I
nte
rna
t
io
na
l J
o
urna
l o
f
E
lect
rica
l a
nd
Co
m
p
ute
r
E
ng
in
ee
ring
(
I
J
E
CE
)
Vo
l.
9
,
No
.
4
,
A
u
g
u
s
t
201
9
,
p
p
.
2
9
6
0
~2
9
6
7
I
SS
N:
2
0
8
8
-
8708
,
DOI
: 1
0
.
1
1
5
9
1
/
i
j
ec
e
.
v9
i
4
.
p
p
2
9
6
0
-
2967
2960
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ia
e
s
co
r
e
.
co
m/
jo
u
r
n
a
ls
/in
d
ex
.
p
h
p
/
I
JE
C
E
Ba
y
esia
n
d
ista
nce
m
etric
l
ea
rning
a
nd
i
ts
a
ppli
ca
tion
in
a
uto
m
a
tic
s
pea
k
e
r
recog
nition s
y
st
e
m
s
Sa
t
y
a
na
nd
Sin
g
h
Co
ll
e
g
e
o
f
En
g
in
e
e
rin
g
S
c
ien
c
e
a
n
d
T
e
c
h
n
o
l
o
g
y
,
F
ij
i
Na
ti
o
n
a
l
U
n
i
v
e
rsit
y
,
F
ij
i
Isla
n
d
,
S
u
v
a
Art
icle
I
nfo
AB
ST
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
M
ar
4
,
2
0
1
9
R
ev
i
s
ed
Mar
2
2
,
2
0
1
9
A
cc
ep
ted
A
p
r
4
,
2
0
1
9
T
h
is
p
a
p
e
r
p
ro
p
o
se
s
sta
te
-
o
f
th
e
-
a
rt
A
u
to
m
a
ti
c
S
p
e
a
k
e
r
Re
c
o
g
n
it
io
n
S
y
ste
m
(A
S
R)
b
a
se
d
o
n
Ba
y
e
sia
n
Dista
n
c
e
L
e
a
rn
in
g
M
e
tri
c
a
s
a
f
e
a
tu
re
e
x
tra
c
to
r.
In
th
is
m
o
d
e
li
n
g
,
I
e
x
p
lo
re
d
th
e
c
o
n
stra
in
ts
o
f
th
e
d
i
sta
n
c
e
b
e
t
w
e
e
n
m
o
d
if
ied
a
n
d
sim
p
li
f
ied
i
-
v
e
c
to
r
p
a
irs
b
y
th
e
sa
m
e
sp
e
a
k
e
r
a
n
d
d
if
f
e
re
n
t
sp
e
a
k
e
rs.
A
n
a
p
p
ro
x
im
a
ti
o
n
o
f
th
e
d
istan
c
e
m
e
tri
c
is
u
se
d
a
s
a
we
ig
h
ted
c
o
v
a
rian
c
e
m
a
tri
x
f
ro
m
th
e
h
ig
h
e
r
e
ig
e
n
v
e
c
t
o
rs
o
f
th
e
c
o
v
a
rian
c
e
m
a
tri
x
,
w
h
ich
is
u
se
d
to
e
stim
a
te
th
e
p
o
ste
rio
r
d
istri
b
u
t
io
n
o
f
th
e
m
e
tri
c
d
istan
c
e
.
G
iv
e
n
a
sp
e
a
k
e
r
tag
,
I
se
lec
t
th
e
d
a
ta
p
a
ir
o
f
th
e
d
if
f
e
r
e
n
t
sp
e
a
k
e
rs
w
it
h
th
e
h
ig
h
e
st
c
o
sin
e
sc
o
re
to
f
o
r
m
a
s
e
t
o
f
sp
e
a
k
e
r
c
o
n
stra
in
ts.
T
h
is
c
o
ll
e
c
ti
o
n
c
a
p
tu
re
s
th
e
m
o
st
d
isc
rim
in
a
ti
n
g
v
a
riab
il
it
y
b
e
t
we
e
n
th
e
sp
e
a
k
e
rs
in
th
e
train
in
g
d
a
ta.
T
h
is
Ba
y
e
sia
n
d
istan
c
e
lea
rn
in
g
a
p
p
ro
a
c
h
a
c
h
iev
e
s
b
e
tt
e
r
p
e
rf
o
rm
a
n
c
e
th
a
n
th
e
m
o
st
a
d
v
a
n
c
e
d
m
e
th
o
d
s.
F
u
rth
e
rm
o
re
,
th
is
m
e
th
o
d
is
in
se
n
siti
v
e
to
n
o
rm
a
li
z
a
ti
o
n
c
o
m
p
a
re
d
to
c
o
sin
e
sc
o
re
s.
T
h
is
m
e
th
o
d
is v
e
r
y
e
ffe
c
ti
v
e
i
n
th
e
c
a
se
o
f
li
m
it
e
d
train
in
g
d
a
ta.
T
h
e
m
o
d
if
ied
su
p
e
rv
ise
d
i
-
v
e
c
to
r
b
a
se
d
A
S
R
s
y
ste
m
is
e
v
a
lu
a
ted
o
n
th
e
NIST
S
RE
2
0
0
8
d
a
tab
a
se
.
T
h
e
b
e
st
p
e
rf
o
r
m
a
n
c
e
o
f
th
e
c
o
m
b
in
e
d
c
o
sin
e
sc
o
re
EE
R
1
.
7
6
7
%
o
b
tai
n
e
d
u
si
n
g
L
DA
2
0
0
+
NCA
2
0
0
+
L
D
A
2
0
0
,
a
n
d
th
e
b
e
st
p
e
rf
o
r
m
a
n
c
e
o
f
Ba
y
e
s_
d
m
l
E
ER
1
.
7
7
5
%
o
b
tai
n
e
d
u
si
n
g
L
D
A
2
0
0
+
NC
A2
0
0
+
L
DA
1
0
0
.
Ba
y
e
sia
n
_
d
m
l
o
v
e
rc
o
m
e
s
th
e
c
o
m
b
in
e
d
n
o
rm
o
f
c
o
sin
e
s
c
o
re
s a
n
d
is t
h
e
b
e
st
re
su
lt
o
f
th
e
sh
o
rt2
-
sh
o
rt3
c
o
n
d
i
ti
o
n
re
p
o
rt
f
o
r
NIST
S
RE
2
0
0
8
d
a
ta.
K
ey
w
o
r
d
s
:
Au
to
m
a
tic
s
p
ea
k
er
r
ec
o
g
n
itio
n
(
A
S
R
)
L
a
n
g
u
a
g
e
r
ec
o
g
n
it
io
n
e
v
alu
a
tio
n
(
L
R
E
)
L
i
n
ea
r
d
is
cr
i
m
i
n
atio
n
an
al
y
s
is
(
L
D
A
)
Neig
h
b
o
r
h
o
o
d
c
om
p
o
n
en
t
an
al
y
s
is
(
N
C
A
)
P
h
o
n
e
r
ec
o
g
n
itio
n
an
d
l
a
n
g
u
a
g
e
m
o
d
ell
in
g
(
P
R
L
M)
Co
p
y
rig
h
t
©
2
0
1
9
In
stit
u
te o
f
A
d
v
a
n
c
e
d
E
n
g
i
n
e
e
rin
g
a
n
d
S
c
ien
c
e
.
Al
l
rig
h
ts re
se
rv
e
d
.
C
o
r
r
e
s
p
o
nd
ing
A
uth
o
r
:
Sat
y
a
n
an
d
Si
n
g
h
,
C
o
lleg
e
o
f
E
n
g
in
ee
r
i
n
g
Scie
n
c
e
an
d
T
ec
h
n
o
lo
g
y
,
Fij
i N
atio
n
al
Un
i
v
er
s
i
t
y
,
Fij
i I
s
lan
d
,
Su
v
a
.
E
m
ail: Sa
t
y
a
n
a
n
d
.
s
in
g
h
@
f
n
u
.
ac
.
f
j
1.
I
NT
RO
D
UCT
I
O
N
Dis
ti
n
g
u
i
s
h
an
i
n
d
iv
id
u
al
b
y
h
is
o
r
h
er
v
o
ice
i
s
a
Go
d
g
i
f
ted
h
u
m
an
q
u
alit
y
m
o
s
t
ta
k
e
as
s
u
r
e
ca
s
e
i
n
r
eg
u
lar
h
u
m
a
n
-
to
-
h
u
m
a
n
ass
o
ciatio
n
/
co
r
r
esp
o
n
d
en
ce
.
T
al
k
in
g
s
o
m
eb
o
d
y
v
ia
p
h
o
n
e
m
o
r
e
o
f
ten
th
a
n
n
o
t
s
tar
ts
b
y
d
is
tin
g
u
is
h
i
n
g
w
h
o
is
talk
i
n
g
an
d
,
at
t
h
e
m
i
n
i
m
u
m
i
n
i
n
s
t
a
n
ce
s
o
f
r
ec
o
g
n
izab
le
s
p
ea
k
er
s
,
a
s
u
b
j
ec
tiv
e
co
n
f
ir
m
atio
n
b
y
th
e
a
u
d
ien
ce
t
h
at
t
h
e
p
er
s
o
n
ali
t
y
is
r
ig
h
t
an
d
t
h
e
d
is
cu
s
s
io
n
ca
n
co
n
ti
n
u
e.
A
S
R
f
r
a
m
e
w
o
r
k
s
h
av
e
d
ev
elo
p
ed
as
a
v
i
tal
m
et
h
o
d
f
o
r
au
t
h
en
tic
atin
g
i
n
d
iv
id
u
alit
y
i
n
s
e
v
er
al
o
n
lin
e
ap
p
licatio
n
s
an
d
i
n
ad
d
itio
n
a
ll
i
n
al
l
b
u
s
i
n
es
s
co
llab
o
r
atio
n
s
,
cr
i
m
i
n
o
l
o
g
y
,
w
h
at
's
m
o
r
e,
la
w
e
n
f
o
r
ce
m
en
t.
A
s
id
e
f
r
o
m
in
d
iv
id
u
al
v
a
lid
atio
n
f
o
r
ac
ce
s
s
co
n
tr
o
l,
s
p
ea
k
er
r
ec
o
g
n
it
io
n
is
a
v
ital
d
e
v
ice
in
la
w
i
m
p
le
m
en
tatio
n
,
n
atio
n
a
l
s
ec
u
r
it
y
,
an
d
leg
a
l scie
n
ce
s
[1
,
2]
.
E
x
tr
ao
r
d
in
ar
il
y
,
p
eo
p
le
r
o
u
ti
n
el
y
id
e
n
ti
f
y
p
eo
p
le
b
y
t
h
eir
v
o
ices
w
it
h
s
tr
ik
i
n
g
p
r
ec
is
io
n
,
p
ar
ticu
lar
l
y
w
h
e
n
th
e
le
v
el
o
f
co
m
m
o
n
alit
y
w
it
h
t
h
e
s
u
b
j
ec
t
is
h
i
g
h
.
No
r
m
al
l
y
,
e
v
e
n
a
s
h
o
r
t
n
o
n
li
n
g
u
i
s
tic
li
n
e,
f
o
r
ex
a
m
p
le,
a
la
u
g
h
q
u
ie
tl
y
,
is
s
u
f
f
icie
n
t
f
o
r
u
s
to
r
ec
o
g
n
ize
a
w
ell
-
k
n
o
w
n
in
d
i
v
id
u
a
l
[
3
]
.
A
SR
r
esear
ch
g
r
o
u
p
s
h
a
v
e
p
r
o
d
u
ce
d
d
if
f
er
e
n
t
tec
h
n
iq
u
es
p
r
ett
y
m
u
c
h
au
to
n
o
m
o
u
s
l
y
f
o
r
f
iv
e
d
ec
ad
es
[
4
]
.
On
th
e
o
th
er
h
an
d
,
n
ati
v
e
s
p
ea
k
er
r
ec
o
g
n
iti
o
n
is
a
r
e
g
u
lar
ca
p
ac
it
y
o
f
p
e
o
p
le
w
h
ic
h
i
s
,
o
n
o
cc
asio
n
,
e
x
tr
e
m
e
l
y
ex
ac
t
an
d
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2
0
8
8
-
8708
B
a
ye
s
ia
n
d
is
ta
n
ce
metric le
a
r
n
in
g
a
n
d
its
a
p
p
lica
tio
n
in
a
u
t
o
m
a
tic
s
p
ea
ke
r
r
ec
o
g
n
itio
n
.
..
(
S
a
tya
n
a
n
d
S
in
g
h
)
2961
ef
f
icien
t.
