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
4
8
8
~
2
4
9
7
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
6
5
5
2488
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
K
erna
l bas
ed
spe
a
ker specifi
c f
ea
t
ure e
x
trac
tion a
n
d it
s
a
pplica
tions
in iT
a
ukei cros
s la
ng
ua
g
e spea
ker re
co
g
nition
Sa
t
y
a
na
nd
Sin
g
h
1
,
P
ra
g
y
a
Si
ng
h
2
1
S
c
h
o
o
l
o
f
El
e
c
tri
c
a
l
a
n
d
El
e
c
tro
n
ics
En
g
i
n
e
e
rin
g
,
F
ij
i
Na
ti
o
n
a
l
Un
i
v
e
rsisty
,
F
ij
i
2
S
c
h
o
o
l
o
f
P
u
b
li
c
He
a
lt
h
a
n
d
P
ri
m
a
ry
Ca
re
,
F
ij
i
Na
ti
o
n
a
l
Un
iv
e
rsi
sty
,
F
i
ji
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
No
v
2
1
,
2
0
1
9
R
ev
is
ed
Ma
y
2
,
2
0
2
0
Acc
ep
ted
May
11
,
2
0
2
0
Ex
trac
ti
o
n
a
n
d
c
las
sifica
ti
o
n
a
lg
o
r
it
h
m
s
b
a
se
d
o
n
k
e
rn
e
l
n
o
n
l
in
e
a
r
f
e
a
tu
re
s
a
re
p
o
p
u
lar
in
t
h
e
n
e
w
d
irec
ti
o
n
o
f
r
e
se
a
rc
h
in
m
a
c
h
i
n
e
lea
rn
i
n
g
.
T
h
i
s
re
se
a
rc
h
p
a
p
e
r
c
o
n
si
d
e
rs
th
e
ir
p
ra
c
ti
c
a
l
a
p
p
li
c
a
ti
o
n
in
t
h
e
iT
a
u
k
e
i
a
u
to
m
a
ti
c
sp
e
a
k
e
r
re
c
o
g
n
it
i
o
n
s
y
ste
m
(ASR
)
fo
r
c
ro
ss
-
lan
g
u
a
g
e
sp
e
e
c
h
re
c
o
g
n
it
io
n
.
S
e
c
o
n
d
,
n
o
n
li
n
e
a
r
s
p
e
a
k
e
r
-
sp
e
c
ifi
c
e
x
tra
c
ti
o
n
m
e
th
o
d
s
su
c
h
a
s
k
e
r
n
e
l
p
ri
n
c
ip
a
l
c
o
m
p
o
n
e
n
t
a
n
a
ly
sis
(K
P
CA),
k
e
rn
e
l
in
d
e
p
e
n
d
e
n
t
c
o
m
p
o
n
e
n
t
a
n
a
l
y
si
s
(KICA
),
a
n
d
k
e
rn
e
l
li
n
e
a
r
d
isc
rimin
a
n
t
a
n
a
l
y
sis
(KLDA
)
a
re
su
m
m
a
rize
d
.
Th
e
c
o
n
v
e
rsi
o
n
e
ffe
c
ts
o
n
s
u
b
se
q
u
e
n
t
c
las
sifica
ti
o
n
s
we
re
tes
ted
in
c
o
n
ju
n
c
ti
o
n
with
G
a
u
ss
ian
m
ix
tu
re
m
o
d
e
li
n
g
(G
M
M
)
lea
rn
in
g
a
lg
o
rit
h
m
s;
in
m
o
st
c
a
se
s,
c
o
m
p
u
tatio
n
s
we
r
e
fo
u
n
d
to
h
a
v
e
a
b
e
n
e
ficia
l
e
ffe
c
t
o
n
c
las
sifica
ti
o
n
p
e
rfo
rm
a
n
c
e
.
A
d
d
i
ti
o
n
a
ll
y
,
th
e
b
e
st
re
su
lt
s
we
re
a
c
h
iev
e
d
b
y
th
e
Ke
rn
e
l
li
n
e
a
r
d
isc
rimin
a
n
t
a
n
a
ly
sis
(KLDA
)
a
lg
o
rit
h
m
.
T
h
e
p
e
rfo
rm
a
n
c
e
o
f
th
e
A
S
R
sy
ste
m
is
e
v
a
lu
a
ted
fo
r
c
lea
r
sp
e
e
c
h
to
a
wid
e
ra
n
g
e
o
f
sp
e
e
c
h
q
u
a
li
t
y
u
sin
g
ATR
Ja
p
a
n
e
se
C
l
a
n
g
u
a
g
e
c
o
r
p
u
s
a
n
d
se
lf
-
re
c
o
rd
e
d
i
Tau
k
e
i
c
o
rp
u
s.
T
h
e
ASR
e
fficie
n
c
y
o
f
KLDA
,
KICA
,
a
n
d
KLDA
tec
h
n
iq
u
e
fo
r
6
se
c
o
f
ATR
Ja
p
a
n
e
se
C
lan
g
u
a
g
e
c
o
rp
u
s
9
9
.
7
%
,
9
9
.
6
%
,
a
n
d
9
9
.
1
%
a
n
d
e
q
u
a
l
e
rro
r
ra
te
(EE
R)
a
re
1
.
9
5
%
,
2
.
3
1
%
,
a
n
d
3
.
4
1
%
re
sp
e
c
ti
v
e
ly
.
T
h
e
EE
R
im
p
ro
v
e
m
e
n
t
o
f
th
e
KLDA
tec
h
n
i
q
u
e
-
b
a
se
d
A
S
R
sy
ste
m
c
o
m
p
a
re
d
wit
h
KICA
a
n
d
KPCA
is 4
.
2
5
%
a
n
d
8
.
5
1
%
re
sp
e
c
ti
v
e
l
y
.
K
ey
w
o
r
d
s
:
Au
to
m
atic
s
p
ea
k
er
r
ec
o
g
n
itio
n
s
y
s
tem
Ker
n
el
in
d
ep
e
n
d
en
t c
o
m
p
o
n
en
t
an
aly
s
is
Ker
n
el
lin
ea
r
d
is
cr
im
in
an
t
an
aly
s
is
Ker
n
el
p
r
in
cip
al
c
o
m
p
o
n
en
t
an
aly
s
is
Prin
cip
al
co
m
p
o
n
en
t a
n
al
y
s
is
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
:
Saty
an
an
d
Sin
g
h
,
Sch
o
o
l o
f
E
lectr
ical
a
n
d
E
lectr
o
n
ics E
n
g
in
ee
r
i
n
g
,
Fij
i N
atio
n
al
Un
iv
er
s
is
ty
,
Fiji
.
E
m
ail:
s
aty
an
an
d
.
s
in
g
h
@
f
n
u
.
a
c.
f
j
1.
I
NT
RO
D
UCT
I
O
N
ASR
i
s
im
p
lem
en
ted
u
s
in
g
v
e
r
y
co
n
v
en
tio
n
al
s
tatis
tical
m
o
d
elin
g
tech
n
i
q
u
es
s
u
ch
as
GM
M
o
r
ANN
m
o
d
elin
g
.
B
u
t in
th
e
p
ast f
ew
y
ea
r
s
,
m
ac
h
in
e
lear
n
in
g
th
eo
r
y
h
as e
v
o
lv
ed
in
to
a
v
a
r
iety
o
f
n
ew
alg
o
r
ith
m
s
f
o
r
lear
n
in
g
an
d
class
if
icatio
n
.
T
h
e
s
o
-
ca
lled
k
er
n
el
-
b
ased
m
eth
o
d
,
in
p
ar
ticu
lar
,
h
as
r
ec
en
tly
b
ec
o
m
e
a
p
r
o
m
is
in
g
n
ew
p
ath
to
s
cien
ce
.
Ke
r
n
el
-
b
ased
class
if
icatio
n
an
d
r
e
g
r
es
s
io
n
tech
n
iq
u
es
lik
e
th
e
well
-
k
n
o
wn
SVM
f
o
u
n
d
a
f
air
ly
s
lo
w
ex
p
r
ess
io
n
.
T
h
a
t
m
ay
b
e
b
ec
a
u
s
e
to
ad
d
r
ess
th
eo
r
etica
l
an
d
p
r
ac
tical
p
r
o
b
lem
s
it
n
ee
d
s
to
b
e
ap
p
lied
to
lar
g
e
task
s
s
u
ch
a
s
s
p
ee
ch
r
ec
o
g
n
itio
n
.
R
ec
en
tl
y
,
h
o
we
v
er
,
m
o
r
e
an
d
m
o
r
e
au
th
o
r
s
h
a
v
e
b
ee
n
co
n
ce
r
n
e
d
ab
o
u
t th
e
u
s
e
o
f
s
u
p
p
o
r
t
v
ec
to
r
m
ac
h
in
es in
s
p
ee
ch
r
ec
o
g
n
itio
n
[
1
]
.
B
esid
es
u
s
in
g
k
er
n
el
-
b
ased
cl
ass
if
ier
s
,
an
alter
n
ate
way
is
t
o
u
s
e
k
e
r
n
el
-
b
ased
tech
n
o
l
o
g
i
es
o
n
ly
to
co
n
v
er
t
th
e
f
ea
tu
r
e
s
p
ac
e
an
d
leav
e
th
e
class
if
icatio
n
jo
b
to
m
o
r
e
co
n
v
en
tio
n
al
m
eth
o
d
s
.
T
h
e
p
u
r
p
o
s
e
o
f
th
is
p
ap
er
is
to
s
tu
d
y
th
e
ap
p
licab
ilit
y
o
f
s
o
m
e
o
f
th
ese
m
eth
o
d
s
to
class
if
y
p
h
o
n
em
es,
u
s
in
g
k
er
n
el
-
b
ased
p
r
e
-
lear
n
in
g
s
p
ea
k
er
-
s
p
ec
if
ic
f
ea
tu
r
e
ex
tr
ac
tio
n
m
eth
o
d
s
to
im
p
r
o
v
e
ASR
class
if
icatio
n
r
ates.
T
h
is
p
ap
er
m
ain
ly
d
is
cu
s
s
es KPC
A
[
2
]
,
KI
C
A
[
3
]
,
KL
DA.
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
K
ern
a
l b
a
s
ed
s
p
ea
ke
r
s
p
ec
ific fe
a
tu
r
e
ex
tr
a
ctio
n
a
n
d
its
a
p
p
l
ica
tio
n
s
i
n
…
(
S
a
tya
n
a
n
d
S
in
g
h
)
2489
Usu
ally
,
a
tr
ad
itio
n
al
ASR
p
r
o
ce
s
s
co
n
s
is
ts
o
f
two
p
h
ase
s
:
a
tr
ain
in
g
p
h
ase,
an
d
a
te
s
t
p
h
ase.
I
n
th
e
tr
ain
in
g
p
h
ase,
th
e
d
ev
i
ce
ex
tr
ac
ts
s
p
ea
k
e
r
-
s
p
ec
if
ic
c
h
ar
ac
ter
is
tics
f
r
o
m
t
h
e
s
p
ee
ch
s
ig
n
al
to
b
e
u
s
ed
to
cr
ea
te
a
s
p
ea
k
er
m
o
d
el
[
1
]
,
wh
er
e
th
e
aim
o
f
th
e
test
p
h
ase
is
to
d
eter
m
in
e
th
e
s
p
ea
k
in
g
s
am
p
les
th
at
f
it
th
e
in
d
iv
id
u
al
o
f
t
h
e
tr
ain
in
g
s
am
p
le.
h
e
o
r
i
g
in
al
s
p
ee
ch
s
ig
n
al
is
tr
an
s
f
o
r
m
ed
in
to
a
v
ec
t
o
r
r
ep
r
esen
tatio
n
o
f
th
e
f
u
n
ctio
n
[
2
]
in
all
a
u
d
io
s
ig
n
al
p
r
o
ce
s
s
in
g
.
