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
4
4
3
~2
4
5
0
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
4
4
3
-
2450
2443
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
H
ig
h
lev
el sp
ea
ke
r specifi
c f
ea
tures
a
s a
n
eff
i
ciency
e
nha
ncing
pa
ra
m
et
ers
in
s
pe
a
k
er
rec
o
g
nition s
y
ste
m
Sa
t
y
a
na
nd
Sin
g
h
S
c
h
o
o
l
o
f
El
e
c
tri
c
a
l
a
n
d
El
e
c
tro
n
i
c
s
En
g
in
e
e
rin
g
,
F
ij
i
Na
ti
o
n
a
l
Un
i
v
e
rsit
y
,
F
ij
i
Isla
n
d
Art
icle
I
nfo
AB
ST
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Dec
5
,
2
0
1
8
R
ev
i
s
ed
J
an
2
1
,
2
0
1
9
A
cc
ep
ted
Mar
11
,
2
0
1
9
In
th
is
p
a
p
e
r,
I
p
re
se
n
t
h
ig
h
-
lev
e
l
sp
e
a
k
e
r
sp
e
c
i
f
ic
fe
a
tu
re
e
x
trac
ti
o
n
c
o
n
sid
e
ri
n
g
in
t
o
n
a
ti
o
n
,
li
n
g
u
ist
ics
rh
y
th
m
,
li
n
g
u
isti
c
s
stre
ss
,
p
r
o
so
d
ic
f
e
a
tu
re
s
d
irec
tl
y
f
ro
m
sp
e
e
c
h
sig
n
a
ls.
I
a
ss
u
m
e
th
a
t
th
e
rh
y
th
m
is
re
late
d
to
lan
g
u
a
g
e
u
n
it
s
su
c
h
a
s
s
y
ll
a
b
les
a
n
d
a
p
p
e
a
rs
a
s
c
h
a
n
g
e
s
in
m
e
a
su
ra
b
le
p
a
ra
m
e
ters
su
c
h
a
s
f
u
n
d
a
m
e
n
t
a
l
f
re
q
u
e
n
c
y
F
o
,
d
u
ra
ti
o
n
,
a
n
d
e
n
e
rg
y
.
In
th
is
w
o
rk
,
th
e
sy
ll
a
b
le
t
y
p
e
f
e
a
tu
re
s
a
re
se
lec
ted
a
s
th
e
b
a
sic
u
n
it
f
o
r
e
x
p
re
ss
in
g
th
e
p
ro
so
d
ic
f
e
a
tu
re
s.
T
h
e
a
p
p
ro
x
im
a
t
e
s
e
g
m
e
n
tatio
n
o
f
c
o
n
ti
n
u
o
u
s
sp
e
e
c
h
t
o
sy
ll
a
b
le
u
n
i
ts
is
a
c
h
iev
e
d
b
y
a
u
to
m
a
ti
c
a
ll
y
lo
c
a
ti
n
g
th
e
v
o
w
e
l
sta
rti
n
g
p
o
in
t.
T
h
e
k
n
o
w
led
g
e
o
f
h
ig
h
-
lev
e
l
sp
e
a
k
e
r’
s
sp
e
c
if
i
c
sp
e
a
k
e
rs
is
u
se
d
a
s
a
re
fe
re
n
c
e
f
o
r
e
x
tra
c
ti
n
g
th
e
p
ro
s
o
d
ic
f
e
a
tu
re
s
o
f
th
e
sp
e
e
c
h
sig
n
a
l.
Hig
h
-
lev
e
l
sp
e
a
k
e
r
-
sp
e
c
i
f
ic
fe
a
tu
re
s
e
x
trac
ted
u
sin
g
th
is
m
e
th
o
d
m
a
y
b
e
u
se
f
u
l
in
a
p
p
li
c
a
ti
o
n
s
su
c
h
a
s
sp
e
a
k
e
r
re
c
o
g
n
it
io
n
w
h
e
re
e
x
p
li
c
it
p
h
o
n
e
m
e
/s
y
ll
a
b
le
b
o
u
n
d
a
ries
a
re
n
o
t
re
a
d
il
y
a
v
a
il
a
b
le.
T
h
e
e
ff
ici
e
n
c
y
o
f
th
e
p
a
rti
c
u
lar
c
h
a
ra
c
teristics
o
f
th
e
sp
e
c
i
f
ic
f
e
a
tu
re
s
u
se
d
f
o
r
a
u
to
m
a
ti
c
sp
e
a
k
e
r
re
c
o
g
n
it
io
n
w
a
s
e
v
a
lu
a
ted
o
n
T
IM
IT
a
n
d
HT
IM
IT
c
o
rp
o
ra
in
it
ially
sa
m
p
led
in
th
e
T
IM
IT
a
t
1
6
k
Hz
to
8
k
Hz
.
In
su
m
m
a
r
y
,
th
e
e
x
p
e
ri
m
e
n
t,
th
e
b
a
sic
d
isc
ri
m
in
a
ti
n
g
s
y
ste
m
,
a
n
d
th
e
HMM
s
y
ste
m
a
re
f
o
r
m
e
d
o
n
T
IM
IT
c
o
rp
u
s
w
it
h
a
se
t
o
f
4
8
p
h
o
n
e
m
e
s.
P
ro
p
o
se
d
A
S
R
sy
ste
m
sh
o
w
s
1
.
9
9
%
,
2
.
1
0
%
,
2
.
1
6
%
a
n
d
2
.
1
9
%
o
f
e
ff
icie
n
c
y
i
m
p
r
o
v
e
m
e
n
ts
c
o
m
p
a
re
to
trad
it
io
n
a
l
A
S
R
sy
ste
m
f
o
r
<
1
0
m
s,
<
2
0
m
s,
<
3
0
m
s
a
n
d
<
4
0
m
s
o
f
1
6
KH
z
T
IM
IT
u
tt
e
ra
n
c
e
s.
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
)
C
o
n
f
id
en
ce
m
ea
s
u
r
e
(
C
M)
Dee
p
n
eu
r
al
n
e
t
w
o
r
k
s
(
DNN)
Gau
s
s
ia
n
m
i
x
er
m
o
d
el
(
GM
M
)
Mel
-
f
r
eq
u
en
c
y
ce
p
s
tr
al
co
ef
f
icie
n
t
s
(
MFC
C
)
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
,
Sch
o
o
l o
f
E
lectr
ical
an
d
E
lectr
o
n
ics E
n
g
i
n
ee
r
in
g
,
Fij
i N
atio
n
al
Un
i
v
er
s
i
t
y
,
Fij
i I
s
lan
d
.
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
T
h
e
lan
g
u
ag
e
is
m
ai
n
l
y
f
o
r
p
ass
in
g
i
n
f
o
r
m
atio
n
f
r
o
m
o
n
e
p
er
s
o
n
to
o
th
er
p
er
s
o
n
i
n
d
a
y
to
d
a
y
lif
e
[
1
]
.
I
t
is
tr
an
s
m
itted
th
r
o
u
g
h
a
s
er
ies
o
f
leg
a
l
s
o
u
n
d
u
n
its
.
T
h
is
s
eq
u
e
n
ce
m
u
s
t
r
esp
ec
t
th
e
co
n
s
tr
ai
n
ts
i
m
p
o
s
ed
b
y
t
h
e
la
n
g
u
a
g
e.
T
h
er
ef
o
r
e,
s
p
ee
ch
a
n
d
la
n
g
u
a
g
e
ar
e
co
m
p
li
m
e
n
ts
o
f
ea
ch
o
th
er
an
d
it
ca
n
n
o
t
be
s
ep
ar
ated
.
B
ec
au
s
e
ea
ch
s
p
ea
k
e
r
h
as
u
n
iq
u
e
p
h
y
s
i
o
lo
g
ical
ch
ar
ac
ter
is
tic
s
o
f
s
p
ee
ch
an
d
s
p
ee
ch
g
en
er
atio
n
s
t
y
le
an
d
s
p
ea
k
er
-
s
p
ec
if
ic
f
ea
t
u
r
es
ar
e
also
in
te
g
r
ated
in
to
th
e
s
p
ee
ch
s
i
g
n
al.
T
h
er
ef
o
r
e,
th
e
s
p
ee
ch
s
ig
n
al
co
n
tain
s
n
o
t
o
n
l
y
t
h
e
ex
p
ec
ted
m
es
s
ag
e
b
u
t
also
t
h
e
lan
g
u
a
g
e
an
d
s
p
ea
k
er
s
p
ec
i
f
ic
c
h
ar
ac
ter
is
tics
.
I
n
ad
d
itio
n
,
th
e
e
m
o
tio
n
al
s
tat
e
o
f
th
e
s
p
ea
k
er
is
al
s
o
tr
an
s
m
itted
t
h
r
o
u
g
h
w
o
r
d
s
[
2
,
3]
.
T
h
e
s
p
ee
ch
m
es
s
ag
e
p
ar
t
is
m
ai
n
l
y
e
x
p
r
ess
ed
as
a
s
er
ies
o
f
le
g
al
s
o
u
n
d
u
n
its
,
ea
ch
co
r
r
esp
o
n
d
in
g
to
t
h
e
m
an
n
er
an
d
lo
ca
tio
n
o
f
s
p
ee
ch
p
r
o
d
u
ctio
n
b
y
a
p
ar
tic
u
lar
s
o
u
n
d
u
n
it.
T
h
e
lan
g
u
ag
e
,
e
m
o
tio
n
s
a
n
d
s
p
ea
k
er
p
ar
ts
o
f
t
h
e
i
n
f
o
r
m
atio
n
co
n
tain
ed
i
n
t
h
e
s
p
ee
ch
s
i
g
n
a
l
ar
e
d
er
iv
ed
u
s
i
n
g
s
ev
er
al
le
v
els
o
f
f
u
n
ctio
n
a
lit
y
.
E
x
i
s
ti
n
g
s
p
ea
k
er
,
lan
g
u
a
g
e,
e
m
o
tio
n
,
a
n
d
s
p
ee
ch
r
ec
o
g
n
i
tio
n
s
y
s
te
m
s
r
el
y
o
n
f
ea
t
u
r
es
d
er
iv
ed
f
r
o
m
t
h
e
s
h
o
r
t
-
ter
m
s
p
ec
tr
al
an
al
y
s
is
.
Ho
w
e
v
er
,
th
e
s
p
ec
tr
al
ch
ar
ac
t
er
is
tics
ar
e
af
f
ec
ted
b
y
ch
a
n
n
el
an
d
n
o
is
e
ch
ar
ac
ter
i
s
tics
.
T
h
is
h
as
p
r
o
m
p
ted
r
esear
ch
er
s
to
e
x
p
lo
r
e
th
e
u
s
e
o
f
ad
d
itio
n
al
f
ea
tu
r
es
t
h
at
m
a
y
p
r
o
v
id
e
ad
d
itio
n
al
e
v
id
e
n
ce
of
a
s
p
ec
tr
u
m
-
b
ased
s
y
s
te
m
.
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
4
4
3
-
2450
2444
Sp
ee
ch
p
r
o
ce
s
s
in
g
r
esear
ch
a
i
m
s
to
i
m
p
le
m
en
t
m
ac
h
i
n
es
c
ap
ab
le
o
f
p
er
f
o
r
m
i
n
g
au
to
m
at
ic
s
p
ee
ch
r
ec
o
g
n
itio
n
,
s
p
ee
ch
s
y
n
t
h
esi
s
,
s
p
ea
k
er
r
ec
o
g
n
itio
n
,
a
n
d
m
a
n
y
o
t
h
er
s
p
ee
ch
p
r
o
ce
s
s
in
g
ta
s
k
s
s
u
ch
a
s
s
p
ee
c
h
r
ec
o
g
n
itio
n
b
y
m
ac
h
in
e
lik
e
a
h
u
m
a
n
[
4
,
5]
.