L
a
test
s
tu
d
ie
s
o
n
m
i
n
d
i
m
a
g
i
n
g
[
5
,
6
]
h
av
e
u
n
co
v
er
ed
n
u
m
er
o
u
s
s
u
b
tle
e
le
m
e
n
t
s
o
n
h
o
w
to
p
er
f
o
r
m
p
s
y
ch
o
lo
g
ical
b
ased
s
p
ea
k
er
r
ec
o
g
n
itio
n
,
w
h
ic
h
m
a
y
m
o
ti
v
a
te
n
e
w
h
ea
d
i
n
g
s
f
o
r
A
S
R
ap
p
r
o
ac
h
.
T
o
b
eg
in
w
i
th
,
co
n
s
id
er
in
g
th
e
g
e
n
er
al
ex
p
lo
r
atio
n
ar
ea
,
it
w
o
u
ld
b
e
h
elp
f
u
l
to
ill
u
m
in
a
t
e
w
h
at
is
in
co
r
p
o
r
ated
b
y
t
h
e
ter
m
s
p
ea
k
er
r
ec
o
g
n
it
io
n
,
w
h
ic
h
c
o
m
p
r
i
s
es
o
f
t
w
o
o
p
tio
n
u
n
d
er
tak
in
g
s
:
s
p
ea
k
er
id
en
ti
f
icatio
n
a
n
d
v
er
i
f
icatio
n
.
I
n
s
p
ea
k
er
r
ec
o
g
n
itio
n
,
th
e
a
s
s
i
g
n
m
e
n
t
is
to
d
is
ti
n
g
u
is
h
an
u
n
k
n
o
w
n
s
p
ea
k
er
f
r
o
m
th
e
d
atab
ase
o
f
k
n
o
w
n
s
p
ea
k
er
s
.
T
h
er
e
ar
e
tw
o
k
i
n
d
s
o
f
ASR
s
y
s
te
m
o
p
en
-
s
et
a
n
d
clo
s
e
-
s
e
t.
I
f
all
s
p
ea
k
er
s
ar
e
k
n
o
w
n
i
n
a
s
et
t
h
en
it
is
ca
lled
as
clo
s
ed
-
s
et
o
n
th
e
o
th
er
h
a
n
d
,
if
t
h
e
te
s
t
s
p
ea
k
er
co
u
ld
lik
e
w
i
s
e,
b
e
f
r
o
m
o
u
t
s
id
e
t
h
e
p
r
ed
ef
i
n
e
d
k
n
o
w
n
s
p
ea
k
er
s
et,
th
i
s
t
u
r
n
s
in
to
a
t
h
e
o
p
en
-
s
et
s
it
u
atio
n
,
a
n
d
,
alo
n
g
t
h
e
s
e
lin
es
,
a
w
o
r
ld
m
o
d
el
o
r
u
n
i
v
er
s
al
b
ac
k
g
r
o
u
n
d
m
o
d
el
(
UB
M)
[
7
]
is
r
eq
u
ir
ed
.
Sp
ea
k
er
id
en
ti
f
icatio
n
ca
n
b
e
b
ased
o
n
a
v
o
ice
s
tr
ea
m
t
h
a
t
is
co
n
te
n
t
d
ep
en
d
e
n
t
o
r
co
n
ten
t
f
r
ee
.
T
h
is
is
m
o
r
e
s
i
g
n
if
ica
n
t
i
n
s
p
ea
k
er
-
v
er
if
icatio
n
s
y
s
te
m
s
i
n
w
h
ic
h
a
clai
m
ed
p
er
s
o
n
s
p
ea
k
s
a
p
r
ed
ef
in
ed
tex
t,
lik
e
a
p
as
s
w
o
r
d
o
r
p
er
s
o
n
al
id
en
ti
f
icatio
n
n
u
m
b
er
(
P
I
N)
,
to
ac
ce
s
s
s
y
s
te
m
.
All
t
h
r
o
u
g
h
t
h
i
s
p
ap
er
,
th
e
e
m
p
h
asi
s
w
ill b
e
o
n
te
x
t
-
i
n
d
ep
en
d
en
t a
u
to
m
at
ic
s
p
ea
k
er
r
ec
o
g
n
itio
n
s
y
s
te
m
s
.
A
f
e
w
alg
o
r
it
h
m
ic
an
d
its
co
m
p
u
tat
io
n
al
ad
v
a
n
ce
s
h
av
e
e
m
p
o
w
er
ed
n
o
te
w
o
r
th
y
A
S
R
p
er
f
o
r
m
an
c
e
in
th
e
c
u
tti
n
g
ed
g
e.
A
p
p
r
o
ac
h
es
u
tili
z
in
g
p
h
o
n
o
tact
ic
d
ata,
p
h
o
n
e
m
e
r
ec
o
g
n
izer
f
o
llo
w
ed
b
y
la
n
g
u
a
g
e
m
o
d
els
P
h
o
n
e
R
e
co
g
n
itio
n
a
n
d
L
a
n
g
u
a
g
e
Mo
d
elin
g
(
P
R
L
M)
p
ar
allel
P
R
L
M
,
h
av
e
b
ee
n
ap
p
ea
r
ed
to
b
e
v
er
y
ef
f
ec
tiv
e
[
8
]
.
I
n
th
i
s
p
h
o
n
o
tac
tic
d
e
m
o
n
s
tr
ati
n
g
s
tr
u
ctu
r
e,
a
n
ar
r
an
g
e
m
e
n
t
o
f
to
k
en
ize
s
i
s
u
tili
ze
d
to
in
ter
p
r
et
th
e
s
p
ee
c
h
i
n
f
o
r
m
atio
n
in
to
to
k
en
s
tr
i
n
g
s
o
r
cr
o
s
s
s
ec
t
io
n
s
w
h
ic
h
ar
e
later
s
co
r
ed
b
y
n
-
g
r
a
m
d
ialec
t
m
o
d
el
s
[
9
]
o
r
m
ap
p
ed
in
to
a
s
ac
k
o
f
tr
ig
r
a
m
s
h
i
g
h
lig
h
t
v
ec
t
o
r
f
o
r
s
u
p
p
o
r
t v
ec
to
r
m
ac
h
in
e
(
SVM)
.
I
n
f
ac
t
tr
ad
itio
n
al
Hid
d
en
Ma
r
k
o
v
m
o
d
el
(
HM
M)
b
ased
s
p
ea
k
er
v
er
i
f
icatio
n
is
g
e
n
er
all
y
b
ased
o
n
th
e
to
k
en
izer
clas
s
m
o
d
el,
all
to
k
e
n
izatio
n
s
co
n
n
ec
ted
h
er
e
to
m
a
k
e
a
s
y
s
te
m
[
10
]
,
f
o
r
ex
a
m
p
le,
Gau
s
s
ia
n
Mix
t
u
r
e
Mo
d
el
(
GM
M)
to
k
e
n
izatio
n
[
11
]
,
u
n
i
v
er
s
al
p
h
o
n
e
r
ec
o
g
n
itio
n
(
UP
R
)
[
12
]
,
ar
tic
u
lato
r
th
e
p
r
o
p
er
ty
b
ased
m
e
th
o
d
o
lo
g
y
[
1
3
]
,
d
ee
p
n
eu
r
al
n
et
w
o
r
k
b
ased
telep
h
o
n
e
r
ec
o
g
n
izer
[
1
4
]
,
j
u
s
t to
g
iv
e
s
o
m
e
ex
a
m
p
les.
W
ith
t
h
e
p
r
ese
n
tatio
n
o
f
s
h
i
f
t
ed
d
elta
-
ce
p
s
tr
al
(
S
DC
)
ac
o
u
s
tic
co
m
p
o
n
e
n
t
s
[
1
5
]
,
p
r
o
m
is
in
g
r
es
u
lt
s
u
tili
zi
n
g
th
e
GM
M
s
y
s
te
m
with
th
e
v
ar
iab
le
i
n
v
e
s
ti
g
atio
n
[1
6
,
1
7
]
,
s
u
p
er
v
ec
to
r
m
o
d
el
[
1
8
]
.
Fu
r
th
er
m
o
r
e,
m
ax
i
m
u
m
m
u
t
u
al
i
n
f
o
r
m
atio
n
(
MM
I
)
b
ased
d
is
cr
i
m
i
n
ati
v
e
tr
ain
i
n
g
[
1
9
]
h
av
e
li
k
e
w
is
e
b
ee
n
ac
c
o
u
n
ted
f
o
r
L
I
D.
I
n
t
h
is
w
o
r
k
,
I
co
n
ce
n
tr
at
e
o
n
th
e
ac
o
u
s
tic
lev
el
f
r
a
m
e
wo
r
k
s
.
A
co
u
s
tic
-
p
h
o
n
etic
m
et
h
o
d
o
lo
g
y
,
w
h
ic
h
i
s
n
o
r
m
all
y
tak
e
n
b
y
s
p
ec
ialis
ts
p
r
ep
ar
ed
o
n
th
i
s
,
r
eq
u
ir
es
q
u
an
tita
tiv
e
ac
o
u
s
tic
e
s
ti
m
atio
n
s
f
r
o
m
s
p
ee
c
h
s
i
g
n
a
l te
s
t
s
,
an
d
b
ased
o
n
f
ac
t e
x
a
m
in
a
tio
n
o
f
th
e
r
es
u
lt.
B
y
a
n
d
lar
g
e,
co
m
p
ar
ab
le
p
h
o
n
etic
u
n
its
ar
e
r
e
m
o
v
ed
f
r
o
m
t
h
e
k
n
o
w
n
an
d
ad
d
r
ess
ed
s
p
ee
c
h
s
i
g
n
al,
a
n
d
d
i
f
f
er
e
n
t
ac
o
u
s
tic
p
ar
a
m
eter
s
m
ea
s
u
r
e
d
f
r
o
m
th
e
s
e
p
o
r
tio
n
s
ar
e
ev
alu
a
ted
.
T
h
e
L
o
g
is
t
ic
R
e
g
r
ess
io
n
(
L
R
)
ca
n
b
e
h
elp
f
u
ll
y
u
tili
ze
d
as a
p
ar
t o
f
th
is
m
et
h
o
d
o
lo
g
y
s
i
n
ce
it
d
ep
en
d
s
o
n
n
u
m
er
ical
p
ar
a
m
eter
s
[
20
].
I
n
s
p
ite
o
f
th
e
f
ac
t
th
at
t
h
e
ac
o
u
s
tic
-
p
h
o
n
etic
a
m
et
h
o
d
o
lo
g
y
is
a
m
o
r
e
tar
g
et
ap
p
r
o
ac
h
,
it
h
as
s
o
m
e
s
u
b
j
ec
tiv
e
co
m
p
o
n
en
t
s
.
Fo
r
in
s
ta
n
ce
,
a
n
ac
o
u
s
tic
–
p
h
o
n
et
ician
m
a
y
d
is
ti
n
g
u
is
h
s
p
ee
c
h
s
i
g
n
al
a
s
b
ein
g
in
f
lu
e
n
ce
d
b
y
an
x
iet
y
a
n
d
a
f
te
r
th
at
p
er
f
o
r
m
t
h
e
o
b
j
ec
tiv
e
e
x
a
m
in
at
io
n
.
I
n
a
n
y
,
th
e
ca
s
e,
w
h
et
h
er
t
h
e
s
p
ea
k
er
w
a
s
r
ea
ll
y
u
n
d
er
an
x
iet
y
at
th
at
m
i
n
u
te
is
a
s
u
b
j
ec
tiv
e
a
m
o
u
n
t
co
n
tr
o
lled
b
y
th
e
in
s
p
ec
to
r
th
r
o
u
g
h
h
is
o
r
h
e
r
ex
p
er
ien
ce
.
A
s
o
n
t
h
e
d
ate,
a
g
g
r
e
g
ate
v
ar
iab
ilit
y
i
-
v
ec
to
r
ASR
m
o
d
elin
g
h
as
ac
h
ie
v
ed
cr
itical
co
n
s
id
er
atio
n
in
b
o
th
L
I
D
a
n
d
SV
ar
ea
s
b
ec
au
s
e
o
f
i
ts
r
e
m
ar
k
ab
le
ef
f
icie
n
c
y
,
les
s
s
y
s
te
m
co
m
p
le
x
it
y
a
n
d
co
m
p
ac
t
s
y
s
te
m
in
s
ize.
I
n
i
-
v
ec
to
r
m
o
d
elin
g
,
i
n
itial,
a
s
o
litar
y
ele
m
en
t i
n
v
es
tig
atio
n
is
u
tili
ze
d
as a
f
r
o
n
t
e
n
d
to
p
r
o
d
u
ce
a
lo
w
d
i
m
en
s
io
n
al
a
g
g
r
eg
ate
v
ar
iab
ilit
y
s
p
ac
e
w
h
ic
h
to
g
et
h
er
m
o
d
els
d
ialec
t,
s
p
ea
k
er
an
d
ch
an
n
el
v
ar
iab
ilit
y
all
to
g
eth
er
.
