
N
(
2
C
o
n
v
+
2
p
o
o
l
)
4
6
2
0
9
2
.
4
Ex
p
:
3
A
g
e
+
3
0
M
F
C
C
+
C
N
N
(
2
C
o
n
v
+
2
p
o
o
l
)
4
6
2
0
9
3
.
9
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.
1
9
,
No
.
2
,
Ap
r
il
202
1
:
51
5
-
52
2
522
7.
CO
NCLU
SI
O
N
I
n
th
is
w
o
r
k
,
a
s
im
p
le
m
ac
h
i
n
e
lear
n
in
g
m
o
d
el
th
at
r
ec
o
g
n
i
ze
s
a
3
3
Am
az
ig
h
letter
s
u
s
in
g
C
NN
an
d
GPU
co
m
p
u
tatio
n
was
d
ev
elo
p
ed
.
W
e
f
o
llo
wed
MFC
C
T
ec
h
n
iq
u
e
to
ex
t
r
ac
t
th
e
f
ea
tu
r
es.
T
h
e
s
y
s
tem
was
b
u
ilt
with
o
p
en
s
o
u
r
ce
T
en
s
o
r
Flo
w
.
T
h
e
s
y
s
tem
o
b
tain
ed
th
e
b
e
s
t
r
esu
lts
wh
en
tr
ain
ed
u
s
in
g
+3
0
ag
e
ca
teg
o
r
y
p
r
o
d
u
ce
d
9
3
.
9
% o
f
ac
cu
r
ac
y
.
Ou
r
wo
r
k
a
n
d
am
o
n
g
t
h
e
f
ir
s
t
co
n
tr
ib
u
tio
n
s
ad
d
r
ess
th
e
Am
a
zig
h
ASR
b
ased
o
n
co
n
v
o
l
u
tio
n
al
n
eu
r
al
n
etwo
r
k
s
an
d
GPU
co
m
p
u
tatio
n
.
I
n
p
er
s
p
ec
tiv
e,
to
o
v
er
co
m
e
th
e
o
v
e
r
f
itti
n
g
p
r
o
b
lem
an
d
im
p
r
o
v
e
th
e
p
e
r
f
o
r
m
an
ce
,
w
e
u
s
e
a
Dr
o
p
wea
k
R
eg
u
lar
izatio
n
f
o
r
C
NNs
in
Am
az
ig
h
r
ec
o
g
n
itio
n
s
y
s
tem
d
r
o
p
wea
k
is
b
ased
o
n
th
e
id
ea
o
f
d
r
o
p
p
in
g
th
e
wea
k
weig
h
ts
in
a
n
e
u
r
al
n
etwo
r
k
,
we
m
ea
n
s
ettin
g
its
v
alu
e
to
ze
r
o
s
o
h
e
ca
n
n
o
t i
n
f
lu
en
ce
o
n
th
e
o
u
tp
u
t o
f
its
p
r
o
ce
s
s
in
g
u
n
it.
RE
F
E
R
E
NC
E
S
[1
]
M
.
Ha
m
id
i,
e
t
a
l
.,
“
Am
a
z
ig
h
d
ig
it
s
t
h
ro
u
g
h
in
tera
c
ti
v
e
sp
e
e
c
h
re
c
o
g
n
i
ti
o
n
sy
ste
m
in
n
o
is
y
e
n
v
iro
n
m
e
n
t,
”
In
ter
n
a
t
io
n
a
l
J
o
u
rn
a
l
o
f
S
p
e
e
c
h
T
e
c
h
n
o
l
o
g
y
,
v
o
l
.
2
3
,
n
o
.
1
,
p
p
.
1
0
1
-
1
0
9
,
De
c
e
m
b
e
r
2
0
2
0
.
[2
]
M.
Telm
e
m
,
e
t
a
l
.
,
“
Am
a
z
ig
h
sp
e
e
c
h
re
c
o
g
n
it
io
n
sy
ste
m
b
a
se
d
o
n
CM
USp
h
in
x
,
”
Pr
o
c
e
e
d
in
g
s
o
f
t
h
e
M
e
d
i
ter
ra
n
e
a
n
S
y
mp
o
si
u
m o
n
S
m
a
rt Ci
ty A
p
p
l
ica
ti
o
n
s S
p
rin
g
e
r
,
p
p
.
3
9
7
-
4
1
0
,
Ja
n
u
a
ry
2
0
1
7
.
[3
]
O.
Zea
lo
u
k
,
e
t
a
l
.,
“
Vo
c
a
l
p
a
ra
m
e
ters
a
n
a
ly
sis o
f
sm
o
k
e
r
u
si
n
g
a
m
a
z
ig
h
lan
g
u
a
g
e
,
”
I
n
ter
n
a
ti
o
n
a
l
J
o
u
rn
a
l
o
f
S
p
e
e
c
h
T
e
c
h
n
o
l
o
g
y
,
v
o
l.
