I
nte
rna
t
io
na
l J
o
urna
l o
f
E
lect
rica
l a
nd
Co
m
pu
t
er
E
ng
ineering
(
I
J
E
CE
)
Vo
l.
1
6
,
No
.
2
,
A
p
r
il
20
2
6
,
p
p
.
806
~
817
I
SS
N:
2088
-
8
7
0
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijece.
v
1
6
i
2
.
pp
8
0
6
-
8
1
7
806
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ec
e.
ia
esco
r
e.
co
m
Ca
sca
ded
speech
enha
ncemen
t
sy
st
em using
dee
p lea
rning
metho
d
K
a
v
it
ha
A
1
,
M
a
hes
h Cha
nd
r
a
2
,
Vij
a
y
K
u
m
a
r
G
up
t
a
3
1,
2
S
c
h
o
o
l
o
f
E
l
e
c
t
r
o
n
i
c
s
a
n
d
C
o
mm
u
n
i
c
a
t
i
o
n
En
g
i
n
e
e
r
i
n
g
,
R
EV
A
U
n
i
v
e
r
si
t
y
,
B
a
n
g
a
l
o
r
e
,
I
n
d
i
a
3
D
e
p
a
r
t
me
n
t
o
f
El
e
c
t
r
o
n
i
c
s a
n
d
C
o
m
mu
n
i
c
a
t
i
o
n
En
g
i
n
e
e
r
i
n
g
,
G
o
v
e
r
n
m
e
n
t
En
g
i
n
e
e
r
i
n
g
C
o
l
l
e
g
e
W
e
s
t
C
h
a
m
p
a
r
a
n
B
i
h
a
r
,
B
i
h
a
r
,
I
n
d
i
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Oct
1
2
,
2
0
2
5
R
ev
is
ed
Dec
5
,
2
0
2
5
Acc
ep
ted
J
an
1
5
,
2
0
2
6
He
re
,
a
two
-
sta
g
e
c
a
sc
a
d
e
d
n
o
ise
m
in
imiz
a
ti
o
n
fr
o
m
n
o
is
y
sp
e
e
c
h
is
p
ro
p
o
se
d
fo
r
n
o
ise
c
a
n
c
e
ll
a
ti
o
n
fr
o
m
h
i
g
h
ly
c
o
rr
u
p
ted
sp
e
e
c
h
sig
n
a
ls.
In
th
e
first
sta
g
e
,
c
o
rr
u
p
te
d
sp
e
e
c
h
is
p
a
ss
e
d
th
ro
u
g
h
sp
e
e
c
h
e
n
h
a
n
c
e
m
e
n
t
sy
ste
m
b
a
se
d
o
n
wa
v
e
let
d
o
m
a
in
a
d
a
p
t
i
v
e
fil
ter
u
si
n
g
l
e
a
st
m
e
a
n
sq
u
a
re
a
lg
o
rit
h
m
(W
DA
F
-
LM
S
)
a
n
d
p
e
rfo
rm
a
n
c
e
is
e
v
a
lu
a
ted
f
o
r
n
o
isy
si
g
n
a
l
c
o
r
ru
p
ted
b
y
b
a
b
b
le
n
o
ise
,
c
a
r
n
o
ise
a
n
d
m
a
c
h
in
e
g
u
n
n
o
ise
s
.
T
h
e
n
th
is
o
u
tp
u
t
i
s
g
iv
e
n
t
o
se
c
o
n
d
sta
g
e
f
o
r
f
u
rt
h
e
r
imp
r
o
v
e
m
e
n
t.
Th
is
is
f
u
ll
y
c
o
n
n
e
c
ted
d
e
e
p
n
e
u
ra
l
n
e
two
rk
u
sin
g
sto
c
h
a
stic
g
ra
d
i
e
n
t
d
e
sc
e
n
t
with
m
o
m
e
n
tu
m
o
p
ti
m
ize
r
(F
CDN
N
-
S
G
DM)
u
se
d
to
imp
ro
v
e
th
e
q
u
a
l
it
y
o
f
sp
e
e
c
h
si
g
n
a
l.
T
h
e
sy
ste
m
is
tes
ted
fo
r
h
i
g
h
l
y
c
o
rr
u
p
te
d
n
o
i
sy
sp
e
e
c
h
sig
n
a
ls
wh
e
re
n
o
ise
sig
n
a
l
p
o
we
r
lev
e
l
is
e
q
u
a
l
to
o
r
m
o
re
t
h
a
n
c
lea
n
sig
n
a
l
p
o
we
r.