A
t
th
at
p
o
in
t,
i
n
s
id
e
th
i
s
i
-
v
ec
to
r
s
p
ac
e,
v
ar
iab
ilit
y
co
s
t
s
tec
h
n
iq
u
es,
f
o
r
ex
a
m
p
le,
W
ith
i
n
-
C
las
s
C
o
v
ar
ian
ce
No
r
m
ali
za
tio
n
(
W
C
C
N)
,
L
i
n
ea
r
Dis
cr
i
m
i
n
ati
v
e
ex
a
m
in
at
io
n
(
L
D
A
)
an
d
Nu
is
a
n
ce
A
t
tr
ib
u
te
P
r
o
j
ec
tio
n
(
NA
P
)
[
18
]
,
ar
e
p
er
f
o
r
m
ed
to
d
i
m
i
n
is
h
t
h
e
v
ar
iab
ilit
y
f
o
r
co
n
s
eq
u
e
n
t
m
o
d
el
in
g
(
e.
g
.
,
u
til
izin
g
SVM,
L
R
,
a
n
d
n
e
u
r
al
n
et
w
o
r
k
a
n
d
p
r
o
b
ab
ilis
tic
li
n
e
ar
d
is
cr
i
m
i
n
ate
a
n
al
y
s
i
s
(
P
L
D
A
)
f
o
r
L
a
n
g
u
a
g
e
I
d
en
tif
icat
io
n
(
L
I
D)
.
I
n
th
i
s
p
ap
er
,
th
e
co
n
v
e
n
tio
n
a
l
i
-
v
ec
to
r
s
ar
e
s
tr
etch
ed
o
u
t
to
lab
el
r
eg
u
lar
ized
r
eg
u
lated
i
-
v
ec
to
r
s
b
y
co
n
n
ec
ti
n
g
t
h
e
m
ar
k
ed
v
ec
to
r
an
d
t
h
e
s
tr
aig
h
t
r
elap
s
e
lattic
e
to
w
ar
d
t
h
e
e
n
d
o
f
th
e
m
ea
n
s
u
p
er
v
ec
to
r
w
h
at
's
m
o
r
e,
t
h
e
i
-
v
ec
to
r
ele
m
e
n
t
s
t
ac
k
in
g
n
et
w
o
r
k
,
s
ep
ar
atel
y
[
21
,
22
].
I
ca
n
let
th
e
ad
d
ed
n
a
m
e
v
ec
to
r
b
e
th
e
p
ar
am
eter
v
ec
to
r
th
at
I
n
ee
d
to
p
er
f
o
r
m
an
d
r
elap
s
e
w
it
h
ag
e
p
ar
alin
g
u
i
s
tic
m
ea
s
u
r
es
to
m
ak
e
t
h
e
p
r
o
p
o
s
ed
s
y
s
te
m
r
ea
s
o
n
ab
le
f
o
r
r
eg
r
es
s
io
n
.
T
h
e
ex
p
lan
a
t
io
n
b
eh
i
n
d
u
tili
zi
n
g
a
lin
ea
r
r
e
g
r
ess
io
n
m
atr
ix
W
i
s
th
at
n
u
m
er
o
u
s
b
ac
k
en
d
cla
s
s
i
f
icat
io
n
m
o
d
u
les
in
L
I
D
an
d
SV
i
s
lin
ea
r
.
A
d
d
itio
n
a
ll
y
,
if
th
e
r
eg
r
ess
io
n
co
n
n
ec
t
io
n
is
n
o
t
li
n
ea
r
,
I
ca
n
u
s
e
n
o
n
-
l
in
ea
r
m
ap
p
in
g
as
a
p
r
ep
r
o
ce
s
s
i
n
g
v
e
n
t
u
r
e
b
ef
o
r
e
cr
ea
ti
n
g
t
h
e
m
ar
k
v
ec
to
r
s
.
T
h
e
co
m
m
i
t
m
e
n
t
w
ei
g
h
t
o
f
ea
ch
s
u
p
er
v
ec
to
r
m
ea
s
u
r
e
m
en
t
a
n
d
ev
er
y
o
b
j
ec
tiv
e
cla
s
s
in
t
h
e
tar
g
et
ca
p
ac
it
y
is
co
n
s
eq
u
e
n
tl
y
ascer
tai
n
ed
b
y
it
er
ativ
e
p
r
ep
ar
in
g
.
T
h
e
co
n
v
e
n
t
io
n
al
i
-
v
ec
to
r
f
r
a
m
e
w
o
r
k
s
er
v
es a
s
o
u
r
b
aseli
n
e.
As
a
f
i
n
al
p
o
in
t,
m
o
ti
v
a
ted
b
y
th
e
ac
h
ie
v
e
m
e
n
t
o
f
s
tr
o
n
g
w
o
r
k
s
f
o
r
co
r
r
u
p
ted
in
f
o
r
m
atio
n
b
ased
SV
j
o
b
s
,
I
lik
e
w
i
s
e
co
n
s
id
er
ed
Ga
m
m
ato
n
e
f
r
eq
u
e
n
c
y
C
ep
s
tr
al
c
o
ef
f
icie
n
t
s
(
GF
C
C
)
f
ea
t
u
r
es
a
n
d
th
e
s
p
ec
tr
o
te
m
p
o
r
al
Gab
o
r
f
ea
tu
r
es
f
o
r
p
o
w
er
f
u
l
L
I
D
as
s
ig
n
m
en
t
o
n
th
e
co
r
r
u
p
ted
i
n
f
o
r
m
atio
n
a
s
ex
tr
a
ex
ec
u
t
io
n
e
n
h
a
n
ce
r
s
tep
s
.
A
t
th
e
p
o
in
t
w
h
en
co
m
b
in
ed
with
c
u
s
to
m
ar
y
MF
C
C
an
d
S
DC
s
p
ea
k
er
s
p
ec
if
i
c
f
ea
t
u
r
e
b
ased
f
r
a
m
e
w
o
r
k
s
,
th
e
g
en
er
al
f
r
a
m
e
w
o
r
k
e
x
ec
u
tio
n
w
a
s
f
u
r
th
er
u
p
g
r
ad
ed
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8708
I
n
t J
E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
9
,
No
.
4
,
A
u
g
u
s
t 2
0
1
9
:
2
9
6
0
-
2967
2962
2.
T
H
E
B
AS
E
L
I
NE
i
-
V
E
C
T
O
R
M
O
DE
L
I
NG
L
et
u
s
co
n
s
id
er
a
m
o
d
el
λ
o
f
C
co
m
p
o
n
e
n
t
GM
M
UB
M
w
it
h
λ
c
=
{
p
c
,
μ
c
,
∑
c
}
w
h
er
e
c
=
1
,
.
,
,
.
.
C
an
d
th
e
s
p
ee
ch
s
i
g
n
a
l
Y=
{
y
1
,
y
2
…….
.
,
y
L
}
h
a
v
i
n
g
L
f
ea
tu
r
e
s
eq
u
en
ce
s
.
UB
M
b
ased
o
n
B
W
s
tatis
tic
s
co
m
p
u
ted
as f
o
llo
w
s
;
N
c
=
∑
P
L
t
=
1
(
c
|
y
t
,
λ
)
(
1
)
w
h
er
e
GM
M
co
m
p
o
n
e
n
t is
c
=
1
,
…
.
,
…
,
…
C
.
an
d
y
t
o
n
λ
c
is
o
cc
u
p
an
c
y
p
r
o
b
ab
ilit
y
co
m
p
o
n
e
n
t
P
(
c
|
y
t
,
)
.
F
c
=
∑
P
L
t
=
1
(
c
|
y
t
,
λ
)
(
y
t
−
μ
c
)
(
2
)
I
ca
n
g
en
er
ate
t
h
e
co
r
r
esp
o
n
d
in
g
ce
n
tr
al
m
ea
n
s
u
p
er
v
ec
to
r
F
̌
b
y
co
n
ca
ten
at
in
g
all
to
g
e
th
er
F
̌
c
as f
o
llo
w
s
;
F
̌
c
=
F
c
=
∑
P
L
t
=
1
(
c
|
y
t
,
λ
)
(
y
t
−
μ
c
)
∑
P
L
t
=
1
(
c
|
y
t
,
λ
)
(
3
)
F
̌
c
ca
n
b
e
p
r
o
j
ec
te
d
o
n
r
ec
tan
g
u
lar
to
tal
v
ar
iab
ilit
y
m
atr
i
x
T
(
lo
w
r
an
k
f
ac
to
r
lo
ad
in
g
)
an
d
i
-
v
ec
to
r
x
as
f
o
llo
w
s
;
F
̌
c
→
T
x
(
4
)
T
h
e
to
tal
v
ar
iab
ilit
y
m
atr
i
x
T
w
it
h
C
co
m
p
o
n
en
t
GM
M
an
d
ac
o
u
s
tic
f
ea
t
u
r
es
in
D
d
im
e
n
s
io
n
ca
n
b
e
r
ep
r
esen
ted
as
C
DX
K
m
atr
ix
s
i
m
il
ar
to
eig
en
v
o
ice
m
atr
ix
V
.
B
y
co
n
s
id
er
in
g
F
̌
c
an
d
th
e
i
-
v
ec
t
o
r
ca
n
b
e
co
m
p
u
ted
as f
o
llo
w
s
;
x
=
{
I
+
T
t
Σ
−
1
NT
}
−
1
T
−
1
Σ
−
1
N
F
̌
(
5
)
w
h
er
e
N
is
CD
X
CD
d
i
m
en
s
io
n
d
ia
g
o
n
al
m
atr
i
x
.
T
o
r
e
d
u
ce
th
e
v
ar
iab
ilit
y
I
ap
p
lied
t
w
o
ch
an
n
el
co
m
p
e
n
s
at
io
n
m
et
h
o
d
s
in
to
tal
v
ar
iab
ilit
y
s
p
ac
e
(
i)
L
D
A
a
n
d
(
ii)
W
C
C
N.
L
D
A
m
in
i
m
izes
in
tr
a
-
clas
s
v
ar
ian
ce
an
d
W
C
C
N
n
o
r
m
al
iz
es
th
e
co
s
in
e
k
er
n
el.
L
e
t
u
s
co
n
s
id
er
t
w
o
i
-
v
ec
to
r
1
an
d
2
th
en
co
s
in
e
k
er
n
el
is
d
ef
in
ed
as
f
o
llo
w
s
to
ad
ap
t e
ith
er
P
L
D
A
o
r
SVM
class
if
ier
.
k
(
x
1
,
x
2
)
=
〈
x
1
,
x
2
〉
‖
x
1
‖
1
‖
x
2
‖
2
(
6
)
3.
I
M
P
L
E
M
E
NT
AT
I
O
N
O
F
M
O
DIFIE
D
AN
D
SUPE
RVI
SE
D
I
-
V
E
C
T
O
R
Usi
n
g
a
co
m
b
i
n
atio
n
o
f
SV
M
class
i
f
ier
s
a
n
d
GM
M
s
u
p
e
r
v
ec
to
r
s
h
a
v
e
b
ee
n
an
e
x
tr
e
m
el
y
f
r
u
it
f
u
l
m
et
h
o
d
o
lo
g
y
f
o
r
A
S
R
.
Ob
s
e
r
v
in
g
t
h
e
w
a
y
th
at
t
h
e
c
h
an
n
el
co
n
s
id
er
s
to
co
n
tai
n
s
p
ea
k
er
-
d
ep
en
d
an
t
d
ata,
th
e
s
p
ea
k
er
a
n
d
ch
an
n
el
ele
m
e
n
ts
w
er
e
co
n
s
o
lid
ated
in
to
a
s
o
litar
y
s
p
ac
e
ter
m
ed
th
e
to
tal
v
ar
iab
ilit
y
s
p
ac
e.
3
.
1
.
L
a
bel
-
re
g
ula
rize
d su
pe
rv
is
e
d i
-
v
ec
t
o
r
L
et
u
s
a
s
s
u
m
e
t
h
at
t
h
e
h
id
d
en
v
ar
iab
le
i
-
v
ec
to
r
is
g
e
n
er
atio
n
m
ea
n
s
u
p
er
v
ec
to
r
.
T
h
e
s
tep
s
in
v
o
lv
e
d
in
lab
el
-
r
e
g
u
lar
izatio
n
as
f
o
llo
w
s
:
Step
I
.
C
o
m
p
u
te
m
u
l
tiv
ar
iate
Gau
s
s
ia
n
d
is
tr
ib
u
tio
n
f
o
r
j
th
u
tter
an
ce
s
o
f
P
(
x
j
)
=
(
0
,
I
)
,
P
(
F
̌
j
|
x
j
)
=
(
T
x
j
,
N
j
−
1
Σ
)
.