L
in
ea
r
p
r
e
d
ictio
n
ce
p
s
tr
al
c
o
ef
f
icien
ts
(
L
PC
C
)
an
d
p
e
r
ce
p
tu
a
l
lin
ea
r
ity
p
r
ed
icted
ce
p
s
tr
u
m
c
o
ef
f
icien
t
(
PLPC
C
)
,
m
al
f
r
eq
u
en
cy
ce
p
s
tr
al
co
ef
f
icien
t
(
MFC
C
)
[
3
]
ap
p
r
o
ac
h
is
m
o
s
t
co
m
m
o
n
ly
u
s
ed
in
th
e
ASR
s
y
s
tem
to
o
b
tain
s
p
ea
k
er
-
s
p
ec
if
ic
f
ea
tu
r
es.
Fo
r
m
o
d
elin
g
,
d
is
cr
im
in
an
t
class
if
ier
s
in
s
u
p
p
o
r
t
v
ec
to
r
m
ac
h
in
e
(
SVM)
[
4
]
r
ep
r
esen
tati
o
n
h
a
v
e
ac
h
iev
ed
i
m
p
r
ess
iv
e
r
esu
lts
in
m
an
y
ASR
s
y
s
tem
s
.
SVM
will
d
ef
in
itely
ef
f
ec
tiv
el
y
tr
ai
n
n
o
n
-
lin
ea
r
b
o
u
n
d
a
r
ies
f
o
r
d
ec
is
io
n
-
m
a
k
in
g
b
y
class
if
y
in
g
in
ter
esti
n
g
s
p
ea
k
er
s
/im
p
o
s
ter
s
as th
ey
ar
e
d
is
tin
ct.
Alth
o
u
g
h
th
ese
f
ea
tu
r
e
ex
t
r
ac
tio
n
tech
n
iq
u
es
ar
e
ef
f
ec
tiv
e
,
n
o
n
-
lin
e
ar
m
ap
p
in
g
o
f
s
p
ee
ch
f
ea
tu
r
es
to
n
ew
s
u
itab
le
s
p
ac
es
m
a
y
g
e
n
er
ate
n
ew
f
ea
tu
r
es
th
at
ca
n
b
etter
id
en
tify
s
p
ee
ch
ca
teg
o
r
ies.
Ker
n
el
-
b
ased
tech
n
o
lo
g
y
h
as
b
ee
n
ap
p
lied
to
a
v
ar
iety
o
f
lear
n
in
g
m
ac
h
in
es,
in
clu
d
in
g
s
u
p
p
o
r
t
v
ec
to
r
m
ac
h
in
es
(
SVM)
,
Ker
n
el
d
is
cr
im
in
an
t
an
aly
s
is
(
KDA)
,
k
er
n
el
p
r
i
n
cip
al
co
m
p
o
n
en
t
an
aly
s
is
(
KPC
A)
[
5
]
.
T
h
e
latter
two
m
eth
o
d
s
ar
e
wid
ely
u
s
ed
in
im
ag
e
r
ec
o
g
n
itio
n
.
T
h
eir
p
er
f
o
r
m
an
ce
in
s
p
ea
k
er
r
ec
o
g
n
itio
n
,
h
o
w
ev
er
,
h
as
n
o
t
b
ee
n
ca
r
ef
u
lly
in
v
esti
g
ated
.
T
h
e
p
u
r
p
o
s
e
o
f
th
is
p
ap
er
is
to
ex
a
m
in
e
th
e
ap
p
licab
ilit
y
o
f
s
o
m
e
o
f
th
ese
m
et
h
o
d
s
t
o
class
if
y
p
h
o
n
e
m
es,
u
s
in
g
k
er
n
el
-
b
ased
f
ea
tu
r
e
ex
tr
ac
tio
n
m
eth
o
d
s
ap
p
lied
b
ef
o
r
e
lear
n
in
g
to
b
o
o
s
t
class
if
icatio
n
lev
els.
E
s
s
en
tially
,
th
is
p
ap
er
d
ea
ls
with
th
e
s
tr
ateg
ies
o
f
KPC
A
,
KI
C
A
[
6
,
7
]
,
KL
DA
[
5
]
,
a
n
d
Ke
r
n
el
s
p
r
in
g
y
d
is
cr
im
in
an
t
an
aly
s
is
(
KSDA)
[
8
]
.
I
n
th
is
wo
r
k
,
KPC
A,
KI
C
A,
an
d
KL
DA
is
u
s
ed
f
o
r
s
p
ea
k
er
s
p
ec
if
ic
f
ea
tu
r
e
ex
tr
ac
tio
n
with
an
ASR
s
y
s
tem
.
W
ith
KPC
A,
s
p
ea
k
er
-
s
p
ec
i
f
ic
f
ea
tu
r
es
ca
n
b
e
ex
p
r
ess
ed
i
n
a
h
ig
h
d
im
en
s
io
n
s
p
ac
e
wh
ich
ca
n
p
o
s
s
ib
ly
g
en
e
r
at
e
m
o
r
e
d
is
tin
g
u
is
h
ab
le
s
p
ea
k
er
f
ea
tu
r
es.
2.
F
UNDA
M
E
N
T
AL
O
F
P
RI
N
CIPAL CO
M
P
O
NE
N
T
AN
AL
YS
I
S
Prin
cip
al
co
m
p
o
n
e
n
t
an
aly
s
is
(
PC
A)
i
s
a
v
er
y
co
m
m
o
n
m
eth
o
d
o
f
d
im
en
s
io
n
ality
r
e
d
u
ctio
n
an
d
f
ea
tu
r
e
ex
tr
ac
tio
n
.
PC
A
attem
p
ts
to
f
in
d
lin
ea
r
s
u
b
s
p
ac
es
th
a
t
ar
e
s
m
aller
in
s
ize
th
an
t
h
e
o
r
ig
in
al
f
ea
tu
r
e
s
p
ac
e,
with
n
ew
f
ea
t
u
r
es
h
a
v
in
g
th
e
l
ar
g
est
v
ar
ian
ce
[
9
]
.
C
o
n
s
id
er
t
h
e
d
ata
s
et
{
}
wh
er
e
=
1
,
2
,
3
,
…
.
,
,
ea
ch
is
a
D
-
d
im
en
s
io
n
al
v
ec
t
o
r
.
No
w
we
p
r
o
ject
t
h
e
d
ata
i
n
to
th
e
-
d
im
en
s
io
n
al
s
u
b
s
p
ac
e,
h
er
e,
<
.
T
h
e
p
r
o
jectio
n
is
r
ep
r
esen
ted
as
=
,
wh
er
e
=
[
1
,
2
,
…
.
,
]
,
=
1
f
o
r
=
1
,
2
,
3
,
…
.
,
.
W
e
wan
t to
m
ax
im
ize
th
e
v
a
r
ian
ce
o
f
{
}
,
wh
ich
is
th
e
tr
ac
e
o
f
th
e
co
v
ar
ian
ce
m
atr
ix
o
f
{
}
.
∗
=
a
r
g
(
)
(
1)
wh
er
e
,
=
1
∑
(
−
̅
)
(
−
̅
)
=
1
(
2
)
an
d
̅
=
1
∑
=
1
(
3
)
C
o
v
ar
ien
ce
m
atr
ix
o
f
{
}
is
th
e
,
s
in
ce
(
)
=
(
)
,
b
y
u
s
in
g
th
e
L
ag
r
an
g
ia
n
m
u
ltip
lier
an
d
tak
in
g
th
e
d
e
r
iv
ativ
e,
we
g
et
,
=
(
4
)
wh
ich
in
d
icate
s
is
th
e
eig
en
v
e
cto
r
o
f
an
d
n
o
w
ca
n
b
e
r
e
p
r
esen
ted
as f
o
llo
ws;
=
∑
(
)
=
1
(
5
)
ca
n
b
e
a
p
p
r
o
x
im
ated
b
y
̃
an
d
ex
p
r
ess
ed
as f
o
llo
ws:
̃
=
∑
(
)
=
1
(
6
)
wh
er
e
is
th
e
eig
en
v
ec
to
r
o
f
co
r
r
esp
o
n
d
in
g
t
o
th
e
k
t
h
lar
g
est
eig
en
v
alu
e.
Stan
d
a
r
d
P
C
A
r
esu
lts
f
o
r
th
e
two
-
s
p
ea
k
er
’
s
au
d
io
d
ata
s
h
o
wn
in
Fig
u
r
e
1
(
a)
.
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
4
8
8
-
2497
2490
2
.
1
.
K
er
nel P
CA
m
et
ho
do
lo
g
y
f
o
r
dim
ens
io
na
lity
re
du
ct
io
n
in ASR
s
y
s
t
em
Stan
d
ar
d
PC
A
o
n
ly
allo
ws
lin
ea
r
s
ize
r
ed
u
ctio
n
.
Ho
wev
er
,
s
tan
d
ar
d
PC
A
is
n
o
t
v
er
y
u
s
ef
u
l
wh
en
th
e
d
ata
h
as
a
m
o
r
e
co
m
p
le
x
s
tr
u
ctu
r
e
th
at
ca
n
n
o
t
b
e
r
e
p
r
esen
ted
well
in
lin
ea
r
s
u
b
s
p
ac
es.
Fo
r
tu
n
ately
,
th
e
k
er
n
el
PC
A
allo
ws
u
s
to
e
x
ten
d
th
e
s
tan
d
a
r
d
PC
A
to
n
o
n
lin
ea
r
d
im
en
s
io
n
ality
r
ed
u
ctio
n
[
1
0
]
.
Ass
u
m
e
th
at
a
s
et
o
f
o
b
s
er
v
ati
o
n
s
is
g
iv
en
∈
ℝ
,
=
1
,
2
,
3
,
…
.
,
.
C
o
n
s
id
er
th
e
in
n
er
d
o
t
p
r
o
d
u
ct
s
p
ac
e
ass
o
ciate
d
with
th
e
in
p
u
t
s
p
ac
e
b
y
a
m
a
p
:
ℝ
→
m
ay
b
e
n
o
n
-
lin
ea
r
.
T
h
e
f
ea
tu
r
e
s
p
ac
e
h
as
an
a
r
b
itr
ar
y
s
ize
an
d
in
s
o
m
e
ca
s
es
h
as
an
in
f
in
ite
d
im
en
s
io
n
.
Her
e,
u
p
p
er
ca
s
e
letter
s
u
s
ed
f
o
r
elem
en
ts
o
f
,
an
d
lo
wer
ca
s
e
letter
s
ar
e
u
s
ed
f
o
r
elem
en
ts
o
f
ℝ
.
Su
p
p
o
s
e
we
ar
e
wo
r
k
in
g
o
n
ce
n
te
r
ed
d
ata
∑
(
)
=
0
=
1
.
I
n
F,
t
h
e
co
v
ar
ian
ce
m
atr
ix
h
as th
e
f
o
r
m
as
f
o
llo
ws:
=
1
∑
(
)
(
)
=
1
(
7
)
E
ig
en
v
alu
es
≥
0
an
d
n
o
n
ze
r
o
eig
e
n
v
ec
to
r
s
∈
\
(
0
)
s
atis
f
y
in
g
=
.