In
p
u
t
sig
n
a
l
-
to
-
n
o
ise
ra
ti
o
(S
NR)
lev
e
l
is
tak
e
n
a
s
0
dB
a
n
d
-
5
t
o
-
13
dB
.
T
h
e
p
r
o
p
o
se
d
sy
ste
m
imp
ro
v
e
d
th
e
q
u
a
li
ty
a
n
d
in
tell
i
g
ib
il
it
y
o
f
sp
e
e
c
h
a
t
a
ll
S
NR
lev
e
ls
fo
r
a
ll
th
re
e
n
o
ise
s.
K
ey
w
o
r
d
s
:
Fu
lly
co
n
n
ec
ted
d
ee
p
n
eu
r
al
n
etwo
r
k
L
ea
s
t m
ea
n
s
q
u
ar
e
alg
o
r
ith
m
SGDM
o
p
tim
izer
Sp
ee
ch
en
h
a
n
ce
m
en
t
W
av
elet
d
o
m
ain
ad
ap
tiv
e
f
ilter
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
:
Vijay
Ku
m
ar
Gu
p
ta
Dep
ar
tm
en
t o
f
E
lectr
o
n
ics an
d
C
o
m
m
u
n
icatio
n
E
n
g
in
ee
r
i
n
g
,
Go
v
er
n
m
e
n
t E
n
g
in
ee
r
in
g
C
o
lleg
e
W
est
C
h
am
p
ar
an
B
ih
ar
R
ailway
Statio
n
,
Op
p
o
s
ite
Ku
m
ar
b
ag
h
,
Ku
m
ar
b
ag
h
,
B
ih
ar
8
4
5
4
5
0
,
I
n
d
ia
E
m
ail:
g
u
p
tav
k
7
6
@
g
m
ail.
co
m
1.
I
NT
RO
D
UCT
I
O
N
I
n
v
o
ice
co
m
m
u
n
icatio
n
s
y
s
tem
s
,
s
p
ee
ch
en
h
an
ce
m
en
t
is
r
eq
u
ir
ed
f
o
r
im
p
r
o
v
in
g
th
e
q
u
ality
an
d
in
tellig
ib
ilit
y
o
f
s
p
ee
c
h
s
ig
n
a
ls
wh
ich
ar
e
ca
p
tu
r
e
d
b
y
t
h
e
s
e
d
ev
ices
u
n
d
er
th
e
p
r
esen
c
e
o
f
b
ac
k
g
r
o
u
n
d
o
r
s
o
m
e
s
p
ec
if
ic
n
o
is
e
[
1
]
alo
n
g
with
clea
n
Hin
d
i
s
p
ee
c
h
d
ata
b
ase
[
2
]
.
T
h
e
g
o
al
o
f
th
ese
s
p
ee
ch
en
h
a
n
ce
m
en
t
s
y
s
tem
s
is
to
m
ak
e
n
o
is
y
s
p
ee
ch
m
o
r
e
clea
r
a
n
d
p
leasan
t
in
lis
ten
in
g
th
r
o
u
g
h
a
u
d
io
d
ev
ice
s
af
ter
ap
p
licatio
n
o
f
n
o
is
e
m
in
im
izatio
n
tech
n
i
q
u
es.
T
h
e
p
er
ce
p
t
u
al
q
u
ality
an
d
in
tellig
ib
ilit
y
o
f
s
p
ee
ch
s
ig
n
al
s
h
o
u
ld
b
e
r
etain
ed
.
On
e
o
f
th
e
im
p
o
r
t
an
t
ap
p
licatio
n
s
is
r
em
o
te
wo
r
k
m
ee
tin
g
s
,
i.e
.
,
telec
o
n
f
er
en
cin
g
o
r
v
i
d
eo
co
n
f
er
e
n
cin
g
wh
er
e
clea
r
v
o
i
ce
f
r
o
m
au
d
i
o
d
ev
ices
is
an
ess
en
tial
r
eq
u
ir
em
en
t
f
o
r
ef
f
ec
t
iv
e
co
m
m
u
n
icatio
n
as we
ll a
s
f
o
r
u
s
er
s
atis
f
ac
tio
n
.
T
r
ad
itio
n
ally
,
ac
c
u
r
ate
n
o
is
e
esti
m
atio
n
-
b
ased
m
eth
o
d
s
s
u
ch
as
s
p
ec
tr
al
s
u
b
tr
ac
tio
n
a
n
d
W
ien
er
f
ilter
in
g
.
wer
e
u
s
ed
f
o
r
d
e
n
o
i
s
in
g
.