W
h
er
e
x
j
=
i
-
v
ec
to
r
o
f
j
th
u
tter
an
ce
s
.
Step
I
I
.
C
o
m
p
u
tatio
n
o
f
p
o
s
ter
io
r
d
is
tr
ib
u
tio
n
o
f
h
id
d
en
v
ar
ia
b
l i
-
v
ec
to
r
P
(
x
j
|
F
̌
j
)
=
{
(
I
+
T
t
Σ
−
1
N
j
T
)
−
1
T
t
Σ
−
1
N
j
F
̌
j
,
(
I
+
T
t
Σ
−
1
N
j
T
)
−
1
}
w
h
er
e
N
j
=
N
v
ec
to
r
.
Step
I
I
I
.
C
o
m
p
u
tatio
n
o
f
d
is
cr
i
m
i
n
ati
v
e
i
-
v
ec
to
r
,
P
(
x
j
)
=
(
0
,
I
)
.
Step
I
V.
R
eg
u
lar
izatio
n
o
f
lab
el
in
f
o
r
m
atio
n
,
P
[
(
F
̌
j
L
j
)
|
x
j
]
=
[
(
T
x
j
W
x
j
)
]
,
[
(
N
j
−
1
Σ
1
n
j
−
1
Σ
2
)
]
w
h
er
e
F
̌
j
=
Me
an
s
u
p
er
v
ec
to
r
,
L
j
=
L
ab
el
v
ec
to
r
,
Σ
1
=
C
D
d
i
m
e
n
s
io
n
m
ea
n
s
u
p
er
v
ec
to
r
,
Σ
2
=
M
d
i
m
en
s
io
n
lab
el
v
ec
to
r
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2
0
8
8
-
8708
B
a
ye
s
ia
n
d
is
ta
n
ce
metric le
a
r
n
in
g
a
n
d
its
a
p
p
lica
tio
n
in
a
u
t
o
m
a
tic
s
p
ea
ke
r
r
ec
o
g
n
itio
n
.
..
(
S
a
tya
n
a
n
d
S
in
g
h
)
2963
Step
V.
Desig
n
i
n
g
o
f
t
w
o
s
u
p
e
r
v
is
ed
lab
el
v
ec
to
r
s
T
y
p
e
1
:
L
i
,
j
=
{
1
for
c
l
a
s
s
i
0
othe
r
wi
s
e
,
T
h
e
class
lab
el
w
ill
b
e
co
r
r
e
ctly
clas
s
i
f
ied
b
y
r
eg
r
ess
io
n
m
atr
i
x
W
.
M
is
d
en
o
ted
as
th
e
d
i
m
e
n
s
io
n
alit
y
o
f
lab
el
v
ec
to
r
L
j
.
T
o
tal
n
u
m
b
er
o
f
s
p
ea
k
er
s
i
n
d
atab
ase
to
r
ec
o
g
n
ize
=
H
th
e
n
L
j
is
H
(
H
=
M
)
d
im
e
n
s
io
n
b
in
ar
y
v
ec
to
r
.
H
−
1
n
u
m
b
er
s
o
f
ele
m
e
n
t
s
w
i
ll
h
a
v
e
“
0
”
an
d
o
n
e
elem
e
n
t
w
ill
h
a
v
e
v
alu
e
"
1"
.
T
y
p
e
2
:
L
j
=
X
̅
sj
,
W
=
I
.
T
h
e
last
iter
atio
n
X
̅
sj
s
p
ec
if
y
th
e
s
a
m
p
le
m
ea
n
v
ec
to
r
an
d
co
m
p
e
l
r
eg
r
ess
io
n
m
atr
i
x
to
b
e
id
en
tit
y
m
atr
i
x
.
Step
VI
.
C
o
m
p
u
tatio
n
o
f
t
h
e
li
k
eli
h
o
o
d
o
f
th
e
to
tal
tr
ain
i
n
g
u
tter
an
ce
s
o
f
ASR
s
y
s
te
m
is
:
∑
ln
{
P
(
F
̌
j
,
L
j
,
x
j
)
}
=
∑
[
ln
{
P
(
(
F
̌
j
L
j
)
|
x
j
)
}
+
ln
{
P
(
x
j
)
}
]
Γ
j
=
1
Γ
j
=
1
Step
VI
I
.
C
o
m
p
u
tatio
n
o
f
o
b
j
e
ctiv
e
f
u
n
ctio
n
J
m
f
o
r
Ma
x
i
m
u
m
L
i
k
eli
h
o
o
d
(
ML
)
:
J
m
=
∑
[
1
2
x
j
t
x
j
+
1
2
(
F
̌
j
−
T
x
j
)
t
∑
N
j
(
F
̌
j
−
T
x
j
)
+
1
2
(
L
j
−
W
xj
)
t
∑
N
j
(
F
̌
j
−
T
x
j
)
−
1
2
ln
(
|
Σ
1
−
t
|
)
−
−
t
2
−
t
1
Γ
j
=
1
1
2
ln
(
|
Σ
2
−
t
|
)
]
Af
ter
s
i
m
p
li
f
y
in
g
o
b
j
ec
tiv
e
f
u
n
ctio
n
eq
u
a
tio
n
.
J
m
=
∑
[
1
2
x
j
t
x
j
+
1
2
(
A
)
t
∑
N
j
(
A
)
+
1
2
(
B
)
t
∑
N
j
(
B
)
−
1
2
ln
(
|
Σ
1
−
t
|
)
−
1
2
ln
(
|
Σ
2
−
t
|
)
−
t
2
−
t
1
]
Γ
j
=
1
w
h
er
e
A
=
F
̌
j
−
T
x
j
an
d
B
=
L
j
−
W
xj
3
.
2
.
M
o
dified
a
nd
s
i
m
p
lifie
d i
-
v
ec
t
o
r
T
h
e
co
s
in
e
d
is
ta
n
ce
s
co
r
e
is
f
ast
an
d
r
o
b
u
s
t,
b
u
t
ad
d
itio
n
al
co
m
p
u
tatio
n
ar
e
r
eq
u
ir
ed
to
s
t
an
d
ar
d
ize
th
e
s
co
r
e.
A
b
etter
g
en
er
ati
o
n
m
o
d
el
w
ill
f
u
ll
y
s
i
m
u
lat
e
th
e
s
p
ee
ch
d
ata
a
n
d
g
e
n
e
r
ate
s
co
r
es
w
it
h
o
u
t
s
tan
d
ar
d
izatio
n
o
r
ca
lib
r
atio
n
.
Featu
r
e
e
x
tr
ac
tio
n
a
n
d
tr
ain
in
g
o
f
ASR
s
y
s
te
m
b
a
s
ed
o
n
i
-
v
ec
to
r
is
co
m
p
u
tatio
n
all
y
v
er
y
ex
p
an
s
iv
e.
L
et
u
s
co
n
s
id
er
th
e
GM
M
s
i
ze
C
,
f
ea
t
u
r
e
d
i
m
e
n
s
io
n
as
D
an
d
f
ac
to
r
lo
ad
in
g
s
ize
as
K
.
A
s
in
g
le
i
-
v
ec
to
r
g
en
er
atio
n
an
d
it
s
co
m
p
u
tatio
n
co
s
t
is
O
[
(
K
)
3
+
(
K
)
2
.
C
+
(
K
.
C
.
D
)
]
.
T
h
e
m
ai
n
o
b
j
ec
tiv
e
is
to
r
ed
ef
in
e
a
n
d
r
e
w
ei
g
h
t
ea
ch
a
n
d
ev
er
y
s
p
ee
ch
d
ata
w
it
h
m
ea
n
s
u
p
er
v
ec
t
o
r
s
o
th
at
i
m
b
ala
n
ce
i
n
i
n
tr
a
-
s
u
p
er
v
ec
to
r
ca
n
b
e
co
m
p
e
n
s
ated
.
T
h
e
s
tep
s
i
n
v
o
l
v
ed
in
s
i
m
p
li
f
icatio
n
o
f
i
-
v
ec
t
o
r
as f
o
llo
w
s
:
S
tep
I
-
i
-
v
ec
to
r
w
it
h
ap
p
r
o
x
i
m
a
ted
co
m
p
u
tat
io
n
al
co
s
t
O
[
(
K
)
3
+
(
K
)
2
.
C
+
(
K
.
C
.
D
)
]
.
S
tep
II
-
s
u
p
er
v
i
s
ed
i
-
v
ec
to
r
w
ith
ap
p
r
o
x
i
m
ated
co
m
p
u
tat
io
n
al
co
s
t
O
[
(
K
)
3
+
(
K
)
2
.
C
+
{
K
.
(
C
.
D
+
M
)
}
]
.
S
tep
III
-
m
o
d
i
f
ied
i
-
v
ec
to
r
w
it
h
o
u
t I
D
w
it
h
ap
p
r
o
x
i
m
a
ted
co
m
p
u
tatio
n
al
co
s
t
O
[
(
K
)
3
+
(
K
.
C
.
D
)
]
.
S
tep
IV
-
m
o
d
i
f
ied
i
-
v
ec
to
r
w
it
h
I
D
w
it
h
ap
p
r
o
x
i
m
ated
co
m
p
u
tatio
n
al
co
s
t
O
[
(
K
.
C
.
D
)
]
.
S
tep
V
-
m
o
d
i
f
ied
an
d
s
u
p
er
v
is
ed
i
-
v
e
cto
r
w
ith
o
u
t
I
D
w
it
h
ap
p
r
o
x
i
m
ated
co
m
p
u
ta
tio
n
al
co
s
t
O
[
(
K
)
3
+
{
K
.
(
C
.
D
+
M
)
}
]
.
S
tep
VI
-
m
o
d
i
f
ied
an
d
s
u
p
er
v
is
ed
i
-
v
e
cto
r
w
i
th
I
D
w
it
h
ap
p
r
o
x
im
a
ted
co
m
p
u
ta
tio
n
al
co
s
t
[
{
.
(
.
+
)
}
]
.
Fo
r
ASR
ap
p
licatio
n
,
I
u
s
e
a
s
i
m
p
le
co
s
i
n
e
d
i
s
tan
ce
clas
s
i
f
ie
r
o
n
s
i
m
p
li
f
ied
a
n
d
m
o
d
if
ie
d
i
-
v
ec
to
r
o
f
th
e
tar
g
e
t sp
ek
er
u
tte
r
an
ce
s
an
d
test
att
u
r
an
ce
s
w
it
h
d
ec
i
s
io
n
t
h
r
esh
o
ld
as f
o
llo
w
s
.
(
,
)
=
(
)
‖
‖
.
‖
‖
<
≥
(
7
)
3
.
3
.
L
inea
r
d
i
s
cr
i
m
a
m
t
An
a
ly
s
i
s
f
o
r
s
ess
io
n c
o
m
pen
s
a
t
io
n
T
r
ea
tin
g
a
s
i
n
g
le
s
p
ea
k
er
a
s
a
class
L
D
A
atte
m
p
ts
to
d
ef
i
n
e
a
n
e
w
ax
is
to
m
i
n
i
m
ize
i
n
tr
a
-
cla
s
s
v
ar
ian
ce
ca
u
s
ed
b
y
s
e
s
s
io
n
/ch
a
n
n
el
ef
f
ec
t
s
an
d
to
m
a
x
i
m
ize
t
h
e
d
i
f
f
er
e
n
ce
s
b
et
w
ee
n
clas
s
es.
I
n
th
e
r
ep
r
esen
tatio
n
o
f
to
tal
v
ar
iab
ilit
y
,
t
h
er
e
is
n
o
clea
r
c
o
m
p
e
n
s
at
io
n
f
o
r
v
ar
iab
ilit
y
b
e
t
w
ee
n
in
ter
s
es
s
io
n
s
.
Ho
w
e
v
er
,
th
e
lo
w
d
i
m
en
s
io
n
al
r
ep
r
esen
tatio
n
allo
w
s
m
a
k
in
g
tech
n
ical
co
m
p
en
s
atio
n
in
th
e
n
e
w
p
lace
,
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8708
I
n
t J
E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
9
,
No
.
4
,
A
u
g
u
s
t 2
0
1
9
:
2
9
6
0
-
2967
2964
w
it
h
t
h
e
b
en
e
f
it
o
f
les
s
co
m
p
u
tat
io
n
al
co
s
t.
I
u
s
e
a
li
n
ea
r
d
is
cr
i
m
i
n
atio
n
an
a
l
y
s
is
(
L
D
A
)
f
o
r
s
e
s
s
io
n
co
m
p
e
n
s
at
io
n
.