I
t
is
well
k
n
o
wn
th
at
all
s
o
lu
tio
n
s
with
≠
0
ar
e
in
th
e
r
an
g
e
o
f
{
(
)
}
=
1
.
T
h
is
h
as
two
co
n
s
eq
u
en
ce
s
.
First,
co
n
s
id
er
a
s
et
o
f
eq
u
atio
n
s
〈
(
)
,
〉
=
〈
(
)
,
〉
,
f
o
r
all
=
1
,
2
,
3
,
…
,
an
d
s
ec
o
n
d
th
er
e
ex
is
t
co
e
f
f
i
cien
ts
,
=
1
,
2
,
3
,
…
,
in
s
u
ch
a
way
th
at
=
∑
(
)
=
1
.
C
o
m
b
in
in
g
〈
(
)
,
〉
=
〈
(
)
,
〉
an
d
=
∑
(
)
=
1
we
g
et
th
e
d
u
al
r
ep
r
esen
tatio
n
o
f
th
e
eig
en
v
alu
e
p
r
o
b
lem
as
1
∑
〈
(
)
,
∑
(
)
〈
(
)
,
(
)
〉
=
1
〉
=
∑
〈
(
)
,
(
)
〉
=
1
=
1
f
o
r
all
=
1
,
2
,
3
,
…
,
.
W
e
ar
e
d
ef
in
in
g
a
m
atr
i
x
b
y
=
(
)
,
(
)
,
th
is
m
ak
es
2
=
.
W
h
er
e
d
en
o
ted
as
a
co
lu
m
n
v
ec
to
r
s
with
1
,
2
,
3
,
…
.
,
en
tr
ies.
L
et
1
≥
2
≥
⋯
b
e
th
e
eig
en
v
alu
e
o
f
,
1
,
2
,
…
,
b
e
th
e
s
et
o
f
co
r
r
esp
o
n
d
in
g
eig
e
n
v
ec
to
r
s
,
an
d
b
e
th
e
last
n
o
n
-
ze
r
o
eig
e
n
v
alu
e.
No
r
m
alizin
g
1
,
2
,
…
,
b
y
n
ee
d
i
n
g
th
e
c
o
r
r
esp
o
n
d
in
g
v
ec
to
r
s
in
b
e
n
o
r
m
alize
d
〈
,
〉
=
1
,
f
o
r
all
=
1
,
2
,
…
,
.
C
o
n
s
id
er
in
g
=
∑
(
)
=
1
an
d
=
,
th
e
n
o
r
m
aliza
tio
n
co
n
d
itio
n
o
f
1
,
2
,
…
,
ca
n
b
e
r
ew
r
itten
as f
o
llo
ws
;
1
=
∑
,
〈
(
)
,
(
)
〉
=
∑
,
,
=
〈
,
〉
=
〈
,
〉
(
8
)
f
o
r
th
e
p
u
r
p
o
s
e
o
f
p
r
in
ci
p
al
co
m
p
o
n
en
t
ex
t
r
ac
tio
n
,
we
n
ee
d
to
co
m
p
u
te
th
e
p
r
o
jectio
n
s
o
n
to
th
e
eig
en
v
ec
to
r
s
in
,
f
o
r
=
1
,
2
,
…
,
.
L
et
b
e
th
e
test
p
o
in
t,
with
a
n
im
a
g
e
(
)
in
.
〈
,
(
)
〉
=
∑
〈
(
)
,
(
)
〉
=
1
(
9
)
〈
,
(
)
〉
n
o
n
lin
ea
r
p
r
in
ci
p
al
co
m
p
o
n
en
t c
o
r
r
esp
o
n
d
in
g
t
o
.
2
.
2
.
Co
m
pu
t
a
t
io
n o
f
co
v
a
ria
nce
m
a
t
rix
a
nd
do
t
pro
du
ct
m
a
t
rix
by
po
s
it
io
nin
g
o
n f
ea
t
ure
s
pa
ce
Fo
r
th
e
s
ak
e
o
f
s
im
p
licity
,
we
ass
u
m
e
th
at
th
e
o
b
s
er
v
atio
n
s
ar
e
a
t
th
e
ce
n
ter
.
T
h
is
is
ea
s
y
to
im
p
lem
en
t
in
th
e
in
p
u
t
s
p
ac
e
b
ec
au
s
e
it
is
n
o
t
p
o
s
s
ib
le
to
ex
p
licitly
ca
lc
u
late
th
e
av
e
r
ag
e
o
f
t
h
e
o
b
s
er
v
atio
n
s
m
ap
p
e
d
with
,
b
u
t
it
is
m
o
r
e
d
if
f
icu
lt
to
u
s
e
.
Ass
u
m
e
th
at
an
y
an
d
a
n
y
s
er
ies
o
f
o
b
s
er
v
atio
n
s
1
,
2
,
…
,
ar
e
g
iv
e
n
th
en
let
u
s
d
ef
in
e
̅
=
1
∑
(
)
=
1
an
d
th
en
th
e
p
o
in
t
̃
(
)
=
(
)
−
̅
will
b
e
ce
n
ter
ed
.
T
h
er
ef
o
r
e
,
th
e
ab
o
v
e
ass
u
m
p
tio
n
h
o
ld
s
,
we
d
ef
in
th
e
co
v
a
r
ian
ce
m
atr
i
x
an
d
th
e
d
o
t
p
r
o
d
u
ct
m
at
r
ix
̃
=
〈
̃
(
)
,
̃
(
)
〉
in
.
W
e
k
n
o
wn
eig
en
v
alu
e
p
r
o
b
lem
s
as
̃
̃
=
̃
̃
with
̃
is
th
e
ex
p
an
s
io
n
co
ef
f
icien
t
o
f
th
e
eig
en
v
ec
to
r
r
elativ
e
to
th
e
ce
n
ter
p
o
in
t
̃
(
)
.
Sin
ce
th
er
e
is
n
o
c
en
tr
al
d
ata,
̃
ca
n
n
o
t
b
e
e
x
p
licitly
ca
lc
u
lated
,
b
u
t
it
ca
n
b
e
r
ep
r
esen
ted
b
y
a
c
o
r
r
esp
o
n
d
in
g
with
o
u
t
a
ce
n
ter
th
e
r
ef
o
r
e
̃
=
〈
̃
(
)
−
,
(
)
−
〉
=
,
−
1
∑
−
1
∑
+
1
2
∑
,
=
1
=
1
=
1
.
W
e
ca
n
g
et
m
o
r
e
co
m
p
ac
t
ex
p
r
e
s
s
io
n
b
y
u
s
in
g
th
e
v
ec
to
r
1
=
(
1
,
…
,
1
)
.
T
h
e
co
m
p
ac
t
ex
p
r
ess
io
n
is
̃
=
−
1
1
1
−
1
1
1
+
1
2
(
1
1
)
1
1
.
W
e
ca
n
ca
lcu
late
̃
f
r
o
m
an
d
s
o
lv
e
th
e
eig
en
v
alu
e
p
r
o
b
lem
.
C
o
n
s
id
er
test
p
o
in
t
p
r
o
jectio
n
o
f
th
e
ce
n
ter
p
o
in
t
o
f
th
e
ce
n
ter
-
im
ag
e
o
f
to
th
e
f
ea
tu
r
e
v
ec
to
r
o
f
th
e
c
o
v
ar
ian
ce
m
atr
ix
is
co
m
p
u
ted
to
f
in
d
its
co
o
r
d
in
ates
[
1
1
]
.
〈
̃
(
)
,
̃
〉
=
〈
(
)
−
⃑
,
̃
〉
=
∑
̃
〈
(
)
−
⃑
,
(
)
−
⃑
〉
=
1
=
∑
̃
=
1
{
(
,
)
−
1
∑
(
,
)
−
1
∑
(
,
)
+
1
2
∑
(
,
)
,
=
1
=
1
=
1
}
(
1
0
)
I
n
tr
o
d
u
cin
g
t
h
e
v
ec
to
r
.
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
K
ern
a
l b
a
s
ed
s
p
ea
ke
r
s
p
ec
ific fe
a
tu
r
e
ex
tr
a
ctio
n
a
n
d
its
a
p
p
l
ica
tio
n
s
i
n
…
(
S
a
tya
n
a
n
d
S
in
g
h
)
2491
=
(
(
,
)
)
1
(
1
1
)
(
〈
̃
(
)
,
̃
〉
)
1
=
̃
−
1
1
̃
−
1
(
1
)
1
̃
+
1
2
(
1
1
)
1
̃
=
(
−
1
1
1
)
̃
−
1
1
(
−
1
1
1
)
̃
=
(
−
1
1
)
(
−
1
1
1
)
̃
(
1
2
)
No
te
th
at
KP
C
A
im
p
licitly
u
s
es
o
n
ly
in
p
u
t
v
ar
iab
les
b
ec
au
s
e
th
e
alg
o
r
ith
m
u
s
es
k
er
n
el
f
u
n
ctio
n
ev
alu
atio
n
to
r
ep
r
esen
t
th
e
r
e
d
u
ctio
n
in
f
ea
t
u
r
e
s
p
ac
e
d
im
en
s
io
n
s
.
T
h
er
ef
o
r
e,
KPC
A
is
u
s
ef
u
l
f
o
r
n
o
n
lin
e
ar
f
ea
tu
r
e
ex
t
r
ac
tio
n
b
y
r
ed
u
cin
g
t
h
e
s
ize;
it d
o
es n
o
t e
x
p
lain
th
e
ch
ar
ac
ter
is
tics
o
f
th
e
in
p
u
t v
ar
iab
le
s
elec
tio
n
.
3.
K
E
RNE
L
-
B
A
SE
D
SPEAK
E
R
SPE
CIFI
C
F
E
A
T
UR
E
E
XT
RAC
T
I
O
N
A
ND
I
T
S AP
P
L
I
CA
T
I
O
N
I
N
ASR
C
las
s
if
icat
io
n
alg
o
r
ith
m
s
m
u
s
t
r
ep
r
esen
t
th
e
o
b
jects
to
b
e
cl
ass
if
ied
as
p
o
in
ts
in
a
m
u
ltid
im
en
s
io
n
al
f
ea
tu
r
e
s
p
ac
e.
Ho
wev
er
,
o
n
e
c
an
ap
p
l
y
o
t
h
er
v
ec
to
r
s
p
ac
e
tr
a
n
s
f
o
r
m
atio
n
s
to
th
e
in
itial
f
ea
t
u
r
es
b
ef
o
r
e
r
u
n
n
in
g
th
e
lear
n
in
g
alg
o
r
ith
m
.
T
h
e
r
e
ar
e
two
r
ea
s
o
n
s
f
o
r
d
o
in
g
t
h
is
.
First,
th
ey
ca
n
im
p
r
o
v
e
th
e
p
er
f
o
r
m
a
n
ce
o
f
class
if
icatio
n
an
d
s
ec
o
n
d
,
th
e
y
ca
n
r
ed
u
ce
th
e
d
ata'
s
d
im
en
s
io
n
ality
.
T
h
e
s
elec
tio
n
o
f
i
n
itial
f
ea
tu
r
es
an
d
th
eir
tr
an
s
f
o
r
m
atio
n
ar
e
s
o
m
etim
e
s
d
ea
lt
with
in
th
e
liter
atu
r
e
u
n
d
er
t
h
e
titl
e
"f
ea
tu
r
e
e
x
tr
ac
tio
n
”.