T
h
e
r
esear
ch
er
s
s
u
cc
ee
d
ed
in
d
e
v
elo
p
in
g
s
p
ee
ch
s
y
s
te
m
s
o
p
er
atin
g
in
a
r
estricte
d
e
n
v
ir
o
n
m
e
n
t.
Ma
n
y
o
f
t
h
ese
s
y
s
te
m
s
r
el
y
s
o
lel
y
o
n
ac
o
u
s
t
ic
m
o
d
els
f
o
r
m
ed
u
s
i
n
g
s
p
ec
tr
al
ch
ar
ac
ter
is
tic
s
.
T
h
ese
ac
o
u
s
tic
m
o
d
els
lac
k
m
u
c
h
h
ig
h
er
-
lev
e
l
i
n
f
o
r
m
atio
n
t
h
at
h
u
m
a
n
s
u
s
e
f
o
r
t
h
e
s
a
m
e
ta
s
k
.
T
h
e
h
i
g
h
est le
v
els
o
f
in
f
o
r
m
atio
n
in
c
lu
d
e
p
r
o
s
o
d
y
,
co
n
tex
t,
an
d
v
o
ca
b
u
lar
y
k
n
o
w
led
g
e.
I
t
is
u
n
d
er
s
to
o
d
th
at
t
h
e
i
n
tr
o
d
u
ctio
n
o
f
t
h
e
k
n
o
w
l
ed
g
e
o
f
p
r
o
s
o
d
y
i
n
to
au
to
m
at
ic
s
p
ea
k
er
r
ec
o
g
n
itio
n
(
A
S
R
)
s
y
s
te
m
o
f
th
e
v
o
ca
l
s
y
s
te
m
s
w
il
l
m
a
k
e
t
h
e
m
m
o
r
e
in
te
l
li
g
en
t
an
d
s
i
m
ilar
to
h
u
m
an
s
[
6
]
.
Var
io
u
s
r
esear
ch
er
s
i
n
t
h
e
p
ast
h
a
v
e
estab
li
s
h
ed
t
h
e
i
m
p
o
r
tan
ce
o
f
p
r
o
s
o
d
ic
f
ea
tu
r
es
f
o
r
s
p
e
ec
h
p
r
o
ce
s
s
in
g
ap
p
licatio
n
s
[
7
]
.
Un
f
o
r
tu
n
atel
y
,
i
n
co
r
p
o
r
atio
n
o
f
p
r
o
s
o
d
y
i
n
to
t
h
e
s
p
ee
ch
s
y
s
te
m
s
h
a
s
to
ad
d
r
ess
s
e
v
er
al
is
s
u
es.
O
n
e
m
aj
o
r
is
s
u
e
is
t
h
e
au
to
m
atic
e
x
tr
ac
tio
n
an
d
r
ep
r
esen
tatio
n
o
f
p
r
o
s
o
d
y
a
n
d
its
ap
p
licatio
n
in
s
p
ea
k
er
r
ec
o
g
n
itio
n
to
en
h
a
n
c
e
th
e
ef
f
icien
c
y
o
f
A
SR
.
O
u
r
f
u
n
d
a
m
e
n
tal
u
n
d
er
s
tan
d
i
n
g
o
f
th
e
p
r
o
ce
s
s
es
in
m
o
s
t
o
f
t
h
e
s
p
ee
ch
p
er
ce
p
tio
n
m
o
d
u
les
i
n
Fi
g
u
r
e
1
is
r
u
d
i
m
e
n
tar
y
at
b
est,
b
u
t
i
t
is
g
e
n
er
all
y
a
g
r
ee
d
th
at
s
o
m
e
p
h
y
s
ical
co
r
r
elate
o
f
ea
c
h
o
f
t
h
e
s
tep
s
i
n
t
h
e
s
p
ee
ch
p
er
ce
p
ti
o
n
m
o
d
el
o
cc
u
r
w
it
h
i
n
t
h
e
h
u
m
an
b
r
ain
,
a
n
d
t
h
u
s
th
e
en
t
ir
e
m
o
d
el
is
u
s
e
f
u
l
f
o
r
th
in
k
i
n
g
ab
o
u
t th
e
p
r
o
ce
s
s
es t
h
at
o
cc
u
r.
Fig
u
r
e
1
.
T
h
e
Sp
ee
ch
Gen
er
ati
o
n
C
h
a
in
o
f
a
No
r
m
al
A
u
d
ito
r
y
S
y
s
te
m
2.
P
RO
SO
DY
H
I
G
H
L
E
V
E
L
SPEAK
E
R
SPEC
I
F
I
C
F
E
A
T
UR
E
S IN
SPEAK
E
R
RE
C
O
G
NI
T
I
O
N
Sh
o
r
t
-
ter
m
ce
p
s
tr
al
f
ea
tu
r
e
s
ar
e
o
f
te
n
r
ef
er
r
ed
to
a
s
lo
w
le
v
el
s
r
ef
lec
ts
th
e
s
p
ea
k
er
's
v
o
ice
r
ath
er
t
h
a
n
ca
p
tu
r
in
g
h
ig
h
le
v
el
s
p
ea
k
er
s
p
ec
if
ic
f
ea
t
u
r
es
,
r
h
y
t
h
m
,
a
n
d
v
o
ca
b
u
lar
y
i
n
f
o
r
m
atio
n
.
Un
f
o
r
tu
n
atel
y
,
s
o
m
e
p
r
o
s
o
d
ic
f
ea
tu
r
es
ar
e
v
er
y
d
if
f
ic
u
lt
to
ca
lc
u
late,
w
h
ile
o
t
h
e
r
s
ar
e
d
if
f
ic
u
lt
to
d
ed
u
ce
s
o
l
el
y
f
r
o
m
ac
o
u
s
t
ics
(
eg
,
th
e
r
o
u
n
d
n
es
s
o
f
lip
s
)
.
As
a
r
es
u
lt,
m
o
r
e
a
n
d
m
o
r
e
f
ea
tu
r
es
ar
e
r
ec
ei
v
i
n
g
i
n
cr
ea
s
i
n
g
atten
t
io
n
o
v
er
t
h
e
p
ast d
ec
ad
e.
Sp
ee
ch
is
tr
an
s
m
itted
t
h
r
o
u
g
h
a
s
er
ies
o
f
le
g
al
s
o
u
n
d
u
n
its
i
n
t
h
e
la
n
g
u
a
g
e.
W
ith
th
e
o
r
d
e
r
o
f
s
o
u
n
d
u
n
i
ts
,
s
o
m
e
b
u
ilt
-
i
n
f
ea
t
u
r
es
g
i
v
e
a
n
atu
r
al
v
o
ice.
T
h
e
p
itch
c
h
an
g
e
p
r
o
v
id
es
i
d
en
ti
f
iab
le
m
elo
d
y
attr
ib
u
tes
f
o
r
s
p
ee
ch
.
T
h
is
co
n
tr
o
lled
m
o
d
u
latio
n
o
f
s
o
u
n
d
is
ca
lled
in
to
n
at
io
n
.
T
h
e
u
n
i
t
o
f
s
o
u
n
d
is
s
h
o
r
ten
ed
o
r
len
g
th
e
n
ed
ac
co
r
d
in
g
to
s
o
m
e
b
asic
m
o
d
es to
g
i
v
e
a
ce
r
tain
r
h
y
t
h
m
to
th
e
v
o
ice.
T
h
er
e
ar
e
f
e
w
s
y
llab
le
s
o
r
w
o
r
d
s
m
a
y
b
e
m
o
r
e
i
m
p
o
r
tan
t
t
h
an
o
t
h
er
s
,
ca
u
s
i
n
g
la
n
g
u
a
g
e
p
r
ess
u
r
e.
T
h
e
in
to
n
atio
n
,
r
h
y
t
h
m
,
an
d
p
r
ess
u
r
e
o
f
s
p
ee
ch
in
cr
ea
s
e
t
h
e
i
n
telli
g
ib
ilit
y
o
f
s
p
ee
c
h
i
n
f
o
r
m
atio
n
,
allo
w
i
n
g
lis
te
n
er
s
to
ea
s
il
y
d
iv
id
e
co
n
t
in
u
o
u
s
s
p
ee
ch
i
n
to
s
en
ten
ce
s
an
d
w
o
r
d
s
[
8
]
.
I
t
also
co
n
v
e
y
s
m
o
r
e
v
o
ca
b
u
lar
y
an
d
n
o
n
-
v
er
b
al
i
n
f
o
r
m
atio
n
s
u
ch
as
v
o
ca
l
to
n
e
s
,
lo
u
d
to
n
e
s
,
ac
ce
n
ts
,
a
n
d
e
m
o
tio
n
s
.
T
h
e
ch
ar
ac
ter
is
tics
th
a
t
m
ak
e
u
s
p
er
ce
iv
e
t
h
ese
e
f
f
ec
t
s
ar
e
co
llectiv
el
y
ca
lled
p
r
o
s
o
d
y
.
H
u
m
an
P
r
o
s
o
d
y
is
u
s
ed
to
o
b
tain
in
f
o
r
m
atio
n
s
u
c
h
a
s
e
m
o
tio
n
s
,
w
o
r
d
/s
en
t
e
n
ce
b
o
u
n
d
ar
ies,
s
p
ea
k
er
c
h
ar
ac
ter
is
tics
an
d
lan
g
u
ag
e
ch
ar
ac
ter
is
tics
,
w
h
ic
h
ar
e
u
s
ed
f
o
r
s
p
ea
k
er
id
en
ti
f
icatio
n
.
E
ac
h
p
r
o
m
p
t
is
a
co
m
p
le
x
p
er
ce
p
tu
al
en
t
it
y
m
ai
n
l
y
r
ep
r
esen
ted
b
y
th
r
e
e
ac
o
u
s
tic
p
ar
a
m
eter
s
: to
n
e,
en
e
r
g
y
,
an
d
d
u
r
atio
n
.
2
.
1
.
I
nto
na
t
i
o
n
a
s
s
pea
k
er
s
pecif
ic
f
ea
t
ures
in ASR
s
y
s
t
e
m
P
itch
is
t
h
e
p
er
ce
iv
ed
p
r
o
p
er
ty
o
f
s
o
u
n
d
a
n
d
ca
n
b
e
d
escr
i
b
ed
as
a
p
er
ce
p
t
io
n
o
f
s
o
u
n
d
r
elativ
e
to
“
p
itc
h
”
[
9
]
.
T
h
e
p
h
y
s
ica
l
co
r
r
elatio
n
o
f
p
itch
is
th
e
f
u
n
d
a
m
e
n
tal
f
r
eq
u
e
n
c
y
(
F
0
)
d
eter
m
i
n
ed
b
y
t
h
e
v
ib
r
atio
n
al
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
E
lec
&
C
o
m
p
E
n
g
IS
SN: 2
0
8
8
-
8708
Hig
h
leve
l sp
ea
ke
r
s
p
ec
ific fe
a
tu
r
es a
s
a
n
efficien
cy
en
h
a
n
cin
g
p
a
r
a
mete
r
s
in
s
p
ea
ke
r
…
(
S
a
tya
n
a
n
d
S
in
g
h
)
2445
r
ate
o
f
th
e
v
o
ca
l
co
r
d
s
.
T
h
e
s
et
o
f
p
itch
ch
an
g
e
s
d
u
r
in
g
s
p
e
ec
h
is
d
ef
i
n
ed
as
in
to
n
a
tio
n
[
10
]
.