T
h
e
ad
ap
tiv
e
n
o
is
e
ca
n
c
ellatio
n
s
y
s
tem
s
u
s
e
p
o
p
u
lar
ad
ap
tiv
e
alg
o
r
ith
m
s
s
u
ch
as
least
m
ea
n
s
q
u
ar
e
(
L
MS
)
,
r
ec
u
r
s
iv
e
least
s
q
u
ar
e
(
R
L
S
)
alg
o
r
ith
m
s
an
d
th
eir
v
ar
ia
n
ts
in
tim
e
d
o
m
ai
n
.
L
MS
alg
o
r
ith
m
is
m
o
s
tly
u
s
ed
d
u
e
to
its
s
im
p
licity
an
d
co
n
v
er
g
en
ce
g
u
ar
an
t
y
in
s
tatio
n
ar
y
en
v
ir
o
n
m
en
ts
,
b
u
t
it
m
ay
tak
e
lo
n
g
tim
e
to
c
o
n
v
er
g
e.
W
h
er
ea
s
R
L
S
alg
o
r
ith
m
s
h
o
ws
Fas
ter
co
n
v
er
g
en
ce
r
a
tes,
b
u
t
it
r
eq
u
ir
es
lar
g
e
co
m
p
u
tatio
n
al
r
eso
u
r
ce
s
,
o
f
ten
it
is
to
o
lar
g
e
f
o
r
r
ea
l
-
tim
e
im
p
lem
en
tatio
n
.
R
L
S
alg
o
r
ith
m
p
er
f
o
r
m
s
b
etter
f
o
r
n
o
n
-
s
tatio
n
ar
y
en
v
i
r
o
n
m
en
ts
.
Af
f
in
e
p
r
o
jectio
n
a
lg
o
r
ith
m
a
n
d
its
v
a
r
ian
ts
ar
e
also
u
s
ed
f
o
r
n
o
is
e
ca
n
ce
llatio
n
s
in
ce
it
h
as
f
ast
co
n
v
er
g
en
ce
lik
e
r
ec
u
r
s
iv
e
R
L
S
an
d
lo
w
c
o
m
p
lex
ity
lik
e
L
MS
alg
o
r
ith
m
.
T
h
e
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:
2088
-
8
7
0
8
C
a
s
ca
d
ed
s
p
ee
ch
en
h
a
n
ce
me
n
t sys
tem
u
s
in
g
d
ee
p
lea
r
n
in
g
meth
o
d
(
K
a
vith
a
A
)
807
tr
an
s
f
o
r
m
d
o
m
ai
n
ad
a
p
tiv
e
al
g
o
r
ith
m
s
o
f
f
er
s
u
p
er
io
r
p
e
r
f
o
r
m
an
ce
o
v
er
co
n
v
en
tio
n
al
ad
ap
tiv
e
alg
o
r
ith
m
s
.
T
h
ese
alg
o
r
ith
m
s
s
h
o
w
lo
wer
co
m
p
u
tatio
n
al
c
o
m
p
lex
ity
an
d
b
etter
co
n
v
er
g
e
n
ce
as
co
m
p
a
r
ed
to
tim
e
d
o
m
ain
alg
o
r
ith
m
s
.
I
m
p
lem
en
tatio
n
o
f
n
o
is
e
ca
n
ce
llatio
n
s
y
s
tem
p
r
o
v
ed
m
o
r
e
ef
f
ec
tiv
e
n
ess
f
o
r
a
p
p
licatio
n
s
h
av
i
n
g
im
p
u
ls
e
r
esp
o
n
s
e
o
f
lo
n
g
d
u
r
a
tio
n
.
Du
e
to
h
i
g
h
ly
co
r
r
elate
d
in
p
u
t
s
ig
n
als,
th
e
co
n
v
er
g
e
n
ce
o
f
L
MS
alg
o
r
ith
m
[
3
]
,
[
4
]
i
n
tim
e
d
o
m
ain
s
ig
n
if
i
ca
n
tly
d
eg
r
ad
es.
B
ec
au
s
e
o
f
o
r
th
o
g
o
n
al
p
r
o
p
er
t
y
,
tr
an
s
f
o
r
m
d
o
m
ain
alg
o
r
ith
m
s
p
r
o
v
id
e
f
aster
co
n
v
er
g
en
ce
s
p
ee
d
with
lo
w
co
m
p
u
tatio
n
a
l
co
m
p
lex
ity
.