Dis
p
la
y
in
g
s
p
ea
k
er
s
as
th
e
cla
s
s
,
L
D
A
att
e
m
p
ts
to
d
ef
i
n
e
a
n
e
w
a
x
i
s
to
r
ed
u
ce
in
ter
-
clas
s
v
ar
iatio
n
d
u
e
to
s
e
s
s
io
n
/ch
a
n
n
el
ef
f
ec
ts
a
n
d
to
m
ax
i
m
ize
d
i
f
f
er
en
ce
s
b
et
w
ee
n
cla
s
s
e
s
.
I
ca
n
d
ef
i
n
e
t
h
e
p
r
o
b
le
m
o
f
o
p
ti
m
izi
n
g
th
e
L
D
A
to
f
i
n
d
d
ir
ec
tio
n
s
o
f
q
th
at
m
ax
i
m
ize
f
is
h
e
r
cr
iter
ia
J
(
q
)
=
q
t
S
b
q
q
t
S
w
q
⁄
.
B
etw
ee
n
-
clas
s
an
d
w
it
h
in
-
cla
s
s
co
v
ar
ia
n
ce
m
atr
i
c
w
s
ar
e
r
ep
r
esen
ted
as
S
b
=
∑
(
w
s
̅
̅
̅
̅
−
w
̅
)
(
w
s
̅
̅
̅
̅
−
w
̅
)
t
S
s
=
1
an
d
S
w
=
∑
1
n
s
∑
∑
(
w
s
̅
̅
̅
̅
−
w
̅
)
(
w
s
̅
̅
̅
̅
−
w
̅
)
t
S
s
=
1
n
s
i
=
1
S
s
=
1
.
W
h
er
e
n
s
is
t
h
e
n
u
m
b
er
o
f
u
tter
an
ce
s
o
f
ea
ch
s
p
ea
k
er
s
,
w
s
̅
̅
̅
̅
=
(
1
n
s
)
∑
w
i
s
n
s
i
=
1
is
t
h
e
m
ea
n
o
f
i
-
v
ec
to
r
f
o
e
ea
ch
s
p
ea
k
er
,
s
p
ea
k
er
p
o
p
u
lat
io
n
is
w
̅
an
d
to
tal
n
u
m
b
er
o
f
s
p
ea
k
er
s
i
s
S
.
P
r
o
j
ec
tio
n
m
atr
i
x
i
s
co
m
p
o
s
ed
b
y
ei
g
e
n
v
ec
to
r
s
o
f
S
w
−
1
S
b
g
en
er
al
m
atr
i
x
to
g
et
t
h
e
o
p
ti
m
izatio
n
.
3
.
4
.
Dis
t
a
nce
m
a
t
ric
lea
rni
ng
t
ec
hn
iq
ue
Usi
n
g
m
o
d
if
ied
an
d
s
u
p
er
v
i
s
ed
as
a
lo
w
-
d
i
m
e
n
s
io
n
al
r
ep
r
esen
tat
io
n
o
f
t
h
e
li
n
g
u
i
s
tic
e
x
p
r
ess
io
n
,
co
s
in
e
d
is
tan
ce
c
lass
if
ier
s
i
s
u
s
ed
to
m
ea
s
u
r
e
t
h
e
d
is
ta
n
ce
b
et
w
ee
n
cla
s
s
i
f
ied
co
s
i
n
e
a
n
d
th
e
tar
g
et
u
s
er
'
s
ac
ce
n
t
an
d
ac
ce
n
t
u
atio
n
o
f
t
h
e
test
.
T
o
d
ef
in
e
th
e
d
i
s
ta
n
ce
b
et
w
ee
n
t
h
e
v
ec
to
r
s
ai
m
in
g
to
f
i
n
d
a
g
o
o
d
d
is
tan
ce
m
etr
ic
i
n
th
e
f
ea
t
u
r
e
s
p
ac
e,
th
e
r
e
is
an
i
m
p
o
r
ta
n
t is
s
u
e
i
n
cla
s
s
if
ica
tio
n
.
I
n
r
ec
e
n
t
y
ea
r
s
,
e
x
te
n
s
i
v
e
r
esear
ch
h
as
b
ee
n
co
n
d
u
cted
o
n
d
is
ta
n
ce
lear
n
in
g
[
23
].
I
ex
p
lo
r
ed
t
w
o
w
a
y
s
to
d
is
ta
n
ce
lear
n
in
g
m
e
tr
ics
t
h
at
ar
e
s
u
p
er
v
i
s
ed
in
t
h
is
p
ap
er
.
Fro
m
n
o
w
o
n
,
I
w
ill
u
s
e
m
o
d
if
ied
a
n
d
s
u
p
er
v
is
ed
to
r
ep
r
esen
t a
n
u
tter
an
ce
.
3
.
5
.
Co
m
po
nent
a
na
ly
s
is
ba
s
e
d o
n neig
hb
o
rho
o
d
m
et
ho
d
Neig
h
b
o
r
h
o
o
d
C
o
m
p
o
n
e
n
t
An
al
y
s
i
s
(
NC
A
)
n
ea
r
t
h
e
r
an
d
o
m
s
elec
tio
n
r
u
le
lear
n
s
d
is
ta
n
ce
m
etr
ics
to
r
ed
u
ce
th
e
a
v
er
ag
e
cla
s
s
i
f
ica
tio
n
er
r
o
r
.
Usi
n
g
tr
an
s
f
o
r
m
a
t
io
n
m
a
tr
ix
B
,
ea
ch
s
i
m
p
li
f
ie
d
an
d
s
u
p
er
v
is
ed
m
o
d
i
f
ied
an
d
s
u
p
er
v
i
s
ed
ch
o
o
s
es
an
o
t
h
er
s
i
m
p
li
f
ied
an
d
s
u
p
er
v
is
ed
m
o
d
i
f
ied
an
d
s
u
p
er
v
is
ed
as
i
ts
n
eig
h
b
o
r
w
i
th
s
o
m
e
p
r
o
b
ab
ilit
y
,
,
w
h
ic
h
is
d
e
f
in
ed
at
E
u
c
lid
ea
n
d
is
ta
n
ce
in
t
h
e
tr
an
s
f
o
r
m
ed
s
p
ac
e:
p
i
,
j
=
ex
p
(
−
‖
Bw
i
−
Bw
j
‖
2
)
∑
‖
Bw
i
−
Bw
k
‖
2
k
≠
i
,
p
i
,
i
=
0
(
8
)
T
h
e
p
r
o
b
ab
ilit
y
o
f
s
i
m
p
lifie
d
an
d
s
u
p
er
v
is
ed
m
o
d
if
ied
an
d
s
u
p
er
v
is
ed
s
elec
t
a
n
eig
h
b
o
r
w
it
h
th
e
s
a
m
e
s
p
ea
k
er
i
s
=
∑
,
∈
,
w
h
er
e
is
a
s
et
o
f
t
h
e
s
a
m
e
s
p
e
ak
er
m
o
d
i
f
ied
an
d
s
u
p
er
v
is
e
d
s.
T
h
e
p
r
o
j
ec
tio
n
m
atr
i
x
B
s
h
o
u
ld
m
a
x
i
m
ize
th
e
n
u
m
b
er
o
f
e
x
p
ec
ted
s
i
m
p
li
f
ied
a
n
d
s
u
p
er
v
is
ed
m
o
d
if
ied
an
d
s
u
p
er
v
i
s
ed
s
t
h
at
s
elec
t
n
ei
g
h
b
o
r
s
f
r
o
m
th
e
s
a
m
e
s
p
ea
k
er
:
=
(
)
=
∑
∑
,
=
∑
∈
(
9
)
3
.
6
.
B
a
y
esia
n dis
t
a
nce
m
et
ric
lea
rning
a
pp
ro
a
c
h
Her
e
I
u
s
ed
co
n
j
u
g
ate
g
r
ad
ie
n
t
m
et
h
o
d
to
o
b
tain
t
h
e
b
est
s
ee
d
.
T
h
e
NC
A
p
r
o
v
id
es
th
e
o
n
e
ti
m
e
esti
m
ate
o
f
t
h
e
d
is
tan
ce
m
e
tr
ic
an
d
ca
n
b
e
u
n
b
elie
v
ab
l
e
w
h
en
t
h
e
n
u
m
b
er
o
f
tr
ain
in
g
d
ata
is
le
s
s
.
I
u
s
e
th
e
B
a
y
e
s
ia
n
s
tr
u
ct
u
r
e
t
o
esti
m
ate
t
h
e
d
is
tr
ib
u
tio
n
af
t
er
th
e
d
is
tan
ce
m
etr
ic
[
8
]
.
L
o
o
k
in
g
at
th
e
s
p
ea
k
er
tag
f
o
r
ea
c
h
u
tter
an
ce
,
I
ca
n
cr
ea
te
t
w
o
s
et
s
o
f
b
ar
r
ier
s
f
o
r
th
e
s
a
m
e
s
p
ea
k
er
S
a
n
d
D.
T
h
e
p
o
s
s
ib
il
it
y
o
f
d
ef
in
i
n
g
t
w
o
u
tter
a
n
ce
s
a
n
d
r
elate
d
to
th
e
s
a
m
e
s
p
ea
k
er
o
r
d
if
f
er
en
t sp
ea
k
er
s
u
n
d
er
a
g
i
v
en
m
atr
ix
:
(
,
|
,
,
)
=
1
1
+
(
,
‖
−
‖
2
−
)
(
1
0
)
T
h
is
p
ar
am
eter
μ
is
th
e
th
r
esh
o
ld
f
o
r
s
ep
ar
atin
g
ex
p
r
ess
io
n
s
f
o
r
th
e
s
a
m
e
s
p
ea
k
er
p
ar
a
m
et
er
s
w
it
h
d
if
f
er
e
n
t
s
p
ea
k
er
s
.
O
n
l
y
w
h
e
n
v
al
u
e
o
f
μ
less
th
a
n
d
i
s
tan
ce
f
r
o
m
m
a
tr
ix
A
,
t
w
o
e
x
p
r
ess
io
n
s
ar
e
li
k
el
y
to
b
e
id
en
ti
f
ied
b
y
th
e
s
a
m
e
s
p
ea
k
er
.
T
h
e
y
i
,
j
is
d
ef
i
n
ed
as f
o
llo
w
s
[
2
4
,
2
5
]
:
,
=
{
+
1
(
,
)
−
1
(
,
)
(
1
1
)
I
u
s
e
t
h
e
NC
A
as
a
p
r
ep
r
o
ce
s
s
i
n
g
tech
n
iq
u
e
s
o
th
at
th
e
v
ec
to
r
s
ca
n
b
e
p
r
o
j
e
cted
in
s
u
ch
a
p
lace
w
h
er
e
th
e
n
ea
r
est
n
eig
h
b
o
r
s
o
f
ea
ch
s
i
m
p
li
f
ied
an
d
s
u
p
er
v
i
s
ed
m
o
d
i
f
ied
an
d
s
u
p
er
v
i
s
ed
ca
n
s
h
ar
e
th
e
s
a
m
e
tag
o
f
h
ig
h
p
r
o
b
ab
ilit
y
.
B
a
y
e
s
ian
d
is
ta
n
ce
lear
n
in
g
m
et
h
o
d
s
ca
n
e
m
u
late
th
e
d
is
ta
n
ce
b
et
w
ee
n
t
h
e
c
ar
r
ier
s
b
etter
an
d
m
o
r
e
d
ep
en
d
ab
le
in
a
n
e
w
p
lace
.
E
x
p
er
i
m
e
n
ta
l
r
esu
lt
s
s
h
o
w
t
h
e
ad
v
an
ta
g
e
s
o
f
t
h
e
B
a
y
e
s
ia
n
d
is
tan
ce
m
etr
ic
lear
n
i
n
g
ap
p
r
o
ac
h
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2
0
8
8
-
8708
B
a
ye
s
ia
n
d
is
ta
n
ce
metric le
a
r
n
in
g
a
n
d
its
a
p
p
lica
tio
n
in
a
u
t
o
m
a
tic
s
p
ea
ke
r
r
ec
o
g
n
itio
n
.
..
(
S
a
tya
n
a
n
d
S
in
g
h
)
2965
4.
E
XP
E
R
I
M
E
NT
A
L
RE
SUL
T
S
E
x
p
er
i
m
e
n
ts
ar
e
p
er
f
o
r
m
ed
o
n
t
h
e
1
3
d
i
f
f
er
en
t
s
p
ea
k
er
d
ete
ctio
n
te
s
ts
th
a
t
ar
e
d
ef
in
ed
b
y
th
e
d
u
r
atio
n
a
n
d
t
y
p
e
o
f
tr
ai
n
i
n
g
an
d
te
s
t
d
at
a
o
n
t
h
e
NI
ST
2
0
0
8
SR
E
d
ataset.