T
h
e
F
0
r
an
g
e
d
ep
en
d
s
o
n
th
e
len
g
t
h
an
d
m
a
s
s
o
f
t
h
e
v
o
ca
l
co
r
d
o
f
a
s
p
ea
k
er
.
Fo
r
m
en
,
it
i
s
u
s
u
a
ll
y
b
et
wee
n
8
0
an
d
2
0
0
Hz,
an
d
f
o
r
w
o
m
e
n
,
b
et
w
ee
n
1
8
0
an
d
4
0
0
Hz
(
a
p
p
r
o
x
im
a
tel
y
)
f
o
r
co
n
v
er
s
atio
n
a
l
s
p
ee
ch
.
I
n
th
is
r
an
g
e,
ea
ch
s
p
ea
k
er
ca
n
p
r
o
d
u
ce
a
n
i
n
cr
ea
s
e
a
n
d
a
d
ec
r
ea
s
e
o
f
F
0
.
T
h
e
d
ir
ec
tio
n
i
n
w
h
ic
h
F
0
ch
a
n
g
e
s
(
u
p
o
r
d
o
w
n
)
is
d
eter
m
i
n
ed
b
y
t
h
e
s
p
ee
ch
p
atter
n
th
at
m
ak
e
s
u
p
th
e
w
o
r
d
.
Fig
u
r
e
2
s
h
o
w
s
t
h
e
ev
o
l
u
t
io
n
0
o
f
th
e
m
a
n
s
tate
m
en
t
i
n
“
Co
lleg
e
o
f
E
n
g
ineering
Science
&
T
ec
hn
o
lo
g
y
”
.
I
t
ca
n
b
e
o
b
s
er
v
ed
th
at
th
e
v
al
u
e
F
0
is
d
is
co
n
ti
n
u
o
u
s
b
ec
au
s
e
o
f
t
h
e
p
er
io
d
ic
ch
ar
ac
ter
o
f
th
e
s
p
ee
ch
d
u
r
i
n
g
th
e
v
o
ca
l
r
eg
io
n
(
v
o
w
el,
n
a
s
al,
s
e
m
i
to
n
e,
co
n
s
o
n
a
n
t,
etc.
)
an
d
th
e
ap
er
io
d
ic
n
at
u
r
e
d
u
r
i
n
g
th
e
s
ile
n
t
r
eg
io
n
.
T
h
e
s
m
all
d
is
t
u
r
b
an
ce
s
i
n
t
h
e
F
0
p
r
o
ce
s
s
ar
e
m
ai
n
l
y
d
u
e
to
th
e
i
n
v
o
l
u
n
tar
y
asp
ec
t
s
o
f
th
e
la
n
g
u
ag
e.
Fig
u
r
e
2
.
Var
iatio
n
o
f
f
o
r
th
e
Utter
an
ce
C
o
lle
g
e
o
f
E
n
g
i
n
ee
r
in
g
Scien
ce
&
T
ec
h
n
o
lo
g
y
w
o
r
d
s
s
tr
ess
ed
2
.
2
.
L
ing
uis
t
ics St
re
s
s
a
s
Sp
ea
k
er
S
pecif
ic
F
ea
t
ures
in A
SR Sy
s
t
e
m
I
n
lin
g
u
is
t
ics,
s
tr
es
s
is
t
h
e
ab
i
lit
y
to
g
iv
e
r
elati
v
e
i
m
p
o
r
ta
n
c
e
to
ce
r
tain
s
y
l
lab
les
o
r
s
en
te
n
ce
s
o
f
a
w
o
r
d
o
r
to
ce
r
tain
w
o
r
d
s
o
f
a
s
en
te
n
ce
.
T
h
er
e
ar
e
p
r
ess
u
r
es
in
m
a
n
y
lan
g
u
ag
e
s
o
f
th
e
w
o
r
ld
.
Stre
s
s
is
a
n
attr
ib
u
te
o
f
th
e
s
tr
u
ct
u
r
al
lan
g
u
ag
e
o
f
a
w
o
r
d
th
at
in
d
icate
s
w
h
ic
h
s
y
llab
le
in
a
w
o
r
d
is
s
tr
o
n
g
er
in
o
n
e
s
e
n
s
e
th
an
a
n
y
o
th
er
s
y
llab
le.
On
e
o
f
th
e
i
m
p
o
r
ta
n
t
th
e
m
es
o
f
t
h
e
r
esear
ch
o
n
s
p
ee
ch
h
as
al
w
a
y
s
b
ee
n
th
e
e
m
p
h
as
i
s
on
th
e
ac
o
u
s
tic
an
d
p
er
ce
p
tu
al
ch
ar
ac
ter
is
tic
s
o
f
attr
ib
u
t
es:
s
y
llab
le
s
ar
e
d
is
tin
g
u
i
s
h
e
d
f
r
o
m
u
n
s
tr
ess
ed
s
y
llab
le
s
t
h
at
s
u
r
r
o
u
n
d
t
h
e
m
,
o
r
in
a
m
o
r
e
co
n
tr
o
llab
le
wa
y
,
th
e
e
m
p
h
asi
s
o
n
s
y
l
lab
les
d
if
f
er
s
f
r
o
m
t
h
e
u
n
ac
ce
n
ted
i
m
p
le
m
en
ta
tio
n
o
f
th
e
s
a
m
e
s
y
llab
le.
T
h
e
i
n
tr
o
d
u
ctio
n
o
f
t
h
e
k
n
o
w
led
g
e
o
f
p
r
o
s
o
d
y
in
to
au
to
m
at
io
n
o
f
t
h
e
A
S
R
s
y
s
te
m
s
w
ill
m
ak
e
t
h
e
m
m
o
r
e
in
tel
li
g
en
t a
n
d
s
i
m
i
lar
to
h
u
m
a
n
s
[
11
]
.
2
.
3
.
L
ing
uis
t
ics
Rhy
t
h
m
a
s
Sp
ea
k
er
Sp
ec
if
ic
F
ea
t
ures
i
n
ASR
Sy
s
t
e
m
T
h
e
r
h
y
th
m
co
r
r
esp
o
n
d
s
to
th
e
to
tal
d
u
r
atio
n
o
f
s
p
ee
ch
.
Sev
er
al
ex
p
er
i
m
en
ts
w
er
e
co
n
d
u
cted
to
s
tu
d
y
t
h
e
r
h
y
t
h
m
ic
p
atter
n
o
f
s
p
ee
ch
b
y
r
ep
lacin
g
t
h
e
o
r
ig
in
al
s
y
llab
le
w
it
h
a
m
ea
n
in
g
le
s
s
s
y
llab
le,
p
r
eser
v
in
g
t
h
e
o
r
ig
in
al
d
u
r
ati
o
n
/
d
u
r
atio
n
a
n
d
th
e
o
r
ig
i
n
a
l
s
tr
ess
p
atter
n
.
Fo
r
e
x
a
m
p
le,
“
MA
N
i
n
ST
R
E
E
T
”
m
i
m
ics
"
ad
Dad
aDa
"
in
w
h
ic
h
ca
p
ital
letter
s
ar
e
ac
ce
n
ted
,
ass
u
m
in
g
t
h
at
t
h
e
s
y
llab
le
i
s
th
e
b
as
ic
u
n
it
o
f
s
p
ee
ch
s
y
n
c
h
r
o
n
izatio
n
.
T
h
is
ca
n
b
e
d
o
n
e
i
n
t
w
o
w
a
y
s
,
e
ith
er
to
p
r
eser
v
e
t
h
e
to
n
e
p
at
ter
n
o
f
t
h
e
o
r
ig
in
a
l
u
tter
an
ce
,
o
r
to
r
e
m
ai
n
m
o
n
o
to
n
o
u
s
.
T
h
is
e
x
p
er
i
m
en
t
d
ea
l
s
w
it
h
t
h
e
te
m
p
o
r
al
m
o
d
els
a
s
s
o
ciate
d
w
it
h
t
h
e
p
er
ce
iv
ed
s
tr
u
c
t
u
r
e
o
f
t
h
e
p
h
o
n
etic
r
h
y
t
h
m
s
ar
e
n
o
lo
n
g
er
e
m
p
h
a
s
i
ze
s
t
h
e
asp
ec
t
s
t
h
at
ar
e
n
o
t
e
n
h
a
n
ci
n
g
ef
f
icien
ic
y
o
f
A
S
R
[
12
]
.
E
v
en
in
t
h
e
ab
s
en
ce
o
f
lan
g
u
ag
e,
b
ab
ies
ar
e
ab
le
to
r
ec
o
g
n
ize
t
h
e
f
a
m
i
l
iar
k
n
o
w
led
g
e
o
f
r
h
y
t
h
m
ic
p
atter
n
s
.
Ho
w
e
v
er
,
d
if
f
er
e
n
t
m
o
d
es
th
at
ca
u
s
e
co
n
ti
n
u
o
u
s
c
h
an
g
es
ca
n
n
o
t
b
e
ea
s
il
y
s
ep
ar
ated
.
H
is
to
g
r
a
m
o
f
al
l
th
e
p
itc
h
p
e
r
io
d
s
f
o
u
n
d
i
n
t
h
e
s
p
ee
c
h
s
i
g
n
al
co
lleg
e
d
i
s
tr
ib
u
ted
ac
o
r
d
in
g
to
th
eir
f
u
n
d
a
m
en
ta
l
f
r
eq
u
e
n
c
y
i
s
s
h
o
w
n
in
F
ig
u
r
e
3
.
W
h
er
e
A
lt
T
x
is
a
h
is
to
g
r
a
m
o
f
all
t
h
e
p
itch
p
er
io
d
s
,
r
eg
u
lar
T
x
is
a
h
is
to
g
r
a
m
o
f
all
th
e
r
e
g
u
l
ar
p
itch
p
er
io
d
s
an
d
Q
x
is
a
h
is
to
g
r
a
m
o
f
th
e
clo
s
ed
q
u
o
tien
t
v
al
u
es
d
i
s
tr
ib
u
ted
ac
o
r
d
in
g
to
th
eir
f
u
n
d
a
m
e
n
tal
f
r
eq
u
en
c
y
F
o
.
T
h
e
clo
s
ed
q
u
o
tien
t
is
an
est
i
m
a
te
o
f
th
er
p
ec
en
tag
e
ti
m
e
t
h
e
v
o
ca
l
f
o
ld
s
r
e
m
ai
n
ed
clo
s
ed
in
ea
c
h
p
itc
h
p
er
io
d
.
J
itter
is
a
m
ea
s
u
r
e
o
f
p
er
io
d
-
to
-
p
er
io
d
f
lu
ct
u
atio
n
s
i
n
d
u
r
atio
n
f
r
o
m
th
e
m
ea
n
tak
e
n
o
v
er
5
p
itch
c
y
cles.
0
0
.
5
1
1
.
5
2
2
.
5
3
-
0
.
4
-
0
.
3
-
0
.
2
-
0
.
1
0
0
.
1
0
.
2
0
.
3
0
.
4
0
.
5
(
a
)
P
l
o
t
o
f
a
c
o
u
s
t
i
c
s
i
g
n
a
l
"
C
o
l
l
e
g
e
o
f
E
n
g
i
n
e
e
r
i
n
g
S
c
i
e
n
c
e
&
T
e
c
h
n
o
l
o
g
y
"
T
i
m
e
[
s
]
A
m
p
l
i
t
u
d
e
(
n
o
r
m
a
l
i
ze
d
)
500
1000
1500
2000
2500
3000
0
50
100
150
200
Hz
T
i
m
e
(
m
s
e
c
s
)
(
b
)
C
o
r
r
e
s
p
o
n
d
i
n
g
v
a
r
i
a
t
i
o
n
o
f
F
0
v
a
l
u
e
s
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
4
4
3
-
2450
2446
Fig
u
r
e
3
.