T
o
av
o
id
m
is
u
n
d
er
s
tan
d
in
g
,
th
is
s
ec
tio
n
d
escr
ib
es
o
n
ly
th
e
latter
an
d
d
escr
ib
es
th
e
f
ir
s
t
f
ea
tu
r
e
s
et.
Ho
p
ef
u
lly
it
will
b
e
m
o
r
e
ef
f
ec
tiv
e
an
d
class
if
icatio
n
will b
e
f
aster
.
T
h
e
a
p
p
r
o
ac
h
to
th
e
ex
tr
ac
tio
n
o
f
f
ea
tu
r
es
m
ay
b
e
eith
er
lin
ea
r
o
r
n
o
n
lin
ea
r
,
b
u
t
th
er
e
is
a
tech
n
iq
u
e
th
at
b
r
ea
k
s
d
o
wn
th
e
b
ar
r
ier
b
etwe
en
th
e
two
f
o
r
m
s
in
s
o
m
e
way
.
T
h
e
k
ey
id
ea
b
eh
in
d
th
e
k
er
n
el
tech
n
iq
u
e
was
o
r
ig
in
ally
p
r
esen
ted
in
[
1
2
]
an
d
ap
p
lied
ag
ain
in
co
n
n
ec
tio
n
with
th
e
g
en
er
al
p
u
r
p
o
s
e
SVM
[
1
3
-
1
5
]
f
o
llo
wed
b
y
o
th
e
r
k
er
n
el
-
b
ased
m
eth
o
d
s
.
3.
1
.
Su
pp
ly
ing
inp
ut
v
a
ria
ble inf
o
rm
a
t
i
o
n into
k
er
nel
P
C
A
A
d
d
itio
n
al
in
f
o
r
m
atio
n
to
t
h
e
KPC
A
r
ep
r
esen
tatio
n
f
o
r
in
ter
p
r
etab
ilit
y
.
W
e
h
av
e
d
e
v
elo
p
e
d
a
p
r
o
ce
s
s
to
p
r
o
ject
a
g
iv
en
in
p
u
t
v
ar
iab
le
in
to
a
s
u
b
s
p
ac
e
s
p
an
n
ed
b
y
f
ea
tu
r
e
v
ec
to
r
s
̃
=
∑
̃
̃
(
1
)
=
1
.
W
e
ca
n
th
in
k
o
f
o
u
r
o
b
s
er
v
atio
n
as
a
r
an
d
o
m
v
ec
to
r
=
(
1
,
2
,
…
.
.
,
)
im
p
lem
en
tatio
n
th
en
t
o
r
ep
r
esen
t
th
e
p
r
o
m
i
n
en
ce
o
f
th
e
in
p
u
t
v
ar
iab
le
in
th
e
KPC
A.
C
o
n
s
id
er
in
g
a
s
et
o
f
p
o
i
n
ts
o
f
m
ath
em
atica
l
f
o
r
m
s
=
+
∈
ℝ
wh
er
e
=
(
0
,
…
.
,
1
,
…
.
,
0
)
o
f
k
th
co
m
p
o
n
en
t
is
eith
er
0
o
r
1
.
Nex
t,
th
e
p
r
o
jectio
n
p
o
in
ts
(
)
o
f
th
ese
im
ag
es
o
n
to
th
e
s
u
b
s
p
ac
e
s
p
an
n
ed
b
y
th
e
f
ea
tu
r
e
v
ec
t
o
r
̃
=
∑
̃
̃
(
1
)
=
1
ca
n
b
e
ca
lcu
lated
.
C
o
n
s
id
er
in
g
i
n
(
1
2
)
th
e
r
o
w
v
ec
to
r
g
iv
es th
e
i
n
d
u
c
tio
n
cu
r
v
e
in
th
e
E
i
g
en
s
p
ac
e
e
x
p
r
ess
ed
in
m
atr
ix
f
o
r
m
:
(
)
1
=
(
−
1
1
)
(
−
1
1
1
)
̃
(
1
3
)
Fu
r
th
er
m
o
r
e
,
b
y
p
r
o
jectin
g
th
e
tan
g
en
t
v
ec
to
r
to
s
=
0
,
we
ca
n
ex
p
r
ess
th
e
m
ax
im
u
m
ch
a
n
g
e
d
ir
ec
tio
n
o
f
(
)
ass
o
ciate
d
with
th
e
v
ar
iab
l
e
.
Ma
tr
ix
f
o
r
m
o
f
th
e
ex
p
r
ess
io
n
r
ep
r
esen
ted
as f
o
llo
ws:
|
=
0
=
|
=
0
(
−
1
1
1
)
̃
(
1
4
)
wh
er
e
|
=
0
=
(
1
|
=
0
,
…
…
.
.
,
|
=
0
)
an
d
|
=
0
=
(
,
)
|
=
0
=
(
∑
(
,
)
=
1
)
|
=
0
=
∑
(
,
)
=
1
|
=
=
(
,
)
|
=
wh
er
e
d
elta
o
f
Kr
o
n
ec
k
er
is
r
ep
r
esen
ted
as
an
d
r
ad
ial
b
asis
k
er
n
el
as
(
,
)
=
(
−
‖
−
‖
2
)
=
(
−
∑
(
−
)
2
=
1
)
.
Af
ter
co
n
s
id
er
i
n
g
=
+
∈
ℝ
:
|
=
0
=
(
,
)
|
=
=
−
2
(
,
)
(
−
)
=
−
2
(
,
)
(
−
)
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
4
8
8
-
2497
2492
wh
er
e
th
e
t
r
ain
in
g
p
o
i
n
t
=
.
T
h
u
s
,
b
y
ap
p
ly
i
n
g
(
1
3
)
,
it
is
p
o
s
s
ib
le
to
lo
ca
lly
r
ep
r
esen
t
a
n
y
g
iv
en
in
p
u
t
v
ar
iab
le
p
lo
t
in
KPC
A.
Fu
r
th
er
m
o
r
e,
b
y
u
s
in
g
(
1
4
)
,
it
is
p
o
s
s
ib
le
to
r
ep
r
esen
t
t
h
e
tan
g
e
n
t
v
e
cto
r
ass
o
ciate
d
with
an
y
g
iv
en
in
p
u
t
v
ar
iab
le
at
ea
c
h
s
am
p
le
p
o
in
t
[
1
6
]
.
T
h
e
r
ef
o
r
e
,
a
v
ec
to
r
f
ield
ca
n
b
e
d
r
aw
n
o
n
KPC
A
in
d
icatin
g
th
e
g
r
o
wth
d
ir
ec
tio
n
o
f
a
g
iv
e
n
v
ar
iab
le.
T
h
er
e
ar
e
s
o
m
e
ex
is
tin
g
tech
n
iq
u
es
to
co
m
p
u
te
z
f
o
r
s
p
ec
if
ic
k
er
n
els
[
1
7
]
.
Fo
r
a
Gau
s
s
ian
k
er
n
el
(
,
)
)
=
(
−
‖
−
‖
2
/
2
2
)
,
m
u
s
t satis
f
y
th
e
f
o
llo
win
g
c
o
n
d
itio
n
;
=
∑
(
‖
−
‖
2
/
2
2
)
=
1
∑
=
1
(
−
‖
−
‖
2
)
/
2
2
(
1
5
)
Ker
n
el
PC
A
r
esu
lts
f
o
r
th
e
two
-
s
p
ea
k
er
’
s
a
u
d
io
d
ata
with
is
s
h
o
wn
in
Fig
u
r
e
1
(
b
)
.
(
a)
(
b
)
Fig
u
r
e
1
.
Stan
d
a
r
d
PC
A
an
d
K
er
n
el
PC
A
r
esu
lts
f
o
r
th
e
two
-
s
p
ea
k
er
’
s
au
d
io
d
ata
;
(
a)
s
tan
d
ar
d
PC
A
an
d
(
b
)
k
er
n
el
PC
A
3
.
2
.
Appl
ica
t
io
n o
f
k
er
nel independ
e
nt
co
m
po
nent
a
na
ly
s
is
(
K
I
CA)
I
n
d
ep
e
n
d
en
t
c
o
m
p
o
n
en
t
an
al
y
s
is
i
s
a
g
en
er
al
s
tatis
tical
ap
p
r
o
ac
h
o
r
ig
in
ally
b
o
r
n
f
r
o
m
th
e
s
tu
d
y
o
f
s
ep
ar
atio
n
f
r
o
m
b
lin
d
s
o
u
r
c
es.
An
o
th
er
a
p
p
licatio
n
o
f
I
C
A
is
th
e
u
n
s
u
p
er
v
is
ed
ex
tr
ac
tio
n
o
f
f
ea
tu
r
es.
T
h
is
is
in
ten
d
ed
t
o
tr
an
s
f
o
r
m
in
p
u
t
d
ata
lin
ea
r
ly
in
to
u
n
c
o
r
r
elate
d
ele
m
en
ts
,
u
s
in
g
at
least
a
d
is
tr
ib
u
tio
n
o
f
th
e
Gau
s
s
ian
s
am
p
le
s
et
[
1
8
]
.
T
h
e
ex
p
la
n
atio
n
f
o
r
th
is
is
th
at
class
if
icati
o
n
o
f
d
ata
i
n
ce
r
tain
d
ir
ec
tio
n
s
wo
u
ld
b
e
s
im
p
ler
.
T
h
is
is
in
ac
c
o
r
d
an
ce
with
t
h
e
m
o
s
t p
o
p
u
lar
s
p
ee
c
h
m
o
d
elin
g
tech
n
iq
u
e,
i.e
.
f
itti
n
g
Gau
s
s
ian
m
ix
tu
r
es o
n
ea
ch
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
K
ern
a
l b
a
s
ed
s
p
ea
ke
r
s
p
ec
ific fe
a
tu
r
e
ex
tr
a
ctio
n
a
n
d
its
a
p
p
l
ica
tio
n
s
i
n
…
(
S
a
tya
n
a
n
d
S
in
g
h
)
2493
class
.
T
h
is
o
b
v
io
u
s
ly
m
ea
n
s
t
h
at
Gau
s
s
ian
m
ix
tu
r
es
ca
n
ap
p
r
o
x
im
ate
th
e
d
is
tr
ib
u
tio
n
s
o
f
th
e
g
r
o
u
p
s
KI
C
A
ex
ten
d
s
th
is
b
y
ass
u
m
in
g
,
o
n
th
e
co
n
tr
ar
y
,
th
at
wh
en
all
class
es
ar
e
f
u
s
ed
,
th
e
d
is
tr
ib
u
tio
n
is
n
o
t
Gau
s
s
ian
;
th
u
s
,
u
s
in
g
n
o
n
-
Ga
u
s
s
ian
is
m
as
a
h
eu
r
is
tic
f
o
r
th
e
u
n
co
n
t
r
o
lled
e
x
tr
ac
tio
n
o
f
f
ea
tu
r
es
wo
u
ld
p
r
ef
er
th
o
s
e
d
ir
ec
tio
n
s
wh
ich
s
ep
ar
ate
class
es.
Sev
er
al
o
b
jectiv
e
f
u
n
ctio
n
s
f
o
r
o
p
tim
al
s
elec
tio
n
o
f
in
d
e
p
en
d
en
t
d
ir
ec
tio
n
s
wer
e
d
escr
ib
ed
u
s
in
g
ap
p
r
o
x
im
ately
eq
u
iv
alen
t
a
p
p
r
o
ac
h
es.