2
1
,
n
o
.
1
,
p
p
.
8
5
-
91
,
2
0
1
8
.
[4
]
M
.
Telm
e
m
,
M
.
,
e
t
a
l
.,
“
Esti
m
a
ti
o
n
o
f
th
e
o
p
t
ima
l
HMM
p
a
ra
m
e
ters
fo
r
a
m
a
z
ig
h
s
p
e
e
c
h
re
c
o
g
n
it
i
o
n
s
y
ste
m
u
si
n
g
CM
U
-
sp
h
i
n
x
,
”
Pr
o
c
e
d
ia
Co
mp
u
t
e
r S
c
ien
c
e
,
v
o
l
.
1
2
7
,
p
p
.
9
2
-
1
0
1
,
2
0
1
8
.
[5
]
A.
Bh
a
n
d
a
re
,
e
t
a
l.
,
“
Ap
p
li
c
a
ti
o
n
s
o
f
c
o
n
v
o
lu
t
io
n
a
l
n
e
u
ra
l
n
e
two
r
k
s,
”
In
ter
n
a
ti
o
n
a
l
J
o
u
rn
a
l
o
f
C
o
mp
u
t
e
r
S
c
ien
c
e
a
n
d
In
fo
rm
a
t
io
n
T
e
c
h
n
o
l
o
g
ies
,
v
o
l.
7
,
n
o
,
5
,
p
p
.
2
2
0
6
-
2
2
1
5
,
2
0
1
6
.
[
6
]
D.P
a
laz
,
e
t
a
l.
,
"
E
n
d
-
To
-
E
n
d
Ac
o
u
stic
M
o
d
e
li
n
g
Us
in
g
C
o
n
v
o
lu
ti
o
n
a
l
Ne
u
ra
l
Ne
two
rk
s
F
o
r
HM
M
-
B
a
se
d
Au
to
m
a
ti
c
S
p
e
e
c
h
Re
c
o
g
n
it
i
o
n
,
"
S
p
e
e
c
h
Co
mm
u
n
ica
ti
o
n
,
v
o
l.
1
0
8
,
p
.
1
5
-
3
2
,
2
0
1
9
.
[7
]
S
.
K.
M
o
h
a
p
a
tra,
e
t
a
l
.,
“
Dia
b
e
tes
d
e
tec
ti
o
n
u
si
n
g
d
e
e
p
n
e
u
ra
l
n
e
tw
o
rk
,
”
I
n
ter
n
a
t
io
n
a
l
Co
n
fer
e
n
c
e
o
n
S
o
ft
Co
m
p
u
t
in
g
S
y
ste
ms
,
v
o
l.
4
,
n
o
.
4
,
De
c
e
m
b
e
r
2
0
1
8
,
p
p
.
2
4
3
-
2
4
6
.
[8
]
H.
K.
P
a
lo
,
e
t
a
l
.,
“
Emo
ti
o
n
a
n
a
l
y
sis
fro
m
sp
e
e
c
h
o
f
d
iffere
n
t
a
g
e
g
ro
u
p
s
,
”
Pro
c
e
e
d
in
g
s
o
f
th
e
S
e
c
o
n
d
In
ter
n
a
ti
o
n
a
l
Co
n
fer
e
n
c
e
o
n
Res
e
a
rc
h
i
n
In
telli
g
e
n
t
a
n
d
Co
m
p
u
ti
n
g
in
En
g
in
e
e
ri
n
g
,
v
o
l.
1
0
,
Ju
n
e
2
0
1
7
,
p
p
.
2
8
3
-
2
8
7
.
[9
]
H.
Zh
a
n
g
,
e
t
a
l
.,
“
Co
m
p
a
riso
n
o
n
n
e
u
ra
l
n
e
tw
o
rk
b
a
se
d
a
c
o
u
st
ic
m
o
d
e
l
in
m
o
n
g
o
li
a
n
sp
e
e
c
h
re
c
o
g
n
it
i
o
n
,
”
Asia
n
L
a
n
g
u
a
g
e
Pr
o
c
e
ss
in
g
(IA
L
P)
,
N
o
v
e
m
b
e
r
2
0
1
6
.
[1
0
]
H.
S
a
to
ri,
e
t
a
l
.
,
“
Vo
ix
c
o
m
p
a
ra
iso
n
e
n
tre
f
u
m
e
u
rs
e
t
n
o
n
-
f
u
m
e
u
rs
a
l'
a
id
e
d
e
la
re
c
o
n
n
a
issa
n
c
e
v
o
c
a
le
HMM
S
y
stè
m
e
,”
In
ter
n
a
t
io
n
a
l
J
o
u
rn
a
l
o
f
S
p
e
e
c
h
T
e
c
h
n
o
l
o
g
y
,
v
o
l
.