H
is
to
g
r
a
m
o
f
all
t
h
e
p
itch
p
er
io
d
s
f
o
u
n
d
i
n
th
e
r
ec
o
r
d
in
g
,
d
is
tr
ib
u
ted
ac
o
r
d
in
g
to
th
eir
f
u
n
d
a
m
e
n
tal
f
r
eq
u
en
c
y
3.
P
RO
B
AB
I
L
I
S
T
I
C
F
O
RM
UL
AT
I
O
N
O
F
H
I
G
H
L
E
V
E
L
SPEAK
E
R
SPEC
I
F
I
C
F
E
A
T
UR
E
S
I
d
en
tify
t
h
e
p
r
o
b
le
m
t
h
at
ca
n
b
e
ex
p
r
ess
ed
as
th
e
m
o
s
t
li
k
el
y
s
p
ea
k
er
o
r
lan
g
u
a
g
e
o
r
e
m
o
tio
n
al
o
r
s
o
u
n
d
B
*
o
f
i
n
p
u
t
s
p
ee
ch
f
r
o
m
k
n
o
w
n
s
p
ea
k
er
s
o
r
lan
g
u
ag
es
o
r
m
o
o
d
s
o
r
s
o
u
n
d
u
n
its
.
L
et
{
B
i
}
,
1
≤
j
≤
N
i
s
d
en
o
ted
th
e
s
et
o
f
clas
s
es
w
h
ich
is
r
ep
r
esen
ti
n
g
lan
g
u
ag
e,
s
p
ea
k
er
an
d
s
o
u
n
t
u
n
it.
T
h
e
o
b
s
er
v
atio
n
d
er
iv
e
d
f
r
o
m
t
h
e
i
n
p
u
t
o
f
s
a
m
p
le
s
p
ee
ch
s
i
g
n
al
i
s
d
en
o
ted
b
y
O
.
T
h
e
p
r
o
b
a
b
ilis
tic
f
o
r
m
u
latio
n
o
f
h
i
g
h
lev
el
s
p
ea
k
er
s
p
ec
if
ic
f
ea
t
u
r
es c
an
b
e
f
o
r
m
u
lated
as f
o
llo
w
s
:
∗
=
(
|
)
(
1
)
W
h
er
e
p
o
s
ter
io
r
p
r
o
b
a
b
ilit
y
o
f
class
B
i
is
r
ep
r
esen
ted
as
P
(
B
i
|
O
)
f
o
r
a
co
n
s
id
er
ed
s
p
ee
ch
s
ig
n
a
l
u
tter
an
ce
o
f
a
s
p
ek
er
ex
p
r
es
s
e
d
in
ter
m
s
o
f
O
.
T
o
r
e
p
r
esen
t
p
r
o
b
ab
ilis
tic
f
o
r
m
u
latio
n
ass
u
m
i
n
g
o
b
s
er
v
a
tio
n
O
b
elo
n
g
i
n
g
to
o
n
o
f
th
e
N
class
es
{
B
i
}
,
1
≤
j
≤
N
.
A
s
p
er
r
u
le
d
ef
in
ed
in
(
1
)
th
e
m
ain
ai
m
is
to
ch
o
o
s
e
th
e
o
b
j
ec
tiv
e
o
f
class
B
i
f
o
r
p
o
s
ter
io
r
p
r
o
b
ab
ilit
y
P
(
B
i
|
O
)
m
u
s
t
b
e
m
a
x
i
m
u
m
f
o
r
a
g
iv
e
n
O
.
A
p
p
l
y
i
n
g
B
ay
es
r
u
l
e
to
o
b
tain
p
o
s
ter
io
r
i p
r
o
b
a
b
ilit
y
,
(
|
)
=
(
|
)
(
)
(
)
(
2
)
W
h
er
e
lik
eli
h
o
o
d
p
r
o
b
a
b
ilit
y
i
s
r
ep
r
esen
ted
as
(
|
)
o
f
w
h
ic
h
i
s
co
r
r
esp
o
n
d
in
g
to
th
e
c
lass
.
T
h
e
p
r
io
r
i p
r
o
b
ab
ilit
y
o
f
t
h
e
cl
ass
is
r
ep
r
esen
ted
as
(
)
.
T
h
en
th
e
p
r
o
b
lem
ca
n
b
e
f
o
r
m
u
lated
a
s
f
o
llo
w
s
:
∗
=
(
|
)
(
)
(
)
(
3
)
T
h
er
e
is
n
o
r
ea
s
o
n
to
c
o
n
s
id
er
o
v
er
lap
p
in
g
th
e
class
,
(
)
ca
n
b
e
co
n
s
id
er
ed
eq
u
al
f
o
r
all
class
es
o
f
d
if
f
er
e
n
t
s
p
ea
k
er
g
r
o
u
p
s
.
Her
e
(
)
b
elo
n
g
s
to
all
class
e
s
,
th
e
p
r
o
b
a
b
ilis
tic
p
r
o
b
lem
ca
n
b
e
s
i
m
p
li
f
ied
to
r
ed
u
ce
th
e
co
m
p
u
tatio
n
al
co
m
p
lex
it
y
a
s
f
o
llo
w
s
:
∗
=
(
|
)
(
4
)
T
h
u
s
s
p
ea
k
er
o
r
lin
g
u
is
tic
o
r
e
m
o
tio
n
al
o
r
s
p
ee
ch
r
ec
o
g
n
i
tio
n
tas
k
s
ar
e
r
eg
ar
d
ed
as
est
i
m
ate
s
o
f
p
o
s
ter
io
r
p
r
o
b
a
b
ilit
ies an
d
ca
n
b
e
r
ed
u
c
ed
to
lik
elih
o
o
d
p
r
o
b
a
b
ilit
y
est
i
m
ate
s
u
n
d
er
s
p
ec
if
ic
as
s
u
m
p
t
io
n
s
.
3
.
1
.
Sp
ea
k
er
s
pecif
ic
f
ea
t
ure
a
s
pect
o
f
ind
iv
idu
a
l sp
ee
ch
s
ig
na
l
Sp
ea
k
er
ch
ar
ac
ter
i
s
tic
s
v
ar
y
d
u
e
to
th
e
d
if
f
er
en
ce
i
n
p
h
y
s
io
lo
g
ical
c
h
ar
ac
ter
is
tic
s
o
f
s
p
ee
c
h
p
r
o
d
u
ctio
n
o
r
g
an
s
a
n
d
ac
q
u
ir
ed
o
r
lear
n
ed
h
ab
its
.
Featu
r
es
o
f
A
SR
ar
e
r
o
u
g
h
l
y
d
i
v
id
ed
in
to
f
o
u
r
g
r
o
u
p
s
o
f
co
n
tin
u
o
u
s
,
q
u
alitati
v
e,
s
p
ec
t
r
al,
an
d
t
ea
g
er
b
ased
en
er
g
y
o
p
er
ato
r
f
ea
tu
r
es,
an
d
p
r
o
s
o
d
ic
f
ea
tu
r
es
ar
e
class
i
f
ied
in
to
ca
te
g
o
r
ies
o
f
co
n
tin
u
o
u
s
s
p
ee
ch
f
ea
t
u
r
es
[
13
]
.
R
h
y
th
m
ic
f
ea
t
u
r
es
ar
e
r
eliab
le
in
d
icato
r
s
o
f
e
m
o
tio
n
a
n
d
ar
e
w
id
el
y
u
s
ed
f
o
r
em
o
tio
n
al
r
ec
o
g
n
itio
n
[
14
]
.
T
h
e
ar
o
u
s
al
s
tate
o
f
t
h
e
s
p
ea
k
er
h
as b
ee
n
s
t
u
d
ied
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
E
lec
&
C
o
m
p
E
n
g
IS
SN: 2
0
8
8
-
8708
Hig
h
leve
l sp
ea
ke
r
s
p
ec
ific fe
a
tu
r
es a
s
a
n
efficien
cy
en
h
a
n
cin
g
p
a
r
a
mete
r
s
in
s
p
ea
ke
r
…
(
S
a
tya
n
a
n
d
S
in
g
h
)
2447
to
in
f
l
u
e
n
ce
th
e
o
v
er
all
en
er
g
y
,
f
r
eq
u
en
c
y
,
an
d
d
u
r
atio
n
o
f
th
e
v
o
ice
p
au
s
e
[
1
5
]
.
E
m
o
tio
n
s
li
k
e
an
g
er
ar
e
ch
ar
ac
ter
ized
b
y
a
h
ig
h
s
p
ee
c
h
r
ate,
b
u
t f
ee
lin
g
s
o
f
s
ad
n
e
s
s
a
r
e
r
elate
d
to
w
h
is
p
er
ed
s
p
ee
d
.
G
au
s
s
ia
n
Mi
x
er
Mo
d
el
(
G
MM
)
[
16
]
an
d
n
eu
r
al
n
e
t
w
o
r
k
[
17
]
w
er
e
s
u
cc
e
s
s
f
u
ll
y
u
s
ed
f
o
r
e
m
o
tio
n
al
r
ec
o
g
n
itio
n
.
SVM
i
s
w
id
el
y
u
s
ed
b
y
r
ese
ar
ch
er
s
to
class
if
y
e
m
o
tio
n
s
[
18
]
.
Dee
p
n
eu
r
al
n
et
w
o
r
k
s
(
DNN
)
ca
n
b
e
u
s
ed
to
o
b
tain
h
i
g
h
er
le
v
el
f
ea
t
u
r
es
f
r
o
m
lo
w
-
lev
el
ac
o
u
s
tic
f
ea
t
u
r
es
an
d
t
h
en
to
o
th
er
clas
s
if
ier
s
f
o
r
e
m
o
tio
n
r
ec
o
g
n
it
i
o
n
.
I
n
[1
9
]
,
f
ea
tu
r
es
o
f
t
h
e
s
e
g
m
e
n
tatio
n
le
v
el
in
cl
u
d
in
g
Mel
-
f
r
eq
u
en
c
y
C
ep
s
tr
a
l
C
o
ef
f
icie
n
ts
(
MF
C
C
)
,
p
itch
-
b
ased
f
ea
t
u
r
es
(
p
itc
h
p
er
io
d
a
n
d
h
ar
m
o
n
ic
to
n
o
is
e
r
atio
)
,
an
d
t
h
eir
d
elta
v
al
u
es
ar
e
ex
tr
ac
ted
.
3
.
2
F
us
ing
hig
her
s
pea
k
er
s
p
ec
if
ic
f
ea
t
ure
into
co
nv
ent
io
na
l A
SR a
pp
lica
t
io
n
T
h
e
p
r
o
s
o
d
ic
m
o
d
el
p
r
o
v
id
es
an
ad
d
itio
n
al
k
n
o
w
led
g
e
s
o
u
r
ce
th
at
th
e
ac
o
u
s
tic
m
o
d
el
ca
n
n
o
t
p
r
o
v
id
e.
T
h
is
m
a
y
h
elp
to
o
v
e
r
co
m
e
s
o
m
e
o
f
t
h
e
m
is
s
id
e
n
ti
f
icatio
n
s
.
T
h
er
ef
o
r
e,
co
m
b
i
n
i
n
g
i
n
f
o
r
m
atio
n
f
r
o
m
m
u
ltip
le
s
o
u
r
ce
s
o
f
e
v
id
e
n
ce
,
k
n
o
w
n
as
f
u
s
io
n
tec
h
n
o
lo
g
y
,
h
as
b
ee
n
w
id
el
y
u
s
ed
in
s
p
ea
k
er
s
,
la
n
g
u
a
g
es
,
e
m
o
tio
n
s
,
an
d
s
p
ee
ch
.