T
h
e
KI
C
A
alg
o
r
ith
m
'
s
g
o
al
its
elf
is
to
f
in
d
s
u
ch
o
b
j
ec
tiv
e
f
u
n
ctio
n
s
as
o
p
tim
ally
as
p
o
s
s
ib
le
[
1
9
]
.
Fo
r
KI
C
A
o
u
tp
u
t
m
o
s
t
iter
a
tiv
e
m
eth
o
d
s
ar
e
av
ailab
le
.
Oth
er
s
n
ee
d
to
b
e
p
r
ep
r
o
ce
s
s
e
d
,
i.e
.
f
o
cu
s
ed
an
d
wh
iten
ed
wh
ile
o
th
er
s
d
o
n
o
t.
Ov
e
r
all,
ex
p
er
ien
ce
s
h
o
ws
th
at
all
o
f
th
ese
alg
o
r
ith
m
s
ca
n
c
o
n
v
er
g
e
f
aste
r
with
o
r
ien
ted
an
d
w
h
itewash
ed
d
ata,
ev
e
n
th
o
s
e
th
at
d
o
n
'
t r
ea
lly
n
ee
d
it
[
2
0
]
.
L
et's f
ir
s
t in
v
esti
g
ate
h
o
w
th
e
ce
n
ter
in
g
a
n
d
wh
iten
i
n
g
p
r
e
-
p
r
o
ce
s
s
in
g
s
tep
s
ca
n
b
e
d
o
n
e
in
th
e
k
er
n
el
f
u
n
ctio
n
s
p
ac
e.
T
o
th
is
en
d
,
allo
w
th
e
k
er
n
el
f
u
n
ctio
n
in
ℱ
to
im
p
licitly
d
ef
in
e
th
e
in
n
er
p
r
o
d
u
ct
with
th
e
ass
o
ciate
d
tr
an
s
f
o
r
m
atio
n
.
Step
o
n
e
C
en
ter
in
g
ℱ
-
Sh
if
tin
g
th
e
d
ata
(
1
)
,
(
2
)
,
…
.
.
,
(
)
alo
n
g
with
its
m
ea
n
{
(
)
}
to
g
et
th
e
d
ata
as f
o
ll
o
ws:
{
′
(
1
)
=
(
1
)
−
{
(
)
}
′
(
2
)
=
(
2
)
−
{
(
)
}
.
.
′
(
)
=
(
)
−
{
(
)
}
(
1
6
)
Step
two
W
h
iten
in
g
i
n
ℱ
.
T
r
an
s
f
o
p
r
m
i
n
g
th
e
ce
n
ter
e
d
s
am
p
le
s
′
(
1
)
,
′
(
2
)
,
…
.
,
′
(
)
v
ia
an
o
r
t
h
o
g
o
n
al
tr
an
s
f
o
r
m
atio
n
in
to
its
v
ec
to
r
s
̂
(
1
)
=
′
(
1
)
,
′
(
2
)
,
…
,
′
(
)
=
(
′
)
.
̂
=
is
th
e
co
v
ar
ian
ce
m
atr
ix
.
B
ec
au
s
e
s
tan
d
ar
d
PC
A
co
n
v
er
ts
th
e
co
v
ar
ia
n
ce
m
a
tr
ix
in
to
a
d
iag
o
n
al
f
o
r
m
ju
s
t
lik
e
its
k
er
n
el
b
ased
eq
u
iv
alen
t,
wh
er
e
th
e
d
ia
g
o
n
a
l
elem
en
ts
ar
e
th
e
u
n
iq
u
e
v
alu
es
o
f
th
e
d
ata
co
v
ar
ian
ce
m
atr
ix
{
̂
(
)
̂
(
)
}
,
all
th
at
r
em
ain
s
is
to
tr
an
s
f
o
r
m
th
e
d
iag
o
n
al
elem
en
t
i
n
to
1
.
B
ased
o
n
th
is
f
in
d
in
g
,
a
s
lig
h
t
m
o
d
if
icatio
n
o
f
th
e
f
o
r
m
u
las
p
r
o
v
id
e
d
in
th
e
KPC
A
s
ec
tio
n
will
o
b
tain
th
e
n
ec
ess
ar
y
wh
iten
in
g
tr
an
s
f
o
r
m
atio
n
[
2
1
]
.
Her
e
(
1
1
)
,
(
2
2
)
,
…
.
.
,
(
)
an
d
1
≥
2
≥
ar
e
th
e
eig
h
p
air
s
o
f
{
̂
(
)
̂
(
)
}
th
en
th
e
tr
an
s
f
o
r
m
atio
n
m
atr
ix
will
tak
e
a
f
o
r
m
[
1
−
1
2
1
,
2
−
1
2
2
,
…
.
.
−
1
2
]
.
Ker
n
el
I
n
d
ep
e
n
d
en
t
co
m
p
o
n
en
t
an
aly
s
is
r
esu
lt
s
f
o
r
th
e
two
-
s
p
ea
k
er
’
s
au
d
io
d
ata
i
s
s
h
o
wn
in
Fig
u
r
e
2
(
a)
.
3
.
3
.
Appl
ica
t
io
n o
f
k
er
nel linea
r
dis
cr
im
ina
nt
a
na
ly
s
is
(
K
L
DA)
L
DA
is
a
co
n
v
en
tio
n
al,
s
u
p
e
r
v
is
ed
m
eth
o
d
o
f
ex
t
r
ac
tin
g
s
p
ea
k
er
-
s
p
ec
if
ic
c
h
ar
ac
ter
is
tics
[
1
9
]
th
at
h
as
p
r
o
v
e
n
to
b
e
o
n
e
o
f
t
h
e
m
o
s
t
ef
f
ec
tiv
e
p
r
e
-
p
r
o
ce
s
s
in
g
class
if
icatio
n
tech
n
iq
u
es.
I
t
h
as
a
ls
o
lo
n
g
b
ee
n
u
s
ed
in
s
p
ee
ch
r
e
co
g
n
itio
n
[
2
2
]
.
T
h
e
m
ain
g
o
al
o
f
L
DA
is
to
f
in
d
a
n
ew
o
r
th
o
g
o
n
al
d
ata
s
et
to
p
r
o
v
id
e
t
h
e
o
p
tim
al
class
s
ep
ar
atio
n
.
I
n
KL
DA
we
ar
e
ess
en
tially
f
o
llo
win
g
th
e
d
is
cu
s
s
io
n
o
f
its
lin
ea
r
co
u
n
ter
p
ar
t,
ex
ce
p
t i
n
th
i
s
ca
s
e
th
is
is
in
ten
d
ed
t
o
h
a
p
p
en
im
p
lici
tly
in
th
e
k
er
n
el
f
ea
t
u
r
e
s
p
ac
e
F.
L
et'
s
s
ay
ag
ain
th
at
a
k
e
r
n
el
f
u
n
ctio
n
with
a
f
ea
tu
r
e
m
ap
an
d
a
k
er
n
el
f
iel
d
s
p
ac
e
h
as
b
ee
n
ch
o
s
en
.
I
n
o
r
d
er
to
d
ef
in
e
th
e
tr
a
n
s
f
o
r
m
atio
n
m
atr
ix
o
f
KL
DA,
we
d
ef
in
e
th
e
o
b
jectiv
e
f
u
n
ctio
n
f
ir
s
t
as
Γ
∶
ℱ
→
ℛ
,
b
ec
au
s
e
o
f
th
e
s
u
p
er
v
is
ed
n
atu
r
e
o
f
th
is
m
eth
o
d
,
it
d
ep
en
d
s
n
o
t o
n
l
y
o
n
t
h
e
test
d
ata
b
u
t a
l
s
o
o
n
th
e
in
d
icato
r
ℒ
.
L
et'
s
d
esc
r
ib
e
u
b
iq
u
ito
u
s
Γ
(
V
)
.
Γ
(
V
)
=
ℬ
,
∈
:
ℱ
\
{
0
}
(
1
7
)
wh
er
e
ℬ
is
th
e
s
ca
tter
m
atr
ix
o
f
th
e
in
ter
class
,
wh
ile
is
th
e
s
ca
tter
m
atr
ix
o
f
th
e
in
ter
class
.
Her
e,
th
e
s
ca
tter
m
atr
ix
ℬ
b
etwe
en
class
es sh
o
ws
th
e
s
ca
tter
o
f
th
e
m
e
an
v
ec
t
o
r
s
ar
o
u
n
d
th
e
o
v
er
all
m
ea
n
v
ec
to
r
.
ℬ
=
∑
(
−
)
=
1
(
−
)
;
=
1
∑
(
)
=
;
=
1
∑
(
)
ℒ
(
)
(
1
8
)
with
th
e
class
lab
el
,
th
e
i
n
-
cla
s
s
s
ca
tter
m
atr
ix
r
ep
r
esen
ts
th
e
weig
h
ted
av
er
ag
e
s
ca
tter
o
f
t
h
e
s
am
p
le
v
ec
to
r
co
v
ar
ian
ce
m
at
r
ices
.
=
∑
;
=
1
=
1
∑
(
(
)
−
)
(
(
−
)
)
ℒ
(
)
=
(
1
9
)
Ker
n
el
lin
ea
r
d
is
cr
im
in
an
t a
n
a
ly
s
is
r
esu
lts
f
o
r
th
e
two
-
s
p
ea
k
er
’
s
au
d
io
d
ata
is
s
h
o
wn
in
Fig
u
r
e
2
(
b
)
.
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
4
8
8
-
2497
2494
(
a)
(
b
)
Fig
u
r
e
2
.
Ker
n
el
in
d
e
p
en
d
e
n
t
co
m
p
o
n
en
t a
n
aly
s
is
an
d
k
er
n
e
l lin
ea
r
d
is
cr
im
in
an
t
an
aly
s
is
r
esu
lts
f
o
r
th
e
two
-
s
p
ea
k
er
’
s
au
d
io
d
ata
;
(
a)
k
er
n
el
i
n
d
ep
en
d
en
t c
o
m
p
o
n
en
t a
n
aly
s
is
an
d
(
b
)
k
er
n
el
ln
ea
r
d
is
cr
im
in
an
t a
n
aly
s
is
4.
E
XP
E
R
I
M
E
N
T
A
L
SE
T
UP
T
o
ev
alu
ate
th
e
e
f
f
icien
cy
o
f
k
er
n
el
-
b
ased
s
p
ea
k
er
-
s
p
ec
if
ic
f
ea
tu
r
e
ex
tr
ac
tio
n
tech
n
iq
u
es,
an
is
o
lated
wo
r
d
r
ec
o
g
n
itio
n
e
x
p
er
im
e
n
t
was
p
er
f
o
r
m
ed
.
T
h
e
e
x
p
er
im
en
t
in
clu
d
es
5
2
0
J
ap
an
ese
w
o
r
d
s
f
r
o
m
th
e
AT
R
J
ap
an
ese
C
lan
g
u
ag
e
s
et
Vo
ic
e
d
atab
ase,
8
0
s
p
ea
k
er
s
(
4
0
m
en
an
d
4
0
Fem
ale)
.
Au
d
io
s
am
p
les
o
f
1
0
iTa
u
k
ei
s
p
ea
k
er
s
wer
e
c
o
llected
at
r
a
n
d
o
m
an
d
u
n
d
er
u
n
f
av
o
u
r
a
b
le
co
n
d
itio
n
s
.