I
p
r
esen
t
r
esu
lt
s
o
n
t
h
e
s
h
o
r
t2
-
s
h
o
r
t3
an
d
1
0
s
ec
-
1
0
s
ec
co
n
d
itio
n
s
f
o
r
t
h
e
tr
ain
i
n
g
a
n
d
te
s
t
v
o
ice
d
ata
o
f
co
n
v
er
s
at
io
n
al
o
f
f
i
v
e
m
in
u
te
s
d
u
r
atio
n
.
I
n
t
h
is
r
esear
c
h
it
i
s
also
u
s
e
d
f
o
r
L
D
A
an
d
N
C
A
tr
ai
n
i
n
g
,
a
n
d
as
t
h
e
i
m
p
o
s
t
o
r
s
et
in
t
h
e
s
co
r
e
n
o
r
m
aliza
t
io
n
s
tep
.
A
6
0
0
-
d
im
en
s
io
n
m
o
d
i
f
ied
an
d
s
u
p
er
v
is
ed
is
e
x
tr
ac
ted
f
r
o
m
ea
c
h
u
t
ter
an
ce
.
T
h
e
E
q
u
al
E
r
r
o
r
R
ate
(
E
E
R
)
an
d
th
e
m
i
n
i
m
u
m
Dete
c
tio
n
C
o
s
t
F
u
n
cti
o
n
(
m
i
n
DC
F)
ar
e
u
s
ed
as
m
et
r
ics
f
o
r
ev
al
u
atio
n
.
C
o
s
i
n
e
s
i
m
ilar
it
y
s
co
r
in
g
an
d
B
ay
e
s
ia
n
d
is
ta
n
ce
m
etr
ic
lea
r
n
in
g
o
n
th
e
s
h
o
r
t2
-
s
h
o
r
t3
co
n
d
itio
n
o
f
t
h
e
NI
ST
2
0
0
8
SR
E
d
ataset.
T
h
e
B
ay
esi
an
d
is
ta
n
ce
m
etr
ic
lear
n
i
n
g
al
g
o
r
ith
m
i
s
r
ef
er
r
ed
to
as
\
B
a
y
e
s
d
m
l,
co
s
i
n
e
s
co
r
e
af
ter
t
h
e
co
m
b
i
n
ed
s
co
r
e
n
o
r
m
aliza
tio
n
f
o
r
C
o
s
i
n
e
Sco
r
e
c
o
m
b
i
n
ed
n
o
r
m
a
n
d
P
L
D
A
w
it
h
Ga
u
s
s
ia
n
GP
L
D
A
.
C
o
n
s
tr
ain
t
s
f
r
o
m
all
p
o
s
s
ib
le
m
o
d
i
f
ied
a
n
d
s
u
p
er
v
is
ed
p
air
s
f
r
o
m
t
h
e
s
a
m
e
s
p
ea
k
er
S,
ap
p
ly
t
h
e
co
s
in
e
s
co
r
e
to
all
p
o
s
s
ib
le
m
o
d
i
f
ied
an
d
s
u
p
er
v
is
ed
p
air
s
f
r
o
m
d
i
f
f
er
en
t
s
p
ea
k
er
s
an
d
s
elec
t
s
th
o
s
e
w
i
th
th
e
h
i
g
h
e
s
t
s
co
r
e
to
f
o
r
m
co
n
s
tr
ai
n
t
D
s
i
n
ce
t
h
ese
p
air
s
ar
e
th
e
lar
g
es
t
d
is
tin
c
tio
n
t
h
at
d
is
ti
n
g
u
i
s
h
e
s
t
h
e
m
etr
ic
d
is
ta
n
ce
.
Sin
ce
th
e
n
u
m
b
er
o
f
all
p
o
s
s
i
b
le
d
if
f
er
e
n
t
s
p
ea
k
er
p
air
s
i
s
v
er
y
lar
g
e,
I
c
h
o
s
e
t
w
i
ce
t
h
e
n
u
m
b
er
o
f
s
i
m
ilar
sp
ea
k
er
p
air
s
f
r
o
m
all
p
o
s
s
ib
le
s
p
ea
k
er
p
air
s
to
f
o
r
m
a
D.
T
h
e
ex
p
er
i
m
e
n
tal
ex
p
er
i
m
e
n
t
s
h
o
w
ed
a
lar
g
e
s
et
o
f
o
f
f
s
et
s
p
ea
k
er
co
n
s
tr
ai
n
ts
(
s
i
m
ilar
s
p
ea
k
er
p
air
s
)
d
o
es
n
o
t
i
m
p
r
o
v
e
p
er
f
o
r
m
a
n
ce
b
u
t
r
eq
u
ir
es
m
o
r
e
ca
lcu
latio
n
s
,
an
d
a
s
m
al
ler
s
et
o
f
s
p
ea
k
er
co
n
s
tr
ai
n
t
s
(
s
a
m
e
s
ize)
w
ith
d
i
f
f
er
e
n
t
s
p
ea
k
er
co
n
s
tr
ain
ts
w
ill
d
eg
r
ad
e
th
e
ASR
p
er
f
o
r
m
an
ce
.
T
h
e
co
m
p
ar
is
o
n
o
f
co
s
i
n
e
s
c
o
r
es,
B
ay
esia
n
_
d
m
l
o
f
NI
ST
2
0
0
8
SR
E
GP
L
D
A
n
o
r
m
aliza
t
io
n
is
s
h
o
w
n
in
T
ab
le
1
.
T
ab
le
1.
C
o
m
p
ar
is
o
n
o
f
co
s
in
e
s
co
r
es,
B
ay
es
ian
_
d
m
l o
f
NI
ST
2
0
0
8
S
R
E
GP
L
DA
n
o
r
m
al
izatio
n
C
o
mb
i
n
a
t
i
o
n
N
o
r
ms
EER
mi
n
D
C
F
L
D
A
2
0
0
C
o
si
n
e
S
c
o
r
e
2
.
5
4
1
%
0
.
0
1
4
4
5
C
o
si
n
e
S
c
o
r
e
c
o
m
b
i
n
e
d
n
o
r
m
1
.
7
9
0
%
0
.
0
0
9
7
B
a
y
e
s d
ml
2
.
1
5
8
%
0
.
0
1
0
6
B
a
y
e
s d
ml
+
z
n
o
r
m
2
.
1
5
9
%
0
.
0
1
0
7
B
a
y
e
s d
ml
+
t
n
o
r
m
2
.
1
5
8
%
0
.
0
1
0
7
G
P
L
D
A
3
.
1
2
%
0
.
0
1
5
6
As
o
n
e
ca
n
s
ee
in
T
ab
le
1
,
th
e
co
s
in
e
s
co
r
e
co
m
b
in
ed
w
it
h
th
e
L
D
A
2
0
0
s
tan
d
ar
d
ca
n
ac
h
iev
e
t
h
e
b
est
r
esu
lts
,
a
n
d
GP
L
D
A
p
er
f
o
r
m
s
th
e
w
o
r
s
t.
Ho
w
e
v
er
,
B
ay
esian
d
m
l
p
er
f
o
r
m
s
b
etter
th
an
th
e
co
s
in
e
s
co
r
e
i
f
th
e
s
co
r
e
is
n
o
t
n
o
r
m
alize
d
.
C
o
m
p
ar
ed
to
th
e
s
tate
-
of
-
t
h
e
-
ar
t
A
S
R
p
er
f
o
r
m
a
n
ce
o
f
th
e
i
n
te
g
r
ated
co
s
in
e
s
co
r
e
s
tan
d
ar
d
,
th
e
g
ap
w
it
h
B
a
y
e
s
ian
_
d
m
l
i
s
v
er
y
s
m
all.
I
n
ad
d
itio
n
,
th
er
e
is
a
l
m
o
s
t
n
o
ad
v
an
tag
e
i
n
t
h
e
n
o
r
m
aliza
t
io
n
o
f
s
co
r
es i
n
B
a
y
esian
_
d
m
l.
B
y
u
n
d
er
s
ta
n
d
in
g
t
h
e
d
if
f
er
en
ce
s
b
et
w
ee
n
t
h
e
C
o
s
i
n
e
Sco
r
e
co
m
b
in
ed
cr
ite
r
ia
an
d
B
a
y
esian
_
d
m
l,
I
co
m
p
ar
e
th
eir
p
er
f
o
r
m
a
n
c
e
to
th
e
d
if
f
er
e
n
t
co
m
b
i
n
at
io
n
s
o
f
tec
h
n
ica
l
p
r
ep
r
o
ce
s
s
in
g
.
P
r
etr
ea
tm
e
n
t
tech
n
iq
u
es
i
n
clu
d
e
L
D
A
a
n
d
NC
A
,
w
h
ic
h
ar
e
ap
p
l
ied
b
ef
o
r
e
th
e
s
co
r
in
g
m
o
d
el.
T
h
e
co
m
p
ar
i
s
o
n
o
f
co
s
i
n
e
s
co
r
es,
B
ay
e
s
ian
_
d
m
l o
f
NI
ST
2
0
0
8
SR
E
d
if
f
er
e
n
t n
o
r
m
aliza
tio
n
is
s
h
o
w
n
in
T
ab
le
2
.
T
ab
le
2
.
C
o
m
p
ar
is
o
n
o
f
co
s
in
e
s
co
r
es,
B
ay
es
ian
_
d
m
l o
f
NI
ST
2
0
0
8
SR
E
d
if
f
er
en
t
n
o
r
m
al
izatio
n
C
o
mb
i
n
a
t
i
o
n
N
o
r
ms
EER
mi
n
D
C
F
L
D
A
2
0
0
1
.
7
9
0
%
0
.
0
0
9
7
N
C
A
1
5
0
+
L
D
A
1
5
0
4
3
.
3
4
5
%
0
.
0
9
8
7
N
C
A
2
0
0
2
.
3
4
5
%
0
.
0
1
3
1
N
C
A
2
0
0
+
L
D
A
1
0
0
2
.
0
1
7
%
0
.
0
0
9
7
N
C
A
2
0
0
+
L
D
A
2
0
0
1
.
7
6
7
%
0
.
0
0
9
5
N
C
A
2
0
0
+
L
D
A
6
0
0
4
1
.
0
7
8
%
0
.
0
1
9
7
N
C
A
2
0
0
4
.
5
6
7
%
0
.
0
2
7
8
B
a
y
e
s_
d
ml
L
D
A
2
0
0
2
.
1
7
6
%
0
.
0
1
0
7
L
D
A
2
0
0
+
N
C
A
1
5
0
+
L
D
A
1
5
0
4
2
.
3
4
5
%
0
.
1
0
0
5
L
D
A
2
0
0
+
N
C
A
2
0
0
3
.
0
3
4
%
0
.
1
7
9
L
D
A
2
0
0
+
N
C
A
2
0
0
+
L
D
A
1
0
0
1
.
7
7
5
%
0
.
0
0
9
7
L
D
A
2
0
0
+
N
C
A
2
0
0
+
L
D
A
2
0
0
1
.
8
1
7
%
0
.
1
1
0
0
L
D
A
6
0
0
+
N
C
A
2
0
0
+
L
D
A
2
0
0
4
3
.
7
8
6
%
0
.
1
0
0
1
N
C
A
2
0
0
+
L
D
A
2
0
0
3
.
5
6
4
%
0
.
0
1
7
8
T
ab
le
2
g
iv
es
u
s
an
id
ea
o
f
h
o
w
N
C
A
a
n
d
L
D
A
r
ep
r
esen
t
h
id
d
en
s
tr
u
c
tu
r
es
i
n
t
h
e
to
tal
v
ar
iatio
n
s
p
ac
e.
T
h
e
w
o
r
s
t
p
er
f
o
r
m
a
n
ce
ap
p
ea
r
ed
in
th
e
s
ec
o
n
d
an
d
s
ix
t
h
li
n
es.
I
n
b
o
th
ca
s
es,
t
h
e
s
ize
o
f
th
e
N
C
A
i
s
d
if
f
er
e
n
t
f
r
o
m
t
h
e
s
ize
o
f
th
e
p
r
ev
io
u
s
L
D
A
.
T
h
at
is
to
s
ay
,
th
e
NC
A
ac
ts
to
r
ed
u
ce
th
e
s
i
ze
an
d
h
as
a
s
er
io
u
s
i
m
p
ac
t
o
n
t
h
e
r
es
u
lt
s
.
I
n
t
h
e
f
o
u
r
th
r
o
w
,
t
h
e
NC
A
2
0
0
o
n
l
y
p
er
f
o
r
m
s
o
n
e
r
o
tatio
n
f
o
llo
win
g
t
h
e
L
D
A
2
0
0
,
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8708
I
n
t J
E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
9
,
No
.