T
y
p
ic
all
y
,
m
a
n
y
d
if
f
er
e
n
t
f
ea
t
u
r
e
s
ets
ar
e
e
x
tr
ac
ted
f
r
o
m
t
h
e
s
p
ee
ch
s
ig
n
al,
th
e
n
a
s
ep
ar
ate
class
if
ier
is
u
s
ed
f
o
r
ea
ch
f
ea
tu
r
e
s
et,
th
e
n
s
u
b
-
s
c
o
r
es
o
r
d
ec
is
io
n
s
ar
e
co
m
b
in
ed
.
T
h
is
m
ea
n
s
th
at
ea
ch
s
p
ea
k
er
s
to
r
es
a
p
l
u
r
alit
y
o
f
s
p
ea
k
er
m
o
d
els
i
n
th
e
d
atab
ase.
I
t
i
s
g
en
er
all
y
b
elie
v
ed
th
at
s
u
cc
es
s
f
u
l
f
u
s
io
n
s
y
s
te
m
s
s
h
o
u
ld
b
e
co
m
b
in
ed
in
to
in
d
ep
en
d
e
n
t
f
ea
t
u
r
es.
P
o
s
s
ib
le
lo
w
-
le
v
el
s
p
ec
tr
al
ch
ar
ac
ter
is
t
ics,
p
r
o
s
o
d
ic
f
u
n
ctio
n
,
ad
v
an
ce
d
f
u
n
ct
io
n
.
T
h
e
s
i
m
p
le
s
t
f
u
s
io
n
m
et
h
o
d
is
to
co
m
b
i
n
e
clas
s
i
f
ie
r
o
u
tp
u
t
s
co
r
es
w
it
h
w
ei
g
h
ted
s
u
m
s
.
T
h
at
is
,
a
g
iv
e
n
s
u
b
s
co
r
e
s
k
is
a
f
u
s
io
n
m
atc
h
o
f
th
e
i
n
d
ex
clas
s
i
f
ier
k
.
=
∑
(
,
)
=
1
ℎ
=
,
=
ℎ
(
5
)
An
o
th
er
w
a
y
to
co
m
b
i
n
e
f
ea
tu
r
es
at
t
h
e
s
co
r
e
le
v
el
i
s
to
u
s
e
co
n
f
id
e
n
ce
m
ea
s
u
r
e.
I
n
[
1
6
]
,
th
e
a
u
th
o
r
co
n
f
ir
m
ed
t
h
at
co
n
f
i
d
e
n
ce
b
as
ed
f
u
s
io
n
co
m
p
le
m
en
tar
y
f
ea
t
u
r
e
m
eth
o
d
o
f
co
m
b
i
n
i
n
g
w
av
elet
m
u
lt
ip
licatio
n
co
ef
f
icie
n
t
s
r
esid
u
al
(
W
OC
O
R
)
an
d
MFC
C
f
u
n
c
tio
n
f
o
r
s
p
ea
k
er
r
ec
o
g
n
itio
n
.
T
h
is
m
etr
ic
is
d
er
iv
ed
f
r
o
m
th
e
lik
eli
h
o
o
d
s
co
r
e
o
b
tain
ed
f
r
o
m
t
h
e
t
w
o
f
ea
t
u
r
es
.
I
n
o
r
d
er
to
ca
m
p
u
t
e
th
e
co
n
f
id
en
ce
m
ea
s
u
r
e
(
C
M)
,
t
h
e
d
is
cr
i
m
i
n
atio
n
ab
ili
t
y
o
f
ea
c
h
f
ea
t
u
r
e
in
a
p
ar
ticu
lar
r
ec
o
g
n
it
io
n
test
is
f
ir
s
t c
alcu
lated
i
s
g
i
v
en
a
s
f
o
llo
w
s
:
=
|
(
/
,
)
|
⁄
(
6
)
w
h
er
e
LL
R
j
=
l
ogP
(
s
j
λ
c
,
j
⁄
)
−
l
ogP
(
s
j
λ
u
,
j
⁄
)
(
7
)
T
h
e
lo
g
-
li
k
eli
h
o
o
d
s
o
f
th
e
cl
i
en
t
m
o
d
el
a
n
d
b
ac
k
g
r
o
u
n
d
m
o
d
el
ar
e
r
ep
r
esen
ted
in
(
6
)
an
d
eq
n
.
(
7
)
r
esp
ec
tiv
el
y
.
T
h
e
co
m
p
u
tat
io
n
o
f
t
h
e
d
is
cr
i
m
i
n
atio
n
r
atio
b
ased
o
n
th
e
v
al
u
e
f
u
n
ctio
n
o
f
ea
ch
tr
ial
is
DR
=
D
1
D
2
⁄
.
Nex
t,
co
n
f
id
an
ce
m
etr
ic
is
ca
m
p
u
ted
b
ased
o
n
th
e
D
R
v
alu
e
as f
o
llo
w
s
:
=
−
(
1
1
+
(
−
)
)
(
8
)
T
h
e
v
alu
es
o
f
α
an
d
β
w
er
e
d
ete
r
m
in
ed
b
y
s
etti
n
g
t
h
e
d
ev
elo
p
m
en
t
d
ata
to
0
.
7
5
an
d
2
,
r
esp
e
ctiv
el
y
.
B
ased
o
n
C
M,
s
co
r
e
lev
el
f
u
s
i
o
n
is
d
o
n
e,
w
h
ic
h
r
ep
r
esen
ted
as f
o
llo
w
s
:
LLR
=
LL
R
1
+
LL
R
2
.
CM
(
9
)
As th
e
f
u
s
io
n
s
co
r
e
co
m
b
in
e
s
w
ei
g
h
ted
LL
R
1
an
d
LL
R
2
,
th
i
s
C
M
b
ase
d
s
co
r
in
g
f
u
s
io
n
m
et
h
o
d
y
ield
s
b
etter
r
esu
lts
i
n
ter
m
s
o
f
f
ix
ed
w
ei
g
h
t f
u
s
io
n
ar
e
r
ep
r
esen
ted
as f
o
llo
w
s
:
2
.
5
4
0
3
,
2
.
4
9
6
0
,
2
.
3
4
0
0
,
1
.
6
5
0
2
,
1
.
5
5
8
5
,
1
.
4
6
6
9
,
1
.
3
7
6
8
,
1
.
2
0
9
7
,
1
.
1
3
7
7
,
1
.
0
3
2
0
,
1
.
0
0
5
4
,
1
.
0
2
3
9
,
1
.
2
8
1
3
,
1
.
4
3
5
8
,
2
.
7
5
5
5
,
2
.
9
1
0
6
,
2
.
9
8
0
7
,
2
.
8
6
1
7
,
1
.
0
1
9
3
,
1
.
0
3
9
2
,
2
.
9
9
7
4
,
2
.
8
3
9
2
,
1
.
0
6
9
6
,
2
.
9
9
0
9
,
2
.
6
7
1
7
,
1
.
0
2
3
1
,
2
.
9
7
8
9
,
1
.
0
2
9
3
,
2
.
9
1
9
1
,
1
.
9
5
6
0
,
1
.
5
8
8
7
,
1
.
1
9
9
3
,
1
.
3
9
7
3
,
1
.
0
5
3
8
,
2
.
9
9
9
0
,
1
.
0
5
7
7
,
1
.
6
9
1
3
,
1
.
0
3
5
7
.
4.
L
AR
G
E
M
ARG
I
N
AP
P
RO
AC
H
F
O
R
L
E
AR
NIN
G
AL
I
G
NM
E
N
T
I
N
ASR
AP
P
L
I
CATI
O
N
A
s
u
p
er
v
is
ed
lear
n
i
n
g
a
lg
o
r
ith
m
f
o
r
ali
g
n
m
e
n
t
r
ec
eiv
es
a
tr
ai
n
in
g
s
et
as
i
n
p
u
t
=
{
(
x
̅
1
,
p
̅
1
,
s
̅
1
)
}
,
…
…
{
(
x
̅
m
−
1
,
p
̅
m
−
1
,
s
̅
m
−
1
)
}
,
{
(
x
̅
m
,
p
̅
m
,
s
̅
m
)
}
w
h
ich
r
et
u
r
n
s
a
ali
n
m
e
n
ed
f
u
n
ctio
n
f
.
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
4
4
3
-
2450
2448
T
o
p
r
o
m
o
te
e
f
f
icie
n
t
al
g
o
r
ith
m
s
,
I
r
estrict
to
a
l
i
m
i
ted
k
in
d
o
f
ali
g
n
m
e
n
t
f
u
n
ctio
n
.
Mo
r
e
s
p
ec
if
icall
y
,
it
i
s
ass
u
m
ed
t
h
at
th
er
e
i
s
a
s
et
o
f
p
r
ed
ef
in
ed
b
asic
ali
g
n
ed
f
ea
t
u
r
e
f
u
n
ctio
n
s
{
φ
j
}
j
=
1
n
.
E
ac
h
b
ase
ali
g
n
m
en
t
s
p
ea
k
er
s
p
ec
if
ic
f
ea
t
u
r
e
is
a
f
u
n
ctio
n
o
f
th
e
f
o
r
m
φ
j
:
∗
.
∗
.
ℕ
∗
ℝ
.
T
h
at
is
,
ea
ch
b
asic
ali
g
n
m
e
n
t
s
p
ea
k
er
s
p
ec
if
i
c
f
ea
t
u
r
e
x
̅
an
d
s
p
ea
k
er
s
p
ec
if
ic
p
h
o
n
e
m
e
s
eq
u
e
n
ce
p
̅
,
to
g
eth
er
w
it
h
t
h
e
ca
n
d
id
ate
ti
m
i
n
g
s
eq
u
en
ce
s
̅
,
r
etu
r
n
s
a
s
ca
lar
v
is
u
all
y
r
ep
r
esen
ti
n
g
t
h
e
co
n
f
id
en
ce
le
v
el
o
f
th
e
s
u
g
g
ested
ti
m
i
n
g
s
eq
u
e
n
ce
s
̅
.
φ
{
x
̅
,
p
̅
,
s
̅
}
is
d
en
o
tin
g
ℝ
n
v
ec
to
r
w
h
o
s
e
j
th
el
m
en
t i
s
φ
j
{
x
̅
,
p
̅
,
s
̅
}
.
I
n
th
is
p
ap
er
I
am
u
s
i
n
g
th
e
t
h
e
a
lig
n
m
e
n
t
f
u
n
ct
io
n
d
ef
i
n
ed
as f
o
llo
w
s
:
(
̅
,
̅
)
=
ma
x
̅
.
{
̅
,
̅
,
̅
}
(
1
0
)
W
ith
th
e
SVM
al
g
o
r
ith
m
f
o
r
b
in
ar
y
clas
s
if
icatio
n
,
th
e
m
et
h
o
d
o
f
s
elec
tin
g
th
e
w
ei
g
h
t
v
ec
to
r
is
b
ased
o
n
th
e
co
n
ce
p
t
o
f
lar
g
e
m
ar
g
i
n
s
ep
ar
atio
n
.
B
u
t
i
n
th
i
s
ca
s
e,
ti
m
i
n
g
is
n
o
t
j
u
s
t
r
i
g
h
t
o
r
w
r
o
n
g
.
T
h
er
ef
o
r
e,
m
y
g
o
al
is
n
o
t
to
s
ep
ar
ate
th
e
r
ig
h
t
t
i
m
in
g
f
r
o
m
t
h
e
w
r
o
n
g
ti
m
i
n
g
,
b
u
t
to
tr
y
to
s
o
r
t
t
h
e
s
eq
u
en
ce
b
y
q
u
a
lit
y
.