T
h
e
a
v
er
ag
e
d
u
r
a
tio
n
o
f
th
e
tr
ain
in
g
s
am
p
les
was
s
ix
s
ec
o
n
d
s
p
er
s
p
ea
k
er
f
o
r
all
1
0
s
p
ea
k
e
r
s
an
d
o
u
t
o
f
twen
ty
u
tter
an
ce
s
o
f
ea
ch
s
p
ea
k
er
j
u
s
t
o
n
e
was
u
s
ed
f
o
r
tr
ain
in
g
p
u
r
p
o
s
e
[
2
3
-
2
7
]
.
Fo
r
m
atch
in
g
p
u
r
p
o
s
e
s
th
e
r
em
ain
in
g
1
9
v
o
ice
s
am
p
les
wer
e
u
s
ed
f
r
o
m
th
e
co
r
p
u
s
.
W
e
h
av
e
r
ec
o
r
d
ed
u
tter
an
ce
s
f
o
r
t
h
is
in
v
esti
g
atio
n
wer
e
at
o
n
e
s
itti
n
g
f
o
r
ea
c
h
s
p
ea
k
er
.
T
h
e
tex
t
f
o
r
th
e
u
tter
an
ce
s
was
r
an
d
o
m
ly
s
elec
ted
b
y
s
p
ea
k
er
.
T
h
e
m
ain
v
o
ice
r
ec
o
r
d
in
g
s
co
n
s
is
t
o
f
b
o
th
m
ale
a
n
d
f
em
ale
s
p
ea
k
er
s
o
f
twen
ty
u
tter
an
ce
o
f
ea
ch
u
s
in
g
s
am
p
lin
g
r
ate
o
f
1
6
k
Hz
with
1
6
b
its
/s
am
p
le.
T
h
r
o
u
g
h
o
u
t
th
e
ex
p
er
im
e
n
t,
1
0
4
0
0
u
tter
a
n
ce
s
wer
e
u
s
ed
as
tr
ain
in
g
d
ata
an
d
th
e
r
em
ai
n
in
g
3
1
,
2
0
0
u
tter
an
ce
s
wer
e
u
s
ed
as
test
d
a
ta.
T
h
e
s
am
p
lin
g
r
ate
o
f
t
h
e
au
d
io
s
ig
n
al
is
1
0
k
Hz.
1
2
Me
l
-
C
ep
s
tr
al
co
ef
f
icien
ts
ex
tr
ac
ted
u
s
in
g
2
5
.
6
m
s
Ham
m
in
g
win
d
o
ws
with
1
0
m
s
s
h
if
ts
[
2
8
-
3
2
]
.
T
h
e
f
ea
tu
r
es
o
f
K
PC
A
wer
e
ex
tr
ac
ted
f
r
o
m
1
3
Me
l
-
ce
p
s
tr
al
c
o
ef
f
icie
n
ts
in
clu
d
in
g
ze
r
o
c
o
ef
f
icien
ts
co
r
r
esp
o
n
d
i
n
g
t
o
3
9
v
ec
to
r
co
ef
f
icien
ts
an
d
th
eir
in
cr
em
en
t
an
d
ac
ce
ler
atio
n
co
ef
f
icien
ts
.
Ar
o
u
n
d
1
,
0
0
0
,
0
0
0
f
r
am
es
wer
e
u
s
ed
as
tr
ain
in
g
d
ata
in
th
is
ex
p
er
im
en
t,
an
d
it
is
co
m
p
u
tat
io
n
ally
im
p
o
s
s
ib
le
to
ca
lcu
late
m
at
r
ix
K
with
th
is
am
o
u
n
t
of
d
ata.
N
f
r
am
es
ar
e
r
an
d
o
m
l
y
p
ick
ed
f
r
o
m
t
h
e
tr
ain
in
g
d
ata
to
r
e
d
u
ce
th
e
n
u
m
b
e
r
o
f
f
r
am
es.
T
h
e
n
u
m
b
er
N=
1
0
2
4
was
ch
o
s
en
t
o
m
ak
e
th
e
s
y
s
tem
co
m
p
u
tatio
n
ally
f
ea
s
ib
le.
T
ab
le
1
r
ep
r
esen
ts
ef
f
icien
c
y
a
n
d
E
E
R
o
f
th
e
ASR
s
y
s
tem
f
o
r
KL
DA,
KI
C
A
an
d
KL
DA
r
e
s
p
ec
tiv
ely
f
o
r
AT
R
J
ap
an
ese
C
lan
g
u
ag
e
.
T
ab
le
2
r
ep
r
esen
ts
ef
f
icien
c
y
an
d
E
E
R
o
f
th
e
ASR
s
y
s
te
m
f
o
r
KL
DA,
KI
C
A
an
d
KL
DA
r
esp
ec
tiv
ely
f
o
r
1
0
iTa
u
k
ei
s
p
ea
k
er
s
c
r
o
s
s
lan
g
u
ag
e
.
Fig
u
r
e
3
s
h
o
w
th
e
eq
u
al
er
r
o
r
r
ate
(
E
E
R
)
o
f
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
K
ern
a
l b
a
s
ed
s
p
ea
ke
r
s
p
ec
ific fe
a
tu
r
e
ex
tr
a
ctio
n
a
n
d
its
a
p
p
l
ica
tio
n
s
i
n
…
(
S
a
tya
n
a
n
d
S
in
g
h
)
2495
KL
DA
KI
C
A,
an
d
KPC
A
b
as
ed
m
o
d
elin
g
tech
n
iq
u
e.
T
h
e
ASR
ef
f
icien
cy
o
f
o
f
KL
DA
KI
C
A,
an
d
KPC
A
b
ased
m
o
d
elin
g
tec
h
n
iq
u
e
ar
e
9
9
.
9
%,
9
9
.
6
%,
an
d
9
8
.
1
%
an
d
E
E
R
ar
e
4
.
7
%,
4
.
9
%
an
d
5
.
1
%
r
esp
ec
tiv
ely
f
o
r
6
s
ec
o
f
au
d
io
s
ig
n
al.
T
h
e
E
E
R
im
p
r
o
v
em
en
t
o
f
KL
DA
tech
n
iq
u
e
b
as
ed
ASR
s
y
s
tem
co
m
p
ar
ed
with
KI
C
A
an
d
KPC
A
is
4
.
2
5
% a
n
d
8
.
5
1
% r
esp
ec
tiv
ely
.
(
a)
(
b
)
Fig
u
r
e
3
.
E
E
R
o
f
KL
DA,
KI
C
A
an
d
KL
DA
tech
n
iq
u
e
f
o
r
6
s
ec
o
f
v
o
ice
d
ata
;
(
a)
AT
R
J
ap
an
ese
C
lan
g
u
ag
e
an
d
(
b
)
iTa
u
k
ei
s
p
ea
k
er
’
s
cr
o
s
s
lan
g
u
ag
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.
18
,
No
.
5
,
Octo
b
e
r
2
0
2
0
:
2
4
8
8
-
2497
2496
T
ab
le
1
.
E
f
f
icien
cy
a
n
d
E
E
R
o
f
th
e
ASR
s
y
s
tem
f
o
r
KL
DA,
KI
C
A
an
d
KL
DA
r
esp
ec
tiv
ely
f
o
r
AT
R
J
ap
an
ese
C
lan
g
u
ag
e
K
LD
A
K
I
C
A
K
P
C
A
Ef
f
i
c
i
e
n
c
y
i
n
%
EER
i
n
%
Ef
f
i
c
i
e
n
c
y
i
n
%
EER
i
n
%
Ef
f
i
c
i
e
n
c
y
i
n
%
EER
i
n
%
6
se
c
9
9
.
7
1
.
9
5
9
9
.
6
2
.
3
1
9
9
.
1
3
.
4
1
4
se
c
9
9
.
5
2
.
2
9
9
9
.
1
3
.
2
0
9
8
.
2
4
.
1
1
2
se
c
9
8
.
8
3
.
2
3
9
8
.
3
4
.
3
2
9
7
.
6
5
.
3
T
ab
le
2
.
E
f
f
icien
cy
a
n
d
E
E
R
o
f
th
e
ASR
s
y
s
tem
f
o
r
KL
DA,
KI
C
A
an
d
KL
DA
r
esp
ec
tiv
ely
f
o
r
1
0
iTa
u
k
ei
s
p
ea
k
er
s
cr
o
s
s
lan
g
u
ag
e
K
LD
A
K
I
C
A
K
P
C
A
Ef
f
i
c
i
e
n
c
y
i
n
%
EER
i
n
%
Ef
f
i
c
i
e
n
c
y
i
n
%
EER
i
n
%
Ef
f
i
c
i
e
n
c
y
i
n
%
EER
i
n
%
6
se
c
9
4
.
9
2
.
0
4
9
4
.
6
3
.
4
9
4
.
1
4
.
1
4
se
c
9
4
.
3
2
.
3
4
9
4
.
1
3
.
7
9
3
.
5
4
.
8
2
se
c
9
3
.
5
3
.
2
.
0
9
3
.
1
4
.
1
9
2
.
6
5
.
3
5.
CO
NCLU
SI
O
N
An
ex
p
er
im
en
tal
ev
al
u
atio
n
o
f
th
e
p
er
f
o
r
m
an
ce
o
f
th
e
ASR
s
y
s
tem
h
as
b
ee
n
d
o
n
e
o
n
6
s
ec
o
f
v
o
ice
d
ata
o
f
AT
R
J
ap
an
ese
C
lan
g
u
ag
e.
Fo
r
t
h
e
1
0
4
0
0
,
v
o
ice
s
a
m
p
les
o
f
th
e
AT
R
J
ap
an
ese
C
lan
g
u
ag
e
s
p
ea
k
er
r
ec
o
g
n
itio
n
ac
c
u
r
ac
y
9
9
.
7
%,
9
9
.
6
%,
an
d
9
9
.
1
%
an
d
eq
u
al
er
r
o
r
r
ate
(
E
E
R
)
is
1
.
9
5
%,
2
.
3
1
%,
an
d
3
.
4
1
%
r
esp
ec
tiv
ely
f
o
r
KL
DA,
KI
C
A,
an
d
KPC
A.
T
h
e
E
E
R
im
p
r
o
v
em
en
t
o
f
th
e
KL
DA
tech
n
iq
u
e
-
b
ased
ASR
s
y
s
tem
co
m
p
ar
ed
with
KI
C
A
an
d
KP
C
A
i
s
4
.
2
5
%
an
d
8
.
5
1
%
r
esp
ec
tiv
ely
.
W
e
f
in
d
th
at
n
o
n
-
lin
ea
r
tr
an
s
f
o
r
m
atio
n
s
u
s
u
ally
lead
to
b
etter
class
if
icatio
n
th
an
n
o
n
-
lin
ea
r
tr
a
n
s
f
o
r
m
atio
n
s
,
an
d
ar
e
th
er
ef
o
r
e
a
p
r
o
m
is
in
g
n
ew
r
esear
ch
d
ir
ec
tio
n
.
W
e
also
f
o
u
n
d
t
h
at
th
e
s
u
p
er
v
is
ed
tr
a
n
s
f
o
r
m
atio
n
s
ar
e
u
s
u
ally
s
tr
o
n
g
e
r
th
an
t
h
o
s
e
n
o
t
s
u
p
er
v
is
ed
.