4
,
A
u
g
u
s
t 2
0
1
9
:
2
9
6
0
-
2967
2966
an
d
th
e
L
D
A
1
0
0
later
f
u
r
th
e
r
r
ed
u
ce
s
t
h
e
s
ize
o
f
t
h
e
f
ea
tu
r
e
s
p
ac
e,
w
h
ic
h
h
as
litt
le
e
f
f
e
ct
o
n
p
er
f
o
r
m
an
ce
.
W
ith
th
e
ex
ce
p
tio
n
o
f
t
h
e
s
ec
o
n
d
an
d
f
o
u
r
th
r
o
w
s
,
th
e
r
esu
l
ts
in
th
e
s
e
v
e
n
th
r
o
w
ar
e
al
m
o
s
t
at
th
e
s
a
m
e
le
v
el
as
th
e
o
th
er
r
o
w
s
,
alt
h
o
u
g
h
th
e
s
ize
o
f
th
e
N
C
A
2
0
0
is
r
ed
u
ce
d
o
n
th
e
6
0
0
-
d
i
m
en
s
i
o
n
al
m
o
d
if
ied
an
d
s
u
p
er
v
i
s
ed
.
T
h
e
r
ea
s
o
n
m
a
y
b
e
th
at
th
i
s
s
ize
r
ed
u
ctio
n
is
d
o
n
e
in
t
h
e
o
r
ig
i
n
al
to
t
al
v
ar
iab
ilit
y
s
p
ac
e,
w
h
ile
t
h
e
s
ec
o
n
d
an
d
f
o
u
r
t
h
lin
e
q
u
o
ta
r
ed
u
ctio
n
s
ar
e
m
a
d
e
in
th
e
r
ed
u
ce
d
f
u
n
ct
io
n
s
p
ac
e
af
ter
th
e
L
D
A
.
I
m
p
r
o
v
e
m
e
n
ts
i
n
t
h
e
f
o
u
r
th
r
o
w
in
d
icate
t
h
at
t
h
e
L
D
A
p
r
o
j
ec
tio
n
co
r
r
ec
ts
th
e
f
u
n
c
tio
n
al
s
p
ac
e
i
n
d
icatio
n
lear
n
ed
f
r
o
m
t
h
e
NC
A
.
T
h
er
ef
o
r
e,
I
ca
n
co
n
clu
d
e
th
at
NC
A
d
id
n
o
t p
lay
a
r
o
le
in
r
ed
u
ci
n
g
th
e
d
i
m
e
n
s
io
n
.
T
h
e
b
est
p
er
f
o
r
m
a
n
ce
o
f
t
h
e
co
m
b
i
n
ed
co
s
i
n
e
s
co
r
e
is
o
b
tain
ed
u
s
in
g
L
D
A
2
0
0
+
NC
A
2
0
0
+
L
D
A
2
0
0
,
an
d
th
e
b
est
p
er
f
o
r
m
an
ce
o
f
B
a
y
es_
d
m
l
w
as
o
b
tain
ed
u
s
in
g
L
D
A
2
0
0
+
NC
A
2
0
0
+
L
D
A
1
0
0
.
B
ay
e
s
ian
_
d
m
l
o
v
er
co
m
es
th
e
co
m
b
in
ed
n
o
r
m
o
f
co
s
i
n
e
s
co
r
es
an
d
is
th
e
b
est
r
es
u
lt
o
f
t
h
e
s
h
o
r
t2
-
s
h
o
r
t3
co
n
d
itio
n
r
ep
o
r
t f
o
r
NI
ST
SR
E
2
0
0
8
d
ata.
W
ith
a
s
p
ec
i
f
ic
e
n
d
g
o
al
to
c
h
ec
k
t
h
e
p
er
f
o
r
m
a
n
ce
o
f
L
D
A
2
0
0
+
N
C
A
2
0
0
+
L
D
A
2
0
0
b
ased
ASR
,
p
r
o
ce
s
s
ed
th
e
r
ea
l
m
atc
h
s
c
o
r
e
s
w
it
h
L
D
A
2
0
0
+
NC
A
2
0
0
+
L
DA
1
0
0
w
it
h
th
e
i
m
p
o
s
to
r
m
atc
h
s
co
r
es.
T
h
e
D
etec
tio
n
E
r
r
o
r
T
r
ad
e
-
o
f
f
(
DE
T
)
o
f
L
D
A
2
0
0
+
N
C
A
2
0
0
+
L
D
A
2
0
0
p
er
f
o
r
m
a
n
ce
w
it
h
NI
ST
2
0
0
8
SR
E
d
ataset
an
E
E
R
v
a
lu
e
o
f
ab
o
u
t
1
.
7
6
7
%
an
d
w
it
h
L
D
A
2
0
0
+
NC
A
2
0
0
+
L
D
A
1
0
0
th
e
E
E
R
o
f
ab
o
u
t
1
.
7
7
6
%
is
s
h
o
w
n
in
F
ig
u
r
e
1.
Fig
u
r
e
1
.
E
q
u
al
er
r
o
r
r
ate
o
f
L
DA
2
0
0
+
NC
A
2
0
0
+
L
D
A
2
0
0
an
d
B
ay
e
s
_
d
m
l
u
s
i
n
g
L
D
A
2
0
0
+
NC
A
2
0
0
+
L
D
A
1
0
0
b
ased
A
S
R
s
y
s
te
m
5.
CO
NCLU
SI
O
N
T
h
is
p
ap
er
p
r
esen
ts
th
e
A
S
R
ap
p
licatio
n
w
it
h
Ba
y
e
s
ia
n
Dis
tan
ce
L
ea
r
n
i
n
g
Me
tr
ic
tech
n
iq
u
e.
T
h
e
tr
ad
itio
n
al
i
-
v
ec
to
r
h
a
s
b
ee
n
m
o
d
i
f
ied
b
y
ca
te
n
ati
n
g
t
h
e
la
b
el
v
ec
to
r
,
li
n
ea
r
r
eg
r
e
s
s
io
n
m
atr
i
x
,
m
ea
n
s
u
p
er
v
ec
to
r
an
d
m
o
d
if
ied
an
d
s
u
p
er
v
is
ed
f
ac
to
r
lo
ad
in
g
m
atr
i
x
.
T
h
e
m
o
d
if
ied
an
d
s
u
p
er
v
i
s
ed
i
-
v
eto
r
in
p
r
o
p
o
s
ed
ASR
h
as
a
v
er
y
h
i
g
h
d
eg
r
ee
o
f
d
is
cr
i
m
i
n
ati
v
el
y
in
o
r
d
er
to
r
eg
u
lar
ize
t
h
e
s
p
ea
k
er
s
p
ec
i
f
i
c
lab
el
in
f
o
r
m
a
tio
n
to
en
h
a
n
ce
th
e
p
er
f
o
r
m
an
ce
o
f
A
S
R
.
T
h
e
p
r
o
p
o
s
ed
Ba
y
esia
n
Dis
ta
n
ce
L
ea
r
n
in
g
Me
tr
ic
tech
n
iq
u
e
in
ASR
ca
n
b
e
u
til
ized
as
an
in
teg
r
ate
d
to
en
h
a
n
ce
t
h
e
e
f
f
icien
c
y
a
n
d
o
u
tp
er
f
o
r
m
t
h
e
tr
ad
itio
n
al
i
-
v
e
ct
o
r
.
T
h
e
p
r
o
p
o
s
ed
A
SR
s
y
s
te
m
h
as
ac
h
ie
v
ed
th
e
b
est
p
er
f
o
r
m
a
n
ce
o
f
its
E
E
R
as
1
.
7
6
7
%
u
s
in
g
L
D
A
2
0
0
+
NC
A
2
0
0
+
L
DA
2
0
0
w
it
h
co
m
b
in
ed
co
s
in
e
s
co
r
e
n
o
r
m
aliza
t
io
n
a
n
d
b
est
p
er
f
o
r
m
an
ce
o
f
B
a
y
es_
d
m
l
w
as
o
b
t
ain
ed
u
s
i
n
g
L
D
A
2
0
0
+
NC
A
2
0
0
+
L
D
A
1
0
0
o
f
its
E
E
R
at
1
.
7
7
6
%.
RE
F
E
R
E
NC
E
S
[1
]
S
.
S
i
n
g
h
,
“
F
o
re
n
sic
a
n
d
A
u
to
m
a
ti
c
S
p
e
a
k
e
r
Re
c
o
g
n
it
io
n
S
y
ste
m
,
”
In
ter
n
a
ti
o
n
a
l
J
o
u
rn
a
l
o
f
El
e
c
trica
l
a
n
d
Co
mp
u
ter
E
n
g
in
e
e
rin
g
,
v
o
l
/i
ss
u
e
:
8
(
5
)
,
p
p
.
2
8
0
4
-
2
8
1
1
,
2
0
1
8
.
[2
]
S.
S
in
g
h
.
“
Hig
h
L
e
v
e
l
S
p
e
a
k
e
r
S
p
e
c
if
ic
F
e
a
tu
re
s
a
s
a
n
E
f
f
i
c
ien
c
y
En
h
a
n
c
in
g
P
a
ra
m
e
ter
s
in
S
p
e
a
k
e
r
Re
c
o
g
n
it
io
n
S
y
st
e
m
,
”
In
ter
n
a
ti
o
n
a
l
J
o
u
r
n
a
l
o
f
El
e
c
trica
l
a
n
d
C
o
mp
u
ter
En
g
i
n
e
e
rin
g
,
v
ol
/
issu
e
:
9
(
4
)
,
2
0
1
9
.
[3
]
S.
S
i
n
g
h
,
“
T
h
e
Ro
le
o
f
S
p
e
e
c
h
T
e
c
h
n
o
lo
g
y
in
Bio
m
e
tri
c
s,
F
o
re
n
sic
s
a
n
d
M
a
n
-
M
a
c
h
in
e
I
n
terf
a
c
e
,
”
In
ter
n
a
t
io
n
a
l
J
o
u
rn
a
l
o
f
El
e
c
trica
l
a
n
d
C
o
mp
u
t
e
r E
n
g
i
n
e
e
rin
g
,
v
o
l
/i
ss
u
e
:
9
(
1
)
,
p
p
.
2
8
1
-
2
8
8
,
2
0
1
9
.
0
.
1
0
.
2
0
.
5
1
2
5
1
0
2
0
4
0
0
.
1
0
.
2
0
.
5
1
2
5
1
0
2
0
4
0
Fa
l
s
e
A
l
a
r
m
proba
bi
l
i
t
y
(
i
n
%
)
M
i
s
s
proba
bi
l
i
t
y
(
i
n
%
)
LD
A
2
0
0
+
N
C
A
2
0
0
+
LD
A
1
0
0
LD
A
2
0
0
+
N
C
A
2
0
0
+
LD
A
2
0
0
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2
0
8
8
-
8708
B
a
ye
s
ia
n
d
is
ta
n
ce
metric le
a
r
n
in
g
a
n
d
its
a
p
p
lica
tio
n
in
a
u
t
o
m
a
tic
s
p
ea
ke
r
r
ec
o
g
n
itio
n
.
..
(
S
a
tya
n
a
n
d
S
in
g
h
)
2967
[4
]
S.
S
in
g
h
,
e
t
a
l
.
,
“
S
h
o
rt
Du
ra
ti
o
n
V
o
ice
Da
ta
S
p
e
a
k
e
r
R
e
c
o
g
n
it
io
n
S
y
st
e
m
U
sin
g
No
v
e
l
F
u
z
z
y
V
e
c
t
o
r
Qu
a
n
ti
z
a
ti
o
n
A
l
g
o
rit
h
m
s
,
”
2
0
1
6
IEE
E
In
ter
n
a
t
io
n
a
l
In
stru
me
n
ta
ti
o
n
a
n
d
M
e
a
su
re
me
n
t
T
e
c
h
n
o
lo
g
y
Co
n
fer
e
n
c
e
,
T
a
ip
e
i,
Tai
w
a
n
pp.
1
-
6
,
2
0
1
6
.
[5
]
P
.
Be
li
n
,
e
t
a
l
.
,
“
Vo
ice
-
se
lec
ti
v
e
a
re
a
s in
h
u
m
a
n
a
u
d
i
to
ry
c
o
rtex
,
”
Na
tu
re
,
v
o
l
.
4
0
3
,
p
p
.
3
0
9
-
3
1
2
,
2
0
0
0
.
[6
]
S.