I
n
th
eo
r
y
,
m
y
m
et
h
o
d
ca
n
b
e
d
escr
ib
ed
as
a
t
w
o
-
s
tep
p
r
o
ce
s
s
:
f
ir
s
t
b
u
ild
a
v
ec
to
r
φ
{
x
̅
,
p
̅
,
s
̅
′
}
in
v
ec
to
r
s
p
ac
e
ℝ
n
in
i
n
s
id
en
t
b
ased
ap
p
r
o
ac
h
(
x
̅
i
,
p
̅
i
)
in
a
tr
ai
n
i
n
g
s
et
an
d
ea
ch
p
o
s
s
ib
le
ti
m
i
n
g
s
eq
u
e
n
ce
s
̅
′
.
Se
co
n
d
I
f
i
n
d
a
v
ec
to
r
th
at
p
r
o
j
ec
ts
th
e
v
ec
to
r
to
an
d
s
o
r
ts
th
e
v
ec
t
o
r
s
w
∈
ℝ
n
co
n
s
tr
u
cted
in
t
h
e
f
ir
s
t
s
te
p
ab
o
v
e
b
ased
o
n
its
q
u
al
it
y
.
I
d
ea
ll
y
,
f
o
r
ea
ch
i
n
s
ta
n
ce
(
x
̅
i
,
p
̅
i
)
an
d
ev
er
y
s
u
g
g
e
s
tab
le
ti
m
i
n
g
to
k
ee
p
t
h
e
f
o
llo
w
i
n
g
co
n
s
tr
ain
ts
:
W
.
φ
(
x
̅
i
,
p
̅
i
,
s
̅
i
)
−
w
.
φ
(
x
̅
i
,
p
̅
i
,
s
̅
′
)
≥
γ
(
s
̅
i
,
s
̅
′
)
(
1
1
)
T
h
e
co
m
p
u
ter
s
i
m
u
l
a
ted
v
ec
to
r
s
w
∈
ℝ
n
co
n
s
tr
u
cted
in
th
e
f
ir
s
t
s
tep
b
ased
o
n
its
q
u
alit
y
is
r
ep
r
esen
ted
as f
o
llo
w
:
0
.
0
0
0
0
,
0
.
0
5
0
5
,
0
.
2
0
2
0
,
0
.
6
5
6
6
,
0
.
7
0
7
1
,
0
.
7
5
7
6
,
0
.
8
0
8
1
,
0
.
9
0
9
1
,
0
.
9
5
9
6
,
1
.
0
6
0
6
,
1
.
1
1
1
1
,
1
.
2
1
2
1
,
1
.
3
6
3
6
,
1
.
4
1
4
1
,
1
.
7
1
7
2
,
1
.
7
6
7
7
,
1
.
8
6
8
7
,
1
.
9
1
9
2
,
2
.
2
2
2
2
,
2
.
2
7
2
7
,
2
.
5
2
5
3
,
2
.
5
7
5
8
,
2
.
7
7
7
8
,
2
.
9
2
9
3
,
2
.
9
7
9
8
,
3
.
0
8
0
8
,
3
.
2
3
2
3
,
3
.
3
3
3
3
,
3
.
4
3
4
3
,
3
.
5
8
5
9
,
3
.
7
3
7
4
,
3
.
9
8
9
9
,
4
.
2
9
2
9
,
4
.
4
4
4
4
,
4
.
5
4
5
5
,
4
.
6
9
7
0
,
4
.
7
9
8
0
,
4
.
9
4
9
5
.
W
h
er
e
γ
(
s
̅
i
,
s
̅
′
)
is
co
s
t
f
u
n
ctio
n
ass
e
s
s
in
g
t
h
e
q
u
al
it
y
o
f
s
eq
u
e
n
ce
s
.
T
h
e
co
n
s
tr
ain
t
o
f
t
h
e
ex
p
r
es
s
i
o
n
in
(
1
0
)
m
ea
n
s
th
at
t
h
e
m
ar
g
i
n
o
f
w
w
it
h
r
esp
ec
t
to
p
o
s
s
ib
le
tim
i
n
g
s
eq
u
e
n
ce
s
̅
′
m
u
s
t
b
e
g
r
ea
t
er
th
an
th
e
co
s
t
o
f
th
e
p
r
ed
ictio
n
s
̅
′
,
n
o
t
t
h
e
tr
u
e
ti
m
i
n
g
s
̅
i
.
Of
co
u
r
s
e,
i
f
th
e
w
r
an
k
i
s
d
i
f
f
er
e
n
t
a
n
d
t
h
e
p
o
s
s
ib
le
ti
m
i
n
g
i
s
ca
lcu
lated
co
r
r
ec
t
ly
,
th
e
m
ar
g
in
r
eq
u
ir
e
m
e
n
t
g
i
v
e
n
b
y
(
1
1
)
ca
n
s
i
m
p
l
y
b
e
s
atis
f
ied
b
y
m
u
ltip
l
y
in
g
w
b
y
a
lar
g
e
s
ca
lar
.
T
h
e
SVM
a
lg
o
r
it
h
m
is
s
u
b
j
ec
ted
to
th
e
co
n
s
tr
ain
ts
g
i
v
e
n
i
n
(
1
1
)
b
y
m
i
n
i
m
i
zin
g
1
2
‖
w
‖
2
w
h
ich
s
o
lv
es
th
is
p
r
o
b
le
m
r
ep
r
ese
n
t
ed
in
(
1
0
)
.
Dec
is
io
n
b
o
u
n
d
ar
i
es
o
f
m
u
ltic
lass
SVM
v
s
co
n
f
id
en
ce
m
ea
s
u
r
e
is
s
h
o
w
n
in
F
ig
u
r
e
4
.
Fig
u
r
e
4
.
Dec
is
io
n
B
o
u
n
d
ar
ies
o
f
Mu
lticla
s
s
SVM
I
n
f
ac
t,
th
er
e
ar
e
ca
s
es
w
h
er
e
th
e
co
n
s
tr
ai
n
t
g
i
v
e
n
b
y
(
1
1
)
c
an
n
o
t
b
e
s
atis
f
ied
.
I
n
o
r
d
er
to
o
v
er
co
m
e
th
is
o
b
s
tacle
,
th
e
f
o
llo
w
i
n
g
h
in
g
e
lo
s
s
f
u
n
ctio
n
i
s
d
ef
in
ed
f
o
r
alig
n
m
en
t
i
n
ac
co
r
d
an
ce
w
it
h
t
h
e
s
o
f
t
SVM
m
et
h
o
d
.
L
ar
g
e
m
ar
g
i
n
e
a
p
p
r
o
ac
h
f
o
r
l
ea
r
n
i
n
g
a
li
g
n
m
e
n
t
o
f
s
p
ea
k
er
1
an
d
s
p
ea
k
er
s
2
i
s
s
h
o
w
n
i
n
Fig
u
r
e
5
.
C
o
m
p
u
ter
s
i
m
u
late
d
co
s
t
f
u
n
ctio
n
γ
(
s
̅
i
,
s
̅
′
)
ass
ess
i
n
g
t
h
e
q
u
alit
y
o
f
s
eq
u
e
n
ce
s
is
r
ep
r
e
s
en
ted
as
f
o
llo
w
s
:
0
.
2
4
2
3
,
0
.
1
4
0
6
,
0
.
1
0
3
1
,
-
0
.
1
8
4
4
,
-
0
.
0
2
3
3
,
-
0
.
0
1
6
2
,
-
0
.
2
7
5
8
,
-
0
.
2
2
6
9
,
-
0
.
2
0
4
6
,
-
0
.
4
6
1
7
,
-
0
.
0
7
6
4
,
-
0
.
5
9
5
1
,
-
0
.
4
0
9
5
,
-
0
.
0
8
7
7
,
0
.
1
7
3
8
,
0
.
3
0
8
0
,
0
.
3
7
0
5
,
0
.
2
6
0
8
,
-
0
.
5
2
4
8
,
-
0
.
5
1
6
1
,
0
.
6
8
9
3
,
0
.
2
6
5
8
,
-
1
.
4
6
2
4
,
0
.
9
4
4
2
,
1
.
6
2
7
1
,
-
3
.
1
5
3
2
,
3
.
4
5
9
3
,
-
3
.
5
7
7
4
,
1
.
8
6
9
1
,
-
0
.
5
8
8
0
,
-
0
.
2
1
7
5
,
0
.
3
4
3
1
,
0
.
2
9
8
6
,
-
0
.
9
5
8
1
,
1
.
1
6
8
6
3
,
-
1
.
7
0
6
3
,
0
.
8
6
2
5
,
-
1
.
1
1
5
0
.
-1
-
0
.
5
0
0
.
5
1
1
.
5
2
0
0
.
2
0
.
4
0
.
6
0
.
8
1
1
.
2
1
.
4
1
.
6
1
.
8
2
D
e
c
i
s
i
on
B
ou
nd
a
r
i
e
s
of
m
ul
t
i
c
l
a
s
s
S
V
M
C
os
t
Fu
nc
t
i
on
C
on
f
i
de
nc
e
M
e
a
s
ure
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
E
lec
&
C
o
m
p
E
n
g
IS
SN: 2
0
8
8
-
8708
Hig
h
leve
l sp
ea
ke
r
s
p
ec
ific fe
a
tu
r
es a
s
a
n
efficien
cy
en
h
a
n
cin
g
p
a
r
a
mete
r
s
in
s
p
ea
ke
r
…
(
S
a
tya
n
a
n
d
S
in
g
h
)
2449
T
ab
le.
1
r
ep
r
esen
ts
b
u
ild
i
n
g
la
r
g
e
m
ar
g
i
n
d
i
s
tan
ce
m
atr
ix
f
o
r
lear
n
i
n
g
ali
g
n
m
e
n
t
w
it
h
r
esp
e
ct
to
co
s
t
f
u
n
ctio
n
,
co
n
f
id
an
ce
m
ea
s
u
r
e,
s
u
p
p
o
r
t
an
d
u
p
s
atu
r
atio
n
.
T
ab
le
1
.
B
u
ild
in
g
lar
g
e
m
ar
g
i
n
d
is
ta
n
ce
m
atr
i
x
f
o
r
lear
n
in
g
alig
n
m
e
n
t
C
o
st
F
u
n
c
t
i
o
n
C
o
n
f
i
d
a
n
c
e
M
e
a
su
r
e
#
S
u
p
p
o
r
t
#
U
p
S
a
t
u
r
a
t
i
o
n
1
.
2
4
5
0
1
.
0
0
0
0
1
0
0
.
0
0
0
0
1
.
0
0
0
0
2
0
-
1
.
0
4
5
1
1
.
0
4
5
1
3
0
-
2
.
4
5
3
9
1
.
3
4
7
9
4
0
-
9
.
7
1
1
3
2
.
9
5
7
5
5
0
-
3
.
5
8
5
6
2
.
6
9
2
2
4
0
-
4
.
0
2
3
6
0
.
1
2
2
1
5
0
-
5
.
8
7
7
9
0
.
4
6
0
9
6
0
-
6
.
7
9
8
7
0
.
1
5
6
7
7
0
-
1
.
7
1
2
5
1
.
5
1
8
8
6
0
-
2
.
4
5
0
1
0
.
4
3
0
8
5
0
-
2
.
4
8
9
2
0
.
0
1
6
0
6
0
-
2
.
5
2
7
3
0
.
0
1
5
3
7
0
-
4
.
9
0
8
0
0
.
9
4
2
0
8
0
-
6
.