W
e
th
in
k
it
wo
u
ld
b
e
wo
r
th
s
ea
r
ch
in
g
f
o
r
o
t
h
er
s
u
p
er
v
is
ed
ap
p
r
o
ac
h
es
wh
ich
co
u
ld
b
e
b
u
ilt
s
im
ilar
ly
to
th
e
KL
DA
o
r
KI
C
A
-
b
ased
ASR
ap
p
licatio
n
m
eth
o
d
o
lo
g
y
.
Su
ch
tr
an
s
f
o
r
m
atio
n
s
s
ig
n
if
ican
tly
im
p
r
o
v
ed
th
e
p
h
o
n
o
lo
g
ical
k
n
o
wled
g
e
ASR
tr
ain
in
g
f
r
am
ewo
r
k
b
y
p
r
o
v
id
in
g
a
co
m
p
r
eh
e
n
s
iv
e
an
d
ac
cu
r
ate
class
if
icatio
n
o
f
s
p
ea
k
in
g
co
n
t
ex
tu
al
f
ea
tu
r
es u
n
iq
u
e
to
r
ea
l
-
t
im
e
s
p
ea
k
er
s
.
RE
F
E
R
E
NC
E
S
[
1
]
S.
S
in
g
h
,
“
S
u
p
p
o
rt
Ve
c
to
r
M
a
c
h
in
e
Ba
se
d
Ap
p
ro
a
c
h
e
s
F
o
r
Re
a
l
Ti
m
e
Au
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
,
p
p
.
8
5
6
1
-
8
5
6
7
,
2
0
1
8
.
[
2
]
B.
S
c
h
ö
lk
o
p
f
,
A.
J
.
S
m
o
la,
a
n
d
K.
R.
M
u
ll
e
r
e
t
a
l.
,
“
Ke
rn
e
l
P
ri
n
c
ip
a
l
Co
m
p
o
n
e
n
t
An
a
l
y
sis
,”
A
d
v
a
n
c
e
s
in
Ke
rn
e
l
M
e
th
o
d
s
:
S
u
p
p
o
rt
Vec
to
r L
e
a
r
n
in
g
,
p
p
.
3
2
7
–
3
5
2
,
1
9
9
9
.
[
3
]
F
.
R.
Ba
c
h
a
n
d
M
.
I.
J
o
rd
a
n
,
“
Ke
rn
e
l
in
d
e
p
e
n
d
e
n
t
c
o
m
p
o
n
e
n
t
a
n
a
ly
sis,”
J
o
u
rn
a
l
o
f
M
a
c
h
i
n
e
L
e
a
r
n
in
g
Res
e
a
rc
h
,
v
o
l.
3
,
n
o
.
2
0
0
2
,
p
p
.
1
-
4
8
,
2
0
0
2
.
[
4
]
G
.
Ba
u
d
a
t
a
n
d
F
.
An
o
u
a
r,
“
G
e
n
e
ra
li
z
e
d
d
isc
rimin
a
n
t
a
n
a
ly
sis
u
si
n
g
a
k
e
rn
e
l
a
p
p
r
o
a
c
h
,
”
Ne
u
ra
l
C
o
mp
u
t
.
,
v
o
l.
1
2
,
n
o
.
1
0
,
p
p
.
2
3
8
5
–
2
4
0
4
,
2
0
0
0
.
[
5
]
A.
Ko
c
so
r
a
n
d
K.
K
o
v
á
c
s
e
t
a
l.
,
“
Ke
rn
e
l
sp
ri
n
g
y
d
isc
rimin
a
n
t
a
n
ly
s
is
a
n
d
it
s
a
p
p
li
c
a
ti
o
n
to
a
p
h
o
n
o
l
o
g
ica
l
a
wa
re
n
e
ss
tea
c
h
in
g
sy
ste
m
,
”
I
n
ter
n
a
t
io
n
a
l
C
o
n
fer
e
n
c
e
o
n
T
e
x
t,
S
p
e
e
c
h
a
n
d
Di
a
lo
g
u
e
,
v
o
l.
2
4
4
8
,
p
p
.
3
2
5
-
3
2
8
,
2
0
0
2
.
[
6
]
F
.
R.
Ba
c
h
a
n
d
M
.
I.
J
o
rd
a
n
,
“
Ke
rn
e
l
in
d
e
p
e
n
d
e
n
t
c
o
m
p
o
n
e
n
t
a
n
a
ly
sis
,
”
J
.
M
a
c
h
in
e
L
e
a
rn
in
g
Res
.
,
v
o
l.
3
,
p
p
.
1
–
4
8
,
2
0
0
2
.
[
7
]
A.
Ko
c
so
r
a
n
d
J.
Csiri
k
,
“
F
a
st
i
n
d
e
p
e
n
d
e
n
t
c
o
m
p
o
n
e
n
t
a
n
a
l
y
sis
i
n
k
e
rn
e
l
fe
a
tu
re
sp
a
c
e
s,”
S
OF
S
E
M
2
0
0
1
:
T
h
e
o
ry
a
n
d
Pra
c
ti
c
e
o
f
In
fo
rm
a
ti
c
s,
2
8
t
h
Co
n
fer
e
n
c
e
o
n
Cu
rr
e
n
t
T
re
n
d
s
i
n
T
h
e
o
ry
a
n
d
Pr
a
c
ti
c
e
o
f
In
f
o
rm
a
ti
c
s
Pi
e
sta
n
y
,
v
o
l.
2
2
3
4
,
p
p
.
2
7
1
–
2
8
1
,
2
0
0
1
.
[
8
]
A.
Ko
c
so
r
a
n
d
K.
K
o
v
á
c
s
e
t
a
l.
,
“
Ke
rn
e
l
sp
ri
n
g
y
d
isc
rimin
a
n
t
a
n
ly
s
is
a
n
d
it
s
a
p
p
li
c
a
ti
o
n
to
a
p
h
o
n
o
l
o
g
ica
l
a
wa
re
n
e
ss
tea
c
h
in
g
sy
ste
m
,
”
I
n
ter
n
a
t
io
n
a
l
C
o
n
fer
e
n
c
e
o
n
T
e
x
t,
S
p
e
e
c
h
a
n
d
Di
a
lo
g
u
e
,
v
o
l.
2
4
4
8
,
p
p
.
3
2
5
–
3
2
8
,
2
0
0
2
.
[
9
]
D.
Lay
,
“
Li
n
e
a
r
Alg
e
b
ra
a
n
d
it
s
a
p
p
li
c
a
ti
o
n
s,
”
4
th
e
d
.
,
P
e
a
rso
n
,
2
0
1
2
[
1
0
]
B.
S
c
h
o
lk
o
p
f
,
A.
J.
S
m
o
la
,
K
.
R.
M
u
ll
e
r,
“
No
n
li
n
e
a
r
Co
m
p
o
n
e
n
t
A
n
a
ly
sis
a
s
a
Ke
rn
e
l
Ei
g
e
n
v
a
lu
e
P
r
o
b
lem
,
”
Ne
u
ra
l
Co
mp
u
t
a
ti
o
n
,
v
o
l.
1
0
,
p
p
.
1
2
9
9
-
1
3
1
9
,
1
9
9
8
.
[
1
1
]
Wein
b
e
rg
e
r,
Kili
a
n
Q.,
P
a
c
k
e
r,
Be
n
jam
in
D.,
a
n
d
S
a
u
l,
Law
re
n
c
e
K.
“
No
n
l
in
e
a
r
d
ime
n
sio
n
a
li
t
y
re
d
u
c
ti
o
n
b
y
se
m
id
e
fin
it
e
p
ro
g
ra
m
m
in
g
a
n
d
k
e
rn
e
l
m
a
tri
x
fa
c
to
riza
ti
o
n
,
”
Pro
c
e
e
d
in
g
s
o
f
th
e
T
e
n
th
I
n
ter
n
a
t
io
n
a
l
W
o
rk
sh
o
p
o
n
Arti
fi
c
ia
l
I
n
telli
g
e
n
c
e
a
n
d
S
t
a
ti
sti
c
s
,
p
p
.
3
8
1
-
3
8
8
,
2
0
0
5
.
[
1
2
]
G
e
r
a
z
o
v
,
B.
;
Iv
a
n
o
v
sk
i,
Z.
“
Ke
rn
e
l
P
o
we
r
F
l
o
w
Orie
n
tatio
n
C
o
e
fficie
n
ts
fo
r
No
ise
-
Ro
b
u
st
S
p
e
e
c
h
Re
c
o
g
n
it
io
n
,
”
IEE
E/
ACM
T
ra
n
s.
A
u
d
i
o
S
p
e
e
c
h
L
a
n
g
.
Pr
o
c
e
ss
,
v
o
l
.
2
3
,
p
p
.
4
0
7
-
4
1
9
,
2
0
1
5
.
[
1
3
]
J.
G
e
ig
e
r
,
B.
S
c
h
u
l
ler
,
G.
Ri
g
o
ll
,
“
Larg
e
-
sc
a
le
a
u
d
io
fe
a
tu
re
e
x
trac
ti
o
n
a
n
d
S
VM
f
o
r
a
c
o
u
stic
sc
e
n
e
c
las
sifica
ti
o
n
,”
Pro
c
e
e
d
in
g
s
o
f
th
e
2
0
1
3
IEE
E
W
o
rk
sh
o
p
o
n
A
p
p
li
c
a
ti
o
n
s
o
f
S
i
g
n
a
l
Pro
c
e
ss
in
g
to
A
u
d
io
a
n
d
Aco
u
st
ics
(W
AS
PA
A),
Ne
w P
a
lt
z
,
NY
,
USA,
2
0
–
2
3
Oc
t
o
b
e
r,
p
p
.
1
-
4
,
2
0
1
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
K
ern
a
l b
a
s
ed
s
p
ea
ke
r
s
p
ec
ific fe
a
tu
r
e
ex
tr
a
ctio
n
a
n
d
its
a
p
p
l
ica
tio
n
s
i
n
…
(
S
a
tya
n
a
n
d
S
in
g
h
)
2497
[
1
4
]
A.
Ra
b
a
o
u
i
,
M.
Da
v
y
,
S.
Ro
ss
i
g
n
a
o
l,
N.
El
l
o
u
z
e
,
“
Us
i
n
g
On
e
-
Clas
s
S
VMs
a
n
d
Wav
e
let
s
fo
r
Au
d
io
S
u
rv
e
il
la
n
c
e
,”
IEE
E
T
ra
n
s.
I
n
f.
Fo
re
n
sic
s S
e
c
u
r
,
v
o
l
.
3
,
7
6
3
-
7
7
5
,
2
0
1
8
.
[
1
5
]
H.
Jia
n
g
,
J.
Ba
i
,
S.
Z
h
a
n
g
,
B.
Xu
,
“
S
VM
-
b
a
se
d
a
u
d
io
sc
e
n
e
c
las
sifica
ti
o
n
,
”
Pro
c
e
e
d
i
n
g
s
o
f
th
e
2
0
0
5
IE
EE
In
ter
n
a
t
io
n
a
l
C
o
n
fer
e
n
c
e
o
n
Na
tu
ra
l
L
a
n
g
u
a
g
e
Pro
c
e
ss
in
g
a
n
d
Kn
o
wled
g
e
En
g
in
e
e
rin
g
,
Wu
h
a
n
,
Ch
in
a
,
p
p
.
1
3
1
-
1
3
6
,
2
0
0
5
.
[
1
6
]
S.
M
i
k
a
,
B.