20
,
n
o
.
4
,
p
p
.
7
7
1
-
7
7
7
,
2
0
1
7
.
[1
1
]
M
.
Am
e
u
r,
e
t
a
l
.
,
“
I
n
it
iati
o
n
la
lan
g
u
e
a
m
a
z
ig
h
,
”
I
n
stit
u
t
Ro
y
a
l
d
e
l
a
Cu
l
tu
re
Ama
zi
g
h
e
,
2
0
0
4
.
[1
2
]
X.
Li
u
,
“
De
e
p
c
o
n
v
o
lu
t
io
n
a
l
a
n
d
LS
TM
n
e
u
ra
l
n
e
tw
o
rk
s
fo
r
a
c
o
u
stic
m
o
d
e
ll
in
g
i
n
a
u
t
o
m
a
ti
c
sp
e
e
c
h
re
c
o
g
n
i
ti
o
n
,
”
p
p
.
1
-
9
,
2
0
1
8
.
[1
3
]
K.
Ku
m
a
r,
e
t
a
l
.,
“
A
Hi
n
d
i
sp
e
e
c
h
re
c
o
g
n
it
i
o
n
sy
ste
m
f
o
r
c
o
n
n
e
c
t
e
d
wo
r
d
s
u
si
n
g
HTK,
”
I
n
ter
n
a
ti
o
n
a
l
J
o
u
r
n
a
l
o
f
Co
mp
u
t
a
ti
o
n
a
l
S
y
ste
ms
En
g
i
n
e
e
rin
g
,
v
o
l.
1
,
n
o
1
,
p
p
.
2
5
-
3
2
,
2
0
1
2
.
[1
4
]
B.
J.
M
o
h
a
n
,
e
t
a
l.
,
"
S
p
e
e
c
h
Re
c
o
g
n
it
i
o
n
Us
in
g
M
F
CC
a
n
d
DTW
,
"
2
0
1
4
In
ter
n
a
ti
o
n
a
l
C
o
n
fer
e
n
c
e
o
n
A
d
v
a
n
c
e
s
i
n
El
e
c
trica
l
En
g
in
e
e
rin
g
(ICAE
E)
,
IEE
E,
2
0
1
4
,
p
p
.
1
-
4
.
[1
5
]
S.
Vie
ira,
S
,
e
t
a
l
.
,
“
Us
in
g
d
e
e
p
l
e
a
rn
in
g
t
o
i
n
v
e
stig
a
te
t
h
e
n
e
u
r
o
i
m
a
g
in
g
c
o
rre
late
s
o
f
p
sy
c
h
iatric
a
n
d
n
e
u
ro
l
o
g
ica
l
d
iso
rd
e
rs:
m
e
th
o
d
s
a
n
d
a
p
p
li
c
a
ti
o
n
s,”
Ne
u
ro
sc
ien
c
e
Bi
o
b
e
h
a
v
io
ra
l
Rev
iews
,
n
o
.
7
4
,
p
p
.
58
-
75,
2
0
1
7
.
[1
6
]
M.
Telm
e
m
,
e
t
a
l
.
,
“
A
c
o
m
p
a
ra
ti
v
e
stu
d
y
o
f
HMM
s
a
n
d
CNN
a
c
o
u
stic
m
o
d
e
l
in
a
m
a
z
ig
h
re
c
o
g
n
it
io
n
sy
ste
m
,
”
Emb
e
d
d
e
d
S
y
ste
ms
a
n
d
Art
if
icia
l
In
telli
g
e
n
c
e
.
S
p
ri
n
g
e
r
,
p
p
.
5
3
3
-
5
4
0
,
Ap
r
il
2
0
2
0
.
[1
7
]
O.
Ab
d
e
l
-
Ha
m
id
,
e
t
a
l
.,
“
Co
n
v
o
lu
ti
o
n
a
l
n
e
u
ra
l
n
e
tw
o
rk
s
fo
r
sp
e
e
c
h
re
c
o
g
n
it
i
o
n
,
”
ACM
T
ra
n
s
a
c
t
io
n
s
o
n
a
u
d
i
o
,
sp
e
e
c
h
,
a
n
d
l
a
n
g
u
a
g
e
p
ro
c
e
ss
in
g
,
v
o
l
.
2
2
,
n
o
1
0
,
p
p
.
1
5
3
3
-
1
5
4
5
,
O
c
to
b
e
r
2
0
1
4
.
[1
8
]
K.
Krish
n
a
,
e
t
a
l
.