2
0
2
6
0
.
2
6
3
8
9
0
-
8
.
2
6
7
5
0
.
3
3
2
9
8
0
-
8
.
3
8
7
6
0
.
0
1
4
5
9
0
-
8
.
9
1
0
7
0
.
0
6
2
4
8
0
-
9
.
9
5
1
3
0
.
1
1
6
8
9
0
Fig
u
r
e
5
.
L
ar
g
e
Ma
r
g
in
e
A
p
p
r
o
ac
h
f
o
r
L
ea
r
n
in
g
A
li
g
n
m
e
n
t
o
f
Sp
ea
k
er
1
an
d
Sp
ea
k
er
s
2
.
5.
E
XP
E
RM
E
NT
AL
S
E
T
UP
F
O
R
ACCURACY A
ND RO
B
UST
N
E
SS
O
F
ASR
AP
P
L
I
CATI
O
N
I
p
er
f
o
r
m
ed
a
s
p
ee
ch
-
to
-
s
p
ee
c
h
ex
p
er
i
m
en
t u
s
in
g
t
h
e
T
I
MI
T
an
d
HT
I
MI
T
co
r
p
u
s
o
r
ig
in
all
y
s
a
m
p
led
in
t
h
e
1
6
k
Hz
T
I
MI
T
co
r
p
at
8
k
Hz.
I
n
all,
th
e
ex
p
er
i
m
en
t,
th
e
b
as
elin
e
d
is
cr
i
m
i
n
ati
v
e
s
y
s
t
e
m
a
n
d
t
h
e
HM
M
s
y
s
te
m
w
er
e
tr
ain
ed
o
n
a
leg
i
b
le,
r
ea
d
a
b
le
T
I
MI
T
co
r
p
u
s
w
it
h
a
s
et
o
f
4
8
p
h
o
n
e
m
es
.
I
d
iv
id
ed
th
e
tr
ain
in
g
p
ar
t
o
f
T
I
MI
T
ex
p
r
ess
io
n
s
o
f
S
A
1
an
d
S
A
2
f
r
o
m
in
d
ep
en
d
en
t
p
ar
t
in
to
th
r
ee
d
is
j
o
in
t
p
ar
t
ex
clu
d
in
g
th
e
ex
p
r
ess
io
n
s
o
f
5
0
0
,
1
0
0
an
d
3
0
9
3
r
esp
ec
tiv
el
y
.
T
h
e
s
ec
o
n
d
an
d
th
ir
d
s
ets
o
f
u
tter
a
n
ce
s
1
0
0
an
d
3
0
9
6
re
s
p
ec
tiv
el
y
f
o
r
m
ed
a
A
S
R
r
ec
o
g
n
itio
n
s
et
a
n
d
ASR
tr
ai
n
i
n
g
s
et
f
o
r
th
e
f
o
r
ce
d
alig
n
m
en
t a
lg
o
r
ith
m
.
A
lt
h
o
u
g
h
T
I
MI
T
w
as
o
r
ig
in
al
l
y
s
a
m
p
led
at
1
6
k
Hz,
co
n
f
ir
m
ed
th
e
d
is
cr
i
m
i
n
atio
n
m
o
d
el
b
y
tr
a
i
n
i
n
g
th
e
8
k
Hz
T
I
MI
T
.
T
h
is
w
as
d
o
n
e
to
ev
alu
ate
th
e
al
g
o
r
ith
m
's
p
er
f
o
r
m
a
n
ce
at
a
m
o
r
e
r
ea
lis
tic
s
a
m
p
li
n
g
r
ate
an
d
to
u
s
e
th
e
tr
ain
ed
A
S
R
m
o
d
e
l
o
n
a
s
u
b
-
s
a
m
p
led
T
I
MI
T
o
n
an
o
th
er
8
k
Hz
s
am
p
led
co
r
p
u
s
.
T
h
e
ex
p
er
i
m
e
n
tal
r
e
s
u
l
ts
o
f
AS
R
a
r
e
s
h
o
w
n
i
n
T
ab
le
2
.
I
t
ca
n
b
e
s
ee
n
t
h
at
th
e
r
es
u
lts
o
f
T
I
MI
T
s
u
b
-
s
a
m
p
les
at
8
k
Hz
s
ti
ll e
x
ce
ed
t
h
e
r
esu
l
ts
r
e
p
o
r
ted
b
y
B
r
u
g
n
ar
a
et
al.
[
2
0
]
an
d
Ho
s
o
m
[
2
1
]
.
T
ab
le
2
.
P
r
ed
ef
in
ed
to
ler
an
ce
s
o
f
T
I
MI
T
co
r
p
u
s
,
th
e
e
f
f
icien
c
y
o
f
p
h
o
n
e
m
e
s
p
o
s
itio
n
ed
co
r
r
ec
tl
y
L
e
n
g
t
h
o
f
S
p
e
e
c
h
C
o
r
p
u
s
L
e
ss t
h
a
n
1
0
ms
L
e
ss t
h
a
n
2
0
ms
L
e
ss t
h
a
n
3
0
ms
L
e
ss t
h
a
n
3
0
ms
1
6
K
H
z
TI
M
I
T
U
t
t
e
r
a
n
c
e
s
D
i
scri
m
i
n
a
t
i
v
e
A
l
i
g
n
me
n
t
7
6
.
5
9
0
.
7
7
9
6
.
4
4
9
9
.
2
3
B
r
u
g
n
a
r
a
e
t
a
l
.
(
1
9
9
3
)
7
5
.
3
8
8
.
9
9
4
.
4
9
7
.
1
H
o
so
m (2
0
0
2
)
9
2
.
6
8
K
H
z
TI
M
I
T
U
t
t
e
r
a
n
c
e
s
D
i
scri
m
i
n
a
t
i
v
e
A
l
i
g
n
me
n
t
8
4
.
2
3
9
4
.
2
1
9
7
.
1
2
9
9
.
1
H
TI
M
I
T
C
B
1
8
3
4
U
t
t
e
r
a
n
c
e
s
D
i
scri
m
i
n
a
t
i
v
e
A
l
i
g
n
me
n
t
7
2
.
5
8
9
.
6
8
9
5
.
8
9
9
7
.
4
-1
-1
-1
-1
-1
-1
-1
-1
-1
-1
-1
-1
-1
-1
-1
-1
-1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
S
pe
a
k
e
r
1
S
pe
a
k
e
r
2
La
r
ge
M
a
r
gi
n
A
pp
r
oa
c
h
f
or
Le
a
r
ni
ng
A
l
i
gn
m
e
nt
-4
-3
-2
-1
0
1
2
3
4
-4
-3
-2
-1
0
1
2
3
4
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
4
4
3
-
2450
2450
6.
CO
NCLU
SI
O
N
ASR
i
s
th
e
u
s
e
o
f
a
m
ac
h
i
n
e
t
o
id
en
tify
a
p
er
s
o
n
f
r
o
m
t
h
e
s
p
o
k
en
w
o
r
d
.
T
h
e
A
S
R
s
y
s
te
m
ca
n
b
e
u
s
ed
i
n
t
w
o
m
o
d
es
t
h
at
r
ec
o
g
n
ize
th
e
id
en
tit
y
r
eq
u
ir
ed
b
y
a
p
ar
ticu
lar
p
er
s
o
n
o
r
v
er
if
ier
.
B
asic
k
n
o
w
led
g
e
o
f
s
p
ea
k
er
r
ec
o
g
n
it
io
n
is
co
v
er
ed
,
s
i
m
p
le
f
u
n
ctio
n
s
an
d
m
ea
s
u
r
es
f
o
r
s
p
ea
k
er
r
ec
o
g
n
i
t
io
n
ar
e
p
r
o
p
o
s
ed
an
d
co
m
p
ar
ed
w
it
h
co
n
v
e
n
tio
n
al
r
ec
o
g
n
itio
n
u
s
i
n
g
s
p
ea
k
er
r
ec
o
g
n
i
tio
n
cr
iter
ia.
P
r
o
p
o
s
al
o
f
th
e
A
S
R
s
y
s
te
m
to
d
is
tin
g
u
is
h
s
p
ea
k
er
s
u
s
e
s
h
i
g
h
-
lev
el
s
p
ea
k
er
-
s
p
ec
if
ic
f
ea
tu
r
es.
T
h
is
m
ea
s
u
r
e
m
e
n
t
ca
n
b
e
in
ter
p
r
eted
as
t
h
e
s
h
ap
e
o
f
in
f
o
r
m
a
tio
n
t
h
eo
r
y
m
ea
s
u
r
e
m
e
n
t
b
y
d
is
cr
i
m
in
a
tio
n
ali
g
n
m
e
n
t.
I
n
f
ac
t,
t
h
e
e
x
p
er
im
e
n
t
s
r
ep
o
r
ted
ab
o
v
e
s
h
o
w
t
h
at
d
is
cr
i
m
i
n
ati
v
e
tr
ain
in
g
r
eq
u
ir
es
les
s
s
a
m
p
l
e
tr
ain
in
g
th
a
n
s
p
ee
ch
-
to
-
p
h
o
n
e
alig
n
m
en
t
b
ased
o
n
HM
M
p
r
o
ce
s
s
es.
T
h
e
p
er
f
o
r
m
a
n
ce
o
f
s
p
ea
k
er
r
ec
o
g
n
iti
o
n
ac
cu
r
ac
y
o
f
t
h
e
A
S
R
s
y
s
t
e
m
p
r
o
p
o
s
ed
in
th
is
p
ap
er
is
9
9
.
2
3
%
an
d
9
9
.
1
%
f
o
r
<
40
of
T
I
MI
T
u
tter
an
ce
s
r
es
p
ec
tiv
el
y
.
P
r
o
p
o
s
ed
A
S
R
s
y
s
t
e
m
s
h
o
ws
1
.
9
9
%,
2
.
1
0
%,
2.
1
6
%
an
d
2
.
1
9
%
o
f
im
p
r
o
v
e
m
e
n
t
s
co
m
p
ar
e
to
tr
ad
itio
n
al
ASR
s
y
s
te
m
f
o
r
<
10
ms
,
<
20
ms
,
<
30
ms
an
d
<
40
ms
o
f
1
6
KHz
T
I
MI
T
u
tter
an
ce
s
.
T
h
e
p
r
o
p
o
s
ed
A
SR
s
y
s
te
m
i
n
tr
o
d
u
ce
d
h
er
e
is
ac
tu
all
y
r
ea
lized
w
i
th
M
A
T
L
A
B
o
n
a
m
o
d
er
ate
p
er
s
o
n
al
co
m
p
u
ter
.
RE
F
E
R
E
NC
E
S
[1
]
S
.
S
in
g
h
,
“
F
o
re
n
sic
a
n
d
A
u
to
m
a
t
ic
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
A
p
p
li
e
d
En
g
in
e
e
rin
g
Res
e
a
rc
h
,
S
S
N
0
9
7
3
-
4
5
6
2
Vo
l
u
m
e
8
,
Nu
m
b
e
r
5
,
2
0
1
8
,
p
p
.
2
8
0
4
-
2
8
1
1
,
2
0
1
8
.
[2
]
S
.
S
in
g
h
,
Ev
a
lu
a
ti
o
n
o
f
S
p
a
rsif
ic
a
ti
o
n
a
lg
o
rit
h
m
a
n
d
Its
A
p
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
,
”
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
,
I
S
S
N:
0
9
7
3
-
4
5
6
2
,
Vo
lu
m
e
1
3
,
N
u
m
b
e
r
1
7
,
p
p
.
1
3
0
1
5
-
1
3
0
2
1
,
2
0
1
8
.
[3
]
S
.