S
c
h
o
lk
o
p
f,
A.
J.
S
m
o
l
a
,
K.
R.
M
u
ll
e
r,
M
.
S
c
h
o
lz,
a
n
d
G
.
Ra
tsc
h
,
“
Ke
rn
e
l
P
CA
a
n
d
d
e
–
n
o
isin
g
in
fe
a
tu
re
sp
a
c
e
s
,”
i
n
M
.
S
.
Ke
a
rn
s,
S
.
A.
S
o
ll
a
,
a
n
d
D.
A.
C
o
h
n
,
e
d
it
o
rs
,
A
d
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
,
v
o
l
u
m
e
1
1
,
p
a
g
e
s
5
3
6
–
5
4
2
.
M
IT
P
re
ss
,
1
9
9
9
.
[
1
7
]
K.
I.
Kim
,
K.
Ju
n
g
,
H.
J.
Kim
,
“
F
a
c
e
Re
c
o
g
n
it
io
n
Us
in
g
Ke
r
n
e
l
P
rin
c
ip
a
l
C
o
m
p
o
n
e
n
t
A
n
a
ly
sis
,”
IEE
E
S
i
g
n
a
l
Pro
c
e
ss
in
g
L
e
tt
e
rs
,
v
o
l.
9
,
n
o
.
2
,
p
p
.
4
0
-
4
2
,
F
e
b
ru
a
ry
2
0
0
2
)
.
[
1
8
]
L.
B.
Alm
e
id
a
,
“
M
IS
EP
-
li
n
e
a
r
a
n
d
n
o
n
l
in
e
a
r
ICA
b
a
se
d
o
n
m
u
t
u
a
l
in
f
o
rm
a
ti
o
n
,”
J
o
u
rn
a
l
o
f
M
a
c
h
in
e
L
e
a
rn
in
g
Res
e
a
rc
h
,
v
o
l.
4
,
n
o
.
2
0
0
2
,
p
p
.
1
2
9
7
-
1
3
1
8
,
2
0
0
3
.
[
1
9
]
K.
Zh
a
n
g
,
L.
Ch
a
n
,
“
No
n
li
n
e
a
r
in
d
e
p
e
n
d
e
n
t
c
o
m
p
o
n
e
n
t
a
n
a
ly
sis
with
m
in
imu
m
n
o
n
li
n
e
a
r
d
isto
rti
o
n
,”
ICM
L
2
0
0
7
,
Co
rv
a
ll
is
,
OR,
US,
p
p
.
1
1
2
7
-
1
1
3
4
,
2
0
0
7
.
[
2
0
]
M.
S.
Ba
rtl
e
tt
,
J.
R.
M
o
v
e
ll
a
n
,
T.
J.
S
e
jn
o
ws
k
i
,
“
F
a
c
e
Re
c
o
g
n
it
i
o
n
b
y
In
d
e
p
e
n
d
e
n
t
Co
m
p
o
n
e
n
t
An
a
ly
sis
,”
IEE
E
T
ra
n
sa
c
ti
o
n
s
o
n
Ne
u
ra
l
Ne
two
rk
s
,
v
o
l.
1
3
,
no
.
6
,
p
p
.
1
4
5
0
-
1
4
6
4
,
2
0
0
2
.
[
2
1
]
F.
R.
Ba
c
h
,
M.
I.
Jo
r
d
a
n
,
“
Ke
rn
e
l
In
d
e
p
e
n
d
e
n
t
C
o
m
p
o
n
e
n
t
An
a
ly
sis
,”
J
.
M
a
c
h
in
e
L
e
a
rn
in
g
Res
,
v
o
l.
3,
n
o
.
2
0
0
2
,
pp.
1
-
4
8
,
2
0
0
2
.
[
2
2
]
O.
S
io
h
a
n
,
“
On
t
h
e
ro
b
u
stn
e
ss
o
f
li
n
e
a
r
d
isc
rimin
a
n
t
a
n
a
l
y
sis
a
s
a
p
re
p
ro
c
e
ss
in
g
ste
p
f
o
r
n
o
isy
sp
e
e
c
h
re
c
o
g
n
it
io
n
,
”
Pro
c
.
ICAS
S
P
,
De
tro
it
,
M
I,
p
p
.
1
2
5
-
1
2
8
,
1
9
9
5
.
[
2
3
]
S
.
S
in
g
h
,
EG
Ra
jan
,
”
Ap
p
li
c
a
ti
o
n
o
f
d
iffere
n
t
fil
ters
i
n
M
e
l
fre
q
u
e
n
c
y
c
e
p
stra
l
c
o
e
fficie
n
ts
fe
a
tu
re
e
x
trac
ti
o
n
a
n
d
fu
z
z
y
v
e
c
to
r
q
u
a
n
t
iza
ti
o
n
a
p
p
ro
a
c
h
in
sp
e
a
k
e
r
re
c
o
g
n
it
io
n
,
”
In
t
e
rn
a
ti
o
n
a
l
J
o
u
rn
a
l
o
f
En
g
i
n
e
e
rin
g
Res
e
a
rc
h
&
T
e
c
h
n
o
l
o
g
y
,
v
o
l.
2
,
n
o
.
6
,
p
p
.
4
1
9
-
4
2
5
,
2
0
1
3
.
[
2
4
]
G
.
S
.
M
o
rriso
n
,
a
n
d
F
.
Ke
ll
y
,
“
A
sta
ti
stica
l
p
ro
c
e
d
u
re
to
a
d
ju
st
fo
r
ti
m
e
-
in
terv
a
l
m
ism
a
tch
in
fo
re
n
sic
v
o
ice
c
o
m
p
a
riso
n
”
S
p
e
e
c
h
Co
mm
u
n
ica
t
io
n
,
v
o
l.
1
1
2
,
p
p
.
1
5
-
2
1
,
2
0
1
9
.
[
2
5
]
S.
S
i
n
g
h
a
n
d
Aje
e
t
S
in
g
h
“
Ac
c
u
ra
c
y
Co
m
p
a
riso
n
u
si
n
g
Diffe
re
n
t
M
o
d
e
li
n
g
Tec
h
n
i
q
u
e
s u
n
d
e
r
Li
m
i
ted
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
rn
a
l
o
f
S
c
ien
c
e
Fro
n
ti
e
r
Res
e
a
rc
h
:
F
M
a
t
h
e
ma
ti
c
s
a
n
d
De
c
isio
n
S
c
ien
c
e
s
,
v
o
l
.
16
,
n
o
.
2
,
p
p
.
1
-
17
,
2
0
1
6
.
[
2
6
]
P
.
Bo
e
rsm
a
a
n
d
D.
Wee
n
i
n
k
,
“
P
ra
a
t:
d
o
in
g
p
h
o
n
e
ti
c
s
b
y
c
o
m
p
u
ter”
,
v
e
rsio
n
6
.
0
.
3
7
,
2
0
2
0
.
A
v
a
il
a
b
le:
h
tt
p
:
//
ww
w.p
ra
a
t.
o
r
g
/,
2
0
2
0
[
2
7
]
S.
S
in
g
h
.
“
Th
e
Ro
le
o
f
S
p
e
e
c
h
Tec
h
n
o
l
o
g
y
in
Bi
o
m
e
tri
c
s,
F
o
re
n
sic
s
a
n
d
M
a
n
-
M
a
c
h
i
n
e
In
terfa
c
e
”
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
C
o
mp
u
t
e
r E
n
g
i
n
e
e
rin
g
,
v
o
l.
9
,
n
o
.
1
,
p
p
.
281
-
2
8
8
,
2
0
1
9
.
[
2
8
]
Ch
e
n
-
Ch
o
u
L
o
,
S
z
u
-
Wei
F
u
,
W
e
n
-
Ch
in
H
u
a
n
g
,
Xin
Wan
g
,
J
u
n
i
c
h
i
Ya
m
a
g
ish
i,
Yu
Tsa
o
,
a
n
d
H
sin
-
M
in
Wan
g
,
“
M
OSNe
t:
De
e
p
Lea
rn
in
g
b
a
se
d
Ob
jec
ti
v
e
As
se
ss
m
e
n
t
fo
r
V
o
ice
Co
n
v
e
rsi
o
n
,
”
In
ter
sp
e
e
c
h
IS
CA
,
p
p
.
1
5
4
2
–
1
5
4
5
,
2
0
1
9
.
[
2
9
]
D.
S
n
y
d
e
r,
D.
G
a
rc
ia
-
Ro
m
e
ro
,
G
.
S
e
ll
,
D.
P
o
v
e
y
,
a
n
d
S
.
Kh
u
d
a
n
p
u
r,
“
X
-
v
e
c
to
rs:
Ro
b
u
st
DN
N
e
m
b
e
d
d
in
g
s
fo
r
sp
e
a
k
e
r
re
c
o
g
n
it
io
n
,
”
ICAS
S
P
2
0
1
8
,
2
0
1
8
.
[
3
0
]
G
a
li
n
a
Lav
re
n
ty
e
v
a
,
S
e
rg
e
y
N
o
v
o
se
lo
v
,
An
d
z
h
u
k
a
e
v
Tse
re
n
,
M
a
rin
a
Vo
lk
o
v
a
,
Arte
m
G
o
rlan
o
v
,
a
n
d
Ale
x
a
n
d
r
Ko
z
lo
v
,
“
S
TC
An
t
isp
o
o
fi
n
g
S
y
ste
m
s
f
o
r
th
e
ASVsp
o
o
f2
0
1
9
Ch
a
ll
e
n
g
e
,
”
Pr
o
c
.
I
n
t
e
rs
p
e
e
c
h
2
0
1
9
,
p
p
.
1
0
3
3
-
1
0
3
7
,
2
0
1
9
.
[
3
1
]
S.
S
i
n
g
h
a
n
d
M
a
n
s
o
u
r
H
.
As
sa
f
,
“
A
P
e
rfe
c
t
Ba
lan
c
e
o
f
S
p
a
rsit
y
a
n
d
Ac
o
u
stic
h
o
le
in
S
p
e
e
c
h
S
ig
n
a
l
a
n
d
Its
Ap
p
li
c
a
ti
o
n
in
S
p
e
a
k
e
r
Re
c
o
g
n
it
io
n
S
y
ste
m
,
”
M
id
d
le
-
Ea
st
J
o
u
rn
a
l
o
f
S
c
ien
ti
fi
c
Res
e
a
rc
h
,
v
o
l.
2
4
,
n
o
.
1
1
,
p
p
.
3
5
2
7
-
3
5
4
1
,
2
0
1
6
.
[
3
2
]
A.
Ale
x
a
n
d
e
r,
O.
F
o
rt
h
,
A.
A.
At
re
y
a
,
F.
Ke
ll
y
,
“
VO
CALIS
E:
A
F
o
re
n
sic
Au
t
o
m
a
ti
c
S
p
e
a
k
e
r
Re
c
o
g
n
it
io
n
S
y
ste
m
S
u
p
p
o
rt
in
g
S
p
e
c
tral,
P
h
o
n
e
ti
c
,
a
n
d
Us
e
r
-
P
ro
v
id
e
d
F
e
a
tu
re
s
,”
Re
se
a
rc
h
a
n
d
De
v
e
l
o
p
m
e
n
t
Ox
fo
r
d
Wa
v
e
Re
se
a
rc
h
Lt
d
,
Un
it
e
d
Ki
n
g
d
o
m
,
2
0
1
6
.
Evaluation Warning : The document was created with Spire.PDF for Python.