S
in
g
h
,
“
Ev
a
lu
a
ti
o
n
o
f
S
p
a
rsif
ica
ti
o
n
a
lg
o
rit
h
m
a
n
d
Its
Ap
p
li
c
a
ti
o
n
i
n
S
p
e
a
k
e
r
Re
c
o
g
n
it
io
n
S
y
ste
m
,
”
In
ter
n
a
t
io
n
a
l
J
o
u
rn
a
l
o
f
A
p
p
li
e
d
En
g
i
n
e
e
rin
g
Res
e
a
rc
h
,
v
ol
/i
ss
u
e
:
13
(
17
)
,
p
p
.
1
3
0
1
5
-
1
3
0
2
1
,
2
0
1
8
.
[7
]
E.
F
o
rm
isa
n
o
,
e
t
a
l
.
,
“
W
h
o
’
is
sa
y
in
g
‘w
h
a
t’?
Bra
in
b
a
se
d
d
e
c
o
d
i
n
g
o
f
h
u
m
a
n
v
o
ice
a
n
d
sp
e
e
c
h
,
”
S
c
i
e
n
c
e
,
v
o
l.
3
2
2
,
p
p
.
9
7
0
-
9
7
3
,
2
0
0
8
.
[8
]
D.
A
.
Re
y
n
o
ld
s,
e
t
a
l
.
,
“
S
p
e
a
k
e
r
v
e
rif
ic
a
ti
o
n
u
si
n
g
v
a
d
a
p
ted
G
a
u
s
sia
n
m
ix
tu
re
m
o
d
e
ls,”
Dig
it
a
l
S
ig
n
a
l
Pro
c
e
ss
,
v
o
l
/i
ss
u
e
:
10
(
1
-
3
)
,
p
p
.
19
-
4
1
,
2
0
0
0
.
[9
]
M.
Zi
ss
m
a
n
,
“
Lan
g
u
a
g
e
id
e
n
t
if
ica
ti
o
n
u
sin
g
p
h
o
n
e
m
e
re
c
o
g
n
it
io
n
a
n
d
p
h
o
n
o
tac
ti
c
lan
g
u
a
g
e
m
o
d
e
li
n
g
,
”
Pro
c
.
ICAS
S
P
,
p
p
.
3
5
0
3
-
3
5
0
6
,
1
9
9
5
.
[1
0
]
S.
S
in
g
h
,
“
S
u
p
p
o
rt
V
e
c
to
r
M
a
c
h
in
e
Ba
se
d
A
p
p
ro
a
c
h
e
s
F
o
r
Re
a
l
T
i
m
e
A
u
to
m
a
ti
c
S
p
e
a
k
e
r
Re
c
o
g
n
it
io
n
S
y
ste
m
,
”
In
ter
n
a
t
io
n
a
l
J
o
u
rn
a
l
o
f
A
p
p
li
e
d
En
g
i
n
e
e
rin
g
Res
e
a
rc
h
,
v
o
l/
issu
e
:
13
(
10
)
,
p
p
.
8
5
6
1
-
8
5
6
7
,
2
0
1
8
.
[1
1
]
S.
S
i
n
g
h
,
e
t
a
l
.
,
“
S
p
e
a
k
e
r
S
p
e
c
if
ic
P
h
o
n
e
S
e
q
u
e
n
c
e
a
n
d
S
u
p
p
o
r
t
V
e
c
to
r
M
a
c
h
i
n
e
s
T
e
lep
h
o
n
ic
Ba
se
d
S
p
e
a
k
e
r
Re
c
o
g
n
it
io
n
S
y
ste
m
,
”
In
ter
n
a
ti
o
n
a
l
J
o
u
rn
a
l
o
f
Ap
p
li
e
d
E
n
g
i
n
e
e
rin
g
Res
e
a
rc
h
,
v
ol
/i
ss
u
e
:
12
(
19
)
,
p
p
.
8
0
2
6
-
8
0
3
3
,
2
0
1
7
.
[1
2
]
H.
L
i,
e
t
a
l
.
,
“
S
p
o
k
e
n
lan
g
u
a
g
e
re
c
o
g
n
it
io
n
:
F
ro
m
f
u
n
d
a
m
e
n
tals
to
p
ra
c
ti
c
e
,
”
Pro
c
e
e
d
in
g
s
o
f
t
h
e
IEE
E
1
0
1
,
p
p
.
1
1
3
6
-
1
1
5
9
,
2
0
1
3
.
[1
3
]
P.
T
.
Ca
rra
sq
u
il
lo
,
e
t
a
l
.
,
“
A
p
p
ro
a
c
h
e
s
to
lan
g
u
a
g
e
id
e
n
ti
f
ica
ti
o
n
u
sin
g
g
a
u
ss
ian
m
ix
tu
re
m
o
d
e
ls
a
n
d
sh
if
te
d
d
e
lt
a
c
e
p
stra
l
f
e
a
tu
re
s,”
Pro
c
.
ICS
L
P
,
p
p
.
8
9
-
9
2
,
2
0
0
2
.
[1
4
]
S
.
M
.
S
in
isc
a
lch
i,
e
t
a
l
.
,
“
Ex
p
lo
it
in
g
c
o
n
tex
t
-
d
e
p
e
n
d
e
n
c
y
a
n
d
a
c
o
u
stic
re
so
lu
t
io
n
o
f
u
n
iv
e
rsa
l
sp
e
e
c
h
a
tt
rib
u
te
m
o
d
e
ls i
n
sp
o
k
e
n
la
n
g
u
a
g
e
re
c
o
g
n
it
io
n
,
”
Pro
c
.
INT
ER
S
P
EE
CH
,
p
p
.
2
7
1
8
-
2
7
2
1
,
2
0
1
0
.
[1
5
]
G
.
Hin
to
n
,
e
t
a
l
.
,
“
De
e
p
n
e
u
ra
l
n
e
tw
o
rk
s
f
o
r
a
c
o
u
stic
m
o
d
e
li
n
g
in
sp
e
e
c
h
re
c
o
g
n
it
io
n
,
”
T
h
e
sh
a
re
d
v
iews
o
f
fo
u
r
re
se
a
rc
h
g
ro
u
p
s.
IEE
E
S
i
g
n
a
l
Pr
o
c
e
ss
in
g
M
a
g
a
z
in
e
,
v
o
l.
2
9
,
p
p
.
8
2
-
9
7
,
2
0
1
2
.
[1
6
]
L
.
De
n
g
a
n
d
X.
L
i,
“
M
a
c
h
in
e
le
a
rn
in
g
p
a
ra
d
ig
m
s
f
o
r
sp
e
e
c
h
re
c
o
g
n
it
io
n
:
A
n
o
v
e
rv
iew
,
”
IEE
E
T
ra
n
sa
c
ti
o
n
s
o
n
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
1
,
p
p
.
1
0
6
0
-
1
0
8
9
,
2
0
1
3
.
[1
7
]
S.
S
i
n
g
h
,
“
S
p
e
a
k
e
r
Re
c
o
g
n
it
io
n
b
y
G
a
u
ss
ian
F
il
ter
Ba
se
d
F
e
a
tu
re
Ex
trac
ti
o
n
a
n
d
P
r
o
p
o
se
d
F
u
z
z
y
V
e
c
to
r
Qu
a
n
ti
z
a
ti
o
n
M
o
d
e
li
n
g
T
e
c
h
n
iq
u
e
,
”
In
ter
n
a
t
io
n
a
l
J
o
u
rn
a
l
o
f
A
p
p
li
e
d
En
g
in
e
e
rin
g
Res
e
a
rc
h
,
v
o
l
/i
ss
u
e
:
13
(
16
)
,
p
p
.
1
2
7
9
8
-
1
2
8
0
4
,
2
0
1
8
.
[1
8
]
P.
Ke
n
n
y
,
e
t
al
.
,
“
A
stu
d
y
o
f
in
ters
p
e
a
k
e
r
v
a
ri
a
b
il
it
y
in
sp
e
a
k
e
r
v
e
r
if
ic
a
ti
o
n
,
”
IEE
E
T
ra
n
sa
c
ti
o
n
s
o
n
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
.
1
6
,
p
p
.
9
8
0
-
9
8
8
,
2
0
0
8
.
[1
9
]
W
.
Ca
m
p
b
e
ll
,
e
t
a
l
.
,
“
S
u
p
p
o
rt
v
e
c
to
r
m
a
c
h
in
e
s
u
sin
g
g
m
m
su
p
e
rv
e
c
to
rs
f
o
r
sp
e
a
k
e
r
v
e
rif
i
c
a
ti
o
n
,
”
IEE
E
S
i
g
n
a
l
Pro
c
e
ss
in
g
L
e
tt
e
rs
,
v
o
l.
1
3
,
p
p
.
3
0
8
-
3
1
1
,
2
0
0
6
.
[2
0
]
L
.
Bu
rg
e
t,
e
t
a
l
.
,
“
Disc
rim
in
a
ti
v
e
train
in
g
tec
h
n
iq
u
e
s f
o
r
a
c
o
u
stic la
n
g
u
a
g
e
id
e
n
ti
f
ica
ti
o
n
,
”
Pro
c
.
ICA
S
S
P
,
2
0
0
6
.
[2
1
]
S.
S
in
g
h
,
e
t
a
l
.
,
“
S
h
o
rt
Du
ra
ti
o
n
V
o
ice
Da
ta
S
p
e
a
k
e
r
R
e
c
o
g
n
it
io
n
S
y
st
e
m
U
sin
g
No
v
e
l
F
u
z
z
y
V
e
c
t
o
r
Qu
a
n
ti
z
a
ti
o
n
A
l
g
o
rit
h
ms
,
”
IEE
E
In
ter
n
a
ti
o
n
a
l
In
s
tru
me
n
ta
ti
o
n
a
n
d
M
e
a
su
re
me
n
t
T
e
c
h
n
o
lo
g
y
Co
n
fer
e
n
c
e
,
2
0
1
6
.
[2
2
]
S
.
S
in
g
h
,
e
t
a
l
.
,
“
Ef
f
icie
n
t
M
o
d
e
ll
in
g
T
e
c
h
n
iq
u
e
b
a
se
d
S
p
e
a
k
e
r
Re
c
o
g
n
i
ti
o
n
u
n
d
e
r
L
im
it
e
d
S
p
e
e
c
h
Da
ta
,
”
In
ter
n
a
t
io
n
a
l
J
o
u
rn
a
l
o
f
Im
a
g
e
,
G
ra
p
h
ics
a
n
d
S
ig
n
a
l
Pro
c
e
ss
in
g
(
IJ
IGS
P)
,
v
ol
/
issu
e
:
8
(
11
)
,
p
p
.
41
-
4
8
,
2
0
1
6
.
[2
3
]
D.
M
a
rti
n
e
z
,
e
t
a
l
.
,
“
L
a
n
g
u
a
g
e
re
c
o
g
n
it
io
n
i
n
iv
e
c
to
rs sp
a
c
e
,
”
Pro
c
.
INT
ER
S
P
EE
CH
,
p
p
.
8
6
1
-
8
6
4
,
2
0
1
1
.
[2
4
]
S.
S
i
n
g
h
,
e
t
a
l
.
,
“
A
No
v
e
l
A
lg
o
rit
h
m
o
f
S
p
a
rse
Re
p
re
se
n
tatio
n
s
f
o
r
S
p
e
e
c
h
Co
m
p
re
ss
io
n
/E
n
h
a
n
c
e
m
e
n
t
a
n
d
Its
A
p
p
li
c
a
ti
o
n
in
S
p
e
a
k
e
r
R
e
c
o
g
n
it
io
n
S
y
ste
m
,
”
In
ter
n
a
ti
o
n
a
l
J
o
u
rn
a
l
o
f
Co
mp
u
ta
ti
o
n
a
l
a
n
d
Ap
p
li
e
d
M
a
th
e
ma
ti
c
s
,
v
o
l
/i
ss
u
e
:
11
(
1
)
,
p
p
.
8
9
-
1
0
4
,
2
0
1
6
.
[2
5
]
S.
S
i
n
g
h
a
n
d
A
.
S
in
g
h
,
“
A
c
c
u
ra
c
y
Co
m
p
a
riso
n
u
sin
g
Dif
fe
re
n
t
M
o
d
e
li
n
g
T
e
c
h
n
iq
u
e
s
u
n
d
e
r
L
i
m
it
e
d
S
p
e
e
c
h
Da
ta o
f
S
p
e
a
k
e
r
Re
c
o
g
n
it
io
n
S
y
ste
m
s
,
”
Glo
b
a
l
J
o
u
r
n
a
l
o
f
S
c
ien
c
e
Fr
o
n
ti
e
r
Res
e
a
rc
h
:
F
M
a
th
e
ma
ti
c
s
a
n
d
De
c
isio
n
S
c
ien
c
e
s
,
v
o
l
/i
ss
u
e
:
16
(
2
)
,
p
p
.
1
-
1
7
,
2
0
1
6
.
Evaluation Warning : The document was created with Spire.PDF for Python.