S
in
g
h
,
“
S
u
p
p
o
rt
V
e
c
to
r
M
a
c
h
i
n
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
r
n
a
l
o
f
Ap
p
li
e
d
En
g
i
n
e
e
rin
g
Res
e
a
rc
h
,
IS
S
N:
0
9
7
3
-
4
5
6
2
,
V
o
lu
m
e
1
3
,
N
u
m
b
e
r
1
0
,
p
p
.
8
5
6
1
-
8
5
6
7
,
2
0
1
8
.
[4
]
S
.
S
in
g
h
“
T
h
e
Ro
le
o
f
S
p
e
e
c
h
T
e
c
h
n
o
l
o
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
i
n
e
In
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
,
2
0
1
8
.
[5
]
S
.
S
in
g
h
,
A
ss
a
f
M
a
n
so
u
r
H,
A
b
h
a
y
Ku
m
a
r
a
n
d
Niti
n
A
g
ra
w
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
t
io
n
a
l
J
o
u
rn
a
l
o
f
A
p
p
l
i
e
d
En
g
in
e
e
rin
g
Res
e
a
rc
h
,
IS
S
N
0
9
7
3
-
4
5
6
2
Vo
lu
m
e
1
2
,
Nu
m
b
e
r
1
9
,
p
p
.
8
0
2
6
-
8
0
3
3
,
2
0
1
7
.
[6
]
A
.
Eri
k
s
so
n
,
“
T
u
to
rial
o
n
f
o
re
n
sic
sp
e
e
c
h
sc
ien
c
e
,
”
in
Pro
c
.
Eu
ro
p
e
a
n
Co
n
f.
S
p
e
e
c
h
Co
m
mu
n
ica
t
io
n
a
n
d
T
e
c
h
n
o
l
o
g
y
,
L
isb
o
n
,
P
o
rtu
g
a
l,
p
p
.
4
0
-
8
0
,
2
0
0
5
.
[7
]
W
a
ib
e
l,
A
.
,
“
Pro
so
d
y
a
n
d
s
p
e
e
c
h
re
c
o
g
n
it
i
o
n
,
”
S
a
n
M
a
teo
:
M
o
rg
a
n
Ka
u
fm
a
n
n
P
u
b
li
sh
e
rs,
1
9
8
8
.
[8
]
S
h
rib
e
rg
,
E.
,
S
to
lck
e
,
A
.
,
Ha
k
k
a
n
i
-
T
u
r,
D.,
&
T
u
r,
G
,
“
P
ro
so
d
y
-
b
a
s
e
d
a
u
to
m
a
ti
c
se
g
m
e
n
tatio
n
o
f
sp
e
e
c
h
in
to
se
n
te
n
c
e
s a
n
d
to
p
ics
,
”
S
p
e
e
c
h
Co
mm
u
n
ica
ti
o
n
,
3
2
,
p
p
.
1
2
7
-
1
5
4
,
2
0
0
0
.
[9
]
No
o
teb
o
o
m
,
S
,
“
T
h
e
p
ro
so
d
y
o
f
sp
e
e
c
h
:
M
e
lo
d
y
a
n
d
rh
y
th
m
,
”
In
T
h
e
h
a
n
d
b
o
o
k
o
f
p
h
o
n
e
ti
c
sc
ien
c
e
s.
Blac
k
w
e
ll
h
a
n
d
b
o
o
k
s in
li
n
g
u
isti
c
s
M
a
l
d
e
n
,
Blac
k
w
e
ll
P
u
b
li
sh
e
rs,
v
o
l
.
5
,
p
p
.
6
4
0
-
6
7
3
,
1
9
9
7
.
[1
0
]
Ha
r
t,
J.,
C
o
ll
ier,
R
.
,
&
Co
h
e
n
,
A
,
“
A
p
e
rc
e
p
tu
a
l
stu
d
y
o
f
in
to
n
a
ti
o
n
,
”
Ca
mb
rid
g
e
,
UK: Ca
mb
ri
d
g
e
U
n
ive
rs
it
y
Pre
ss
,
1
9
9
0
.
[1
1
]
S
h
r
ib
e
rg
,
E
.
,
&
S
t
o
lck
e
,
A
,
“
D
irec
t
m
o
d
e
li
n
g
o
f
p
ro
so
d
y
:
A
n
o
v
e
rv
ie
w
o
f
a
p
p
li
c
a
ti
o
n
s
in
a
u
t
o
m
a
ti
c
sp
e
e
c
h
p
ro
c
e
ss
in
g
,
”
In
S
p
e
e
c
h
Pr
o
so
d
y
,
Na
ra
,
Ja
p
a
n
,
p
p
.
1
-
8
,
2
0
0
4
.
[1
2
]
Ra
y
m
o
n
d
W
.
M
.
Ng
,
T
a
n
Lee
,
Ch
e
u
n
g
-
Ch
i
L
e
u
n
g
,
Bi
n
M
a
,
H
a
izh
o
u
L
i
,
“
S
p
o
k
e
n
L
a
n
g
u
a
g
e
Re
c
o
g
n
it
io
n
W
it
h
P
r
o
so
d
ic
F
e
a
tu
re
s
,
”
IEE
E
T
r
a
n
sa
c
ti
o
n
s
o
n
A
u
d
io
,
S
p
e
e
c
h
,
a
n
d
L
a
n
g
u
a
g
e
Pr
o
c
e
ss
in
g
,
V
o
l.
21
,
Iss
u
e
.
9
,
pp
-
1
8
4
1
-
1
8
5
3
,
S
e
p
t.
2
0
1
3
.
[1
3
]
El
Ay
a
d
i,
M
.
,
Ka
m
e
l,
M
.
S
.
,
&
Ka
rra
y
,
F
,
“
S
u
rv
e
y
o
n
sp
e
e
c
h
e
m
o
ti
o
n
re
c
o
g
n
it
i
o
n
:
F
e
a
tu
re
s,
c
las
si
f
i
c
a
ti
o
n
sc
h
e
m
e
s,
a
n
d
d
a
ta
b
a
se
s,”
Pa
tt
e
rn
Rec
o
g
n
it
i
o
n
,
v
o
l.
4
4
(
3
)
,
p
p
.
5
7
2
-
5
8
7
,
2
0
1
1
.
[1
4
]
Bu
ss
o
,
C.
,
L
e
e
,
S
.
,
&
Na
ra
y
a
n
a
n
,
S
,
“
A
n
a
l
y
sis o
f
e
m
o
ti
o
n
a
ll
y
sa
li
e
n
t
a
sp
e
c
ts o
f
f
u
n
d
a
m
e
n
tal
f
re
q
u
e
n
c
y
f
o
r
e
m
o
ti
o
n
d
e
tec
ti
o
n
,
”
I
EE
E
T
r
a
n
s
a
c
ti
o
n
s o
n
Au
d
io
,
S
p
e
e
c
h
,
a
n
d
L
a
n
g
u
a
g
e
Pr
o
c
e
ss
in
g
,
v
o
l
.
1
7
(4
)
,
p
p
.
5
8
2
-
5
9
6
,
2
0
0
9
.
[1
5
]
L
u
e
n
g
o
,
I.
,
Na
v
a
s,
E.
,
He
r
n
á
e
z
,
I.
,
&
S
á
n
c
h
e
z
,
“
A
u
to
m
a
ti
c
e
m
o
ti
o
n
re
c
o
g
n
it
io
n
u
sin
g
p
ro
s
o
d
i
c
p
a
ra
m
e
t
e
rs
,”
In
Pro
c
e
e
d
i
n
g
s
o
f
I
n
ter
sp
e
e
c
h
.
p
p
.
4
9
3
-
4
9
6
,
2
0
0
5
.
[1
6
]
Ili
o
u
,
T
.
,
&
A
n
a
g
n
o
sto
p
o
u
lo
s,
C.
-
N,
“
S
tatisti
c
a
l
e
v
a
lu
a
ti
o
n
o
f
sp
e
e
c
h
fe
a
tu
re
s
f
o
r
e
m
o
ti
o
n
re
c
o
g
n
it
io
n
,
”
In
Pro
c
e
e
d
in
g
s
o
f
F
o
u
rth
In
te
rn
a
ti
o
n
a
l
C
o
n
fer
e
n
c
e
o
n
Dig
i
ta
l
T
e
lec
o
mm
u
n
ica
t
io
n
s
(
ICDT’0
9
)
,
v
o
l
.
1
,
p
p
.
1
2
1
-
1
2
6
,
2
0
0
9
.
[1
7
]
L
u
e
n
g
o
,
I.
,
Ev
a
,
N.,
&
He
rn
á
e
z
,
I
,
“
F
e
a
tu
re
a
n
a
ly
sis
a
n
d
e
v
a
lu
a
ti
o
n
f
o
r
a
u
to
m
a
ti
c
e
m
o
ti
o
n
i
d
e
n
ti
f
ica
ti
o
n
in
sp
e
e
c
h
,
”
IEE
E
T
ra
n
sa
c
ti
o
n
s
o
n
M
u
lt
ime
d
i
a
,
v
o
l.
1
2
(6
)
,
4
9
0
-
5
01
,
2
0
1
0
.
[1
8
]
Ha
n
,
K.,
Do
n
g
,
Y.,
&
T
a
sh
e
v
,
I,
“
S
p
e
e
c
h
e
m
o
ti
o
n
re
c
o
g
n
it
io
n
u
sin
g
d
e
e
p
n
e
u
ra
l
n
e
tw
o
rk
a
n
d
e
x
tre
m
e
le
a
rn
in
g
m
a
c
h
in
e
,
”
In
Pro
c
e
e
d
i
n
g
s
o
f
I
n
ter
sp
e
e
c
h
,
p
p
.
2
2
3
-
2
2
7
,
2
0
1
4
.
[1
9
]
Zh
e
n
g
,
N.,
L
e
e
,
T
.
,
&
Ch
in
g
,
P
.
-
C,
“
In
teg
ra
ti
o
n
o
f
c
o
m
p
le
m
e
n
tar
y
a
c
o
u
stic
f
e
a
t
u
re
s
f
o
r
sp
e
a
k
e
r
re
c
o
g
n
it
io
n
,
”
IEE
E
S
i
g
n
a
l
Pro
c
e
ss
in
g
L
e
tt
e
rs
,
v
o
l.
1
4
(3
),
p
p
.
1
8
1
-
8
4
,
2
0
0
7
.
[2
0
]
F
.
Bru
g
n
a
ra
,
D.
F
a
lav
i
g
n
a
,
a
n
d
M
.
Om
o
lo
g
o
,
“
A
u
to
m
a
ti
c
se
g
m
e
n
tatio
n
a
n
d
lab
e
li
n
g
o
f
sp
e
e
c
h
b
a
se
d
o
n
h
id
d
e
n
M
a
rk
o
v
m
o
d
e
ls,”
S
p
e
e
c
h
C
o
mm
u
n
.
,
v
o
l
.
1
2
,
p
p
.
3
5
7
-
3
7
0
,
1
9
9
3
.
[2
1
]
J.
-
P
.
Ho
so
m
,
“
A
u
to
m
a
ti
c
p
h
o
n
e
m
e
a
li
g
n
m
e
n
t
b
a
se
d
o
n
a
c
o
u
stic
-
p
h
o
n
e
ti
c
m
o
d
e
li
n
g
,
”
in
Pro
c
.
7
th
I
n
t
.
Co
n
f
.
S
p
o
k
e
n
L
a
n
g
u
a
g
e
Pr
o
c
e
ss
in
g
,
p
p
.
3
5
7
-
3
6
0
,
2
0
0
2
.
Evaluation Warning : The document was created with Spire.PDF for Python.