,
“
A
stu
d
y
o
f
a
ll
-
c
o
n
v
o
lu
ti
o
n
a
l
e
n
c
o
d
e
rs
fo
r
c
o
n
n
e
c
ti
o
n
ist
tem
p
o
ra
l
c
las
sifica
ti
o
n
,
”
2
0
1
8
IE
EE
In
ter
n
a
t
io
n
a
l
C
o
n
fer
e
n
c
e
o
n
Ac
o
u
stics
,
S
p
e
e
c
h
a
n
d
S
ig
n
a
l
Pro
c
e
ss
in
g
(IC
AS
S
P)
,
2
0
1
8
.
[1
9
]
D.
H.
Hu
b
e
l
,
e
t
a
l
.,
“
Re
c
e
p
ti
v
e
fie
ld
s
,
b
i
n
o
c
u
lar
in
tera
c
ti
o
n
a
n
d
fu
n
c
ti
o
n
a
l
a
rc
h
i
tec
tu
re
in
t
h
e
c
a
t'
s
v
isu
a
l
c
o
rtex
,
”
T
h
e
J
o
u
rn
a
l
o
f
p
h
y
sio
l
o
g
y
,
v
o
l.
1
6
0
,
n
o
1
,
p
p
.
1
0
6
,
Ja
n
u
a
ry
1
9
6
2
.
[2
0
]
A.
Kh
a
n
,
e
t
a
l
.
,
“
A
su
rv
e
y
o
f
t
h
e
re
c
e
n
t
a
rc
h
it
e
c
tu
re
s
o
f
d
e
e
p
c
o
n
v
o
lu
ti
o
n
a
l
n
e
u
ra
l
n
e
two
r
k
s,
”
Arti
fi
c
ia
l
In
tell
ig
e
n
c
e
Rev
iew
,
2
0
2
0
.
[
2
1
]
B
r
a
n
d
o
n
R
o
h
r
e
r
,
“
H
o
w
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
u
r
a
l
n
e
t
w
o
r
k
s
w
o
r
k
,
”
[
O
n
l
i
n
e
]
.
A
v
a
i
l
a
b
l
e
:
h
t
t
p
s
:
/
/
b
r
o
h
r
e
r
.
g
i
t
h
u
b
.
i
o
/
h
o
w
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
u
r
a
l
n
e
t
w
o
r
k
s
w
o
r
k
:
h
t
m
l
.
[2
2
]
[On
li
n
e
].
A
v
a
il
a
b
le:
h
tt
p
s://
ww
w.
ten
so
ro
w.
o
rg
/
tu
t
o
rials/se
q
u
e
n
c
e
s/a
u
d
io
re
c
o
g
n
i
ti
o
n
.
[2
3
]
G
o
ld
sb
o
ro
u
g
h
,
P
.
,
“
A t
o
u
r
o
f
Ten
so
ro
w
,
”
a
rX
iv p
re
p
rin
t
a
rX
iv:1
6
1
0
.
0
1
1
7
8
,
2
0
1
6
.
[2
4
]
Ten
so
rF
l
o
w,
“
In
sta
ll
Ten
so
rF
lo
w
2
,
”
[On
li
n
e
].
Av
a
il
a
b
le:
h
tt
p
s:/
/www
.
ten
so
ro
w.
o
rg
/
in
sta
ll/
.
[2
5
]
W.
Q.
Z
h
e
n
g
,
e
t
a
l.
,
“
An
e
x
p
e
ri
m
e
n
tal
stu
d
y
o
f
sp
e
e
c
h
e
m
o
t
io
n
re
c
o
g
n
it
i
o
n
b
a
se
d
o
n
d
e
e
p
c
o
n
v
o
lu
ti
o
n
a
l
n
e
u
ra
l
n
e
two
rk
s
,
”
2
0
1
5
I
n
ter
n
a
t
io
n
a
l
c
o
n
fer
e
n
c
e
o
n
a
_
e
c
ti
v
e
c
o
mp
u
ti
n
g
a
n
d
i
n
telli
g
e
n
t
i
n
ter
a
c
ti
o
n
(ACII
)
,
S
e
p
tem
b
e
r
2
0
1
5
,
p
p
.
8
2
7
-
831
.
[2
6
]
H.
Zh
a
n
g
,
e
t
a
l
.,
“
Ro
b
u
st
s
o
u
n
d
e
v
e
n
t
re
c
o
g
n
it
i
o
n
u
si
n
g
c
o
n
v
o
lu
t
i
o
n
a
l
n
e
u
ra
l
n
e
two
r
k
s,”
2
0
1
5
I
EE
E
in
ter
n
a
ti
o
n
a
l
c
o
n
fer
e
n
c
e
o
n
a
c
o
u
stics
,
sp
e
e
c
h
a
n
d
si
g
n
a
l
p
ro
c
e
ss
in
g
(IC
AS
S
P)
,
2
0
1
5
,
p
p
.
5
5
9
-
5
6
3
.
Evaluation Warning : The document was created with Spire.PDF for Python.