T
h
is
is
s
in
ce
tr
an
s
f
o
r
m
d
o
m
ain
ad
ap
tiv
e
f
ilter
s
[
5
]
,
[
6
]
u
s
es
d
ec
o
r
r
elatio
n
p
r
o
p
er
ties
o
f
s
o
m
e
well
-
k
n
o
wn
s
ig
n
al
tr
a
n
s
f
o
r
m
s
,
s
u
c
h
as
th
e
d
is
cr
ete
Fo
u
r
ier
tr
an
s
f
o
r
m
(
D
FT)
,
d
is
cr
ete
co
s
in
e
tr
an
s
f
o
r
m
(
DC
T
)
an
d
wav
elet
tr
an
s
f
o
r
m
(
W
T
)
d
o
m
ain
[
7
]
,
[
8
]
.
W
T
d
o
m
ain
L
MS
New
to
n
ad
ap
tiv
e
f
ilter
in
g
alg
o
r
it
h
m
[
9
]
h
as
p
r
o
v
ed
b
etter
th
an
o
th
er
d
o
m
ain
alg
o
r
ith
m
s
f
o
r
b
o
t
h
f
ir
s
t
o
r
d
e
r
an
d
s
ec
o
n
d
o
r
d
er
au
to
r
e
g
r
e
s
s
iv
e
(
AR
)
p
r
o
ce
s
s
.
I
n
th
is
p
ap
er
th
e
an
aly
s
is
o
f
s
tab
ilit
y
,
m
is
ad
ju
s
tm
en
t,
an
d
co
n
v
er
g
en
ce
p
er
f
o
r
m
an
ce
h
as
b
ee
n
d
o
n
e.
T
h
e
c
o
ef
f
icien
ts
b
elo
n
g
in
g
to
ce
r
tain
s
u
b
-
b
an
d
s
o
f
W
T
b
ased
L
MS
[
1
0
]
ar
e
d
y
n
a
m
ically
s
elec
ted
f
o
r
th
e
u
p
d
ate
b
ased
o
n
lar
g
es
t
d
ec
r
em
en
t
o
f
t
h
e
m
ea
n
-
s
q
u
ar
e
d
ev
iatio
n
.
I
t
r
es
u
lted
in
a
f
ast
co
n
v
er
g
en
ce
s
p
ee
d
an
d
a
lo
w
s
tead
y
-
s
tate
e
r
r
o
r
as
co
m
p
a
r
ed
t
o
s
im
p
le
W
T
b
ased
L
MS.
Oth
er
r
esear
ch
er
s
h
av
e
also
p
r
o
v
ed
th
at
W
T
d
o
m
ain
f
ilter
s
[
1
1
]
,
[
1
2
]
h
a
v
e
also
p
r
o
v
e
d
th
eir
s
u
p
e
r
io
r
ity
f
o
r
a
p
p
licatio
n
s
in
s
p
ee
ch
n
o
is
e
r
ed
u
ctio
n
an
d
ac
o
u
s
tic
ec
h
o
ca
n
ce
l
latio
n
.
I
n
th
e
r
ec
en
t
p
ast
y
ea
r
s
,
m
a
n
y
d
ee
p
lear
n
in
g
n
etwo
r
k
-
b
a
s
ed
s
p
ee
ch
en
h
an
ce
m
e
n
t
s
y
s
tem
s
h
av
e
s
h
o
wn
th
eir
s
u
p
er
io
r
ity
o
v
er
t
h
e
tr
ad
itio
n
al
s
p
ee
ch
d
en
o
is
in
g
m
eth
o
d
s
.
I
n
d
ee
p
lear
n
in
g
m
o
d
els,
d
atab
ases
ar
e
u
s
ed
f
o
r
co
m
p
lex
m
a
p
p
in
g
s
f
r
o
m
n
o
is
e
to
clea
n
s
p
ee
ch
d
i
r
ec
tly
.
I
n
r
ev
iew
p
ap
er
o
n
s
p
ee
ch
en
h
a
n
ce
m
en
t
[
1
3
]
,
r
ev
iewe
r
s
h
av
e
th
e
o
p
in
io
n
th
at
c
o
n
v
o
l
u
tio
n
al
n
eu
r
al
n
etwo
r
k
s
(
C
NN
)
ar
e
b
etter
f
o
r
s
p
ee
ch
en
h
an
ce
m
e
n
t
as
C
NN
is
m
o
r
e
ef
f
ec
tiv
e
in
lear
n
in
g
tem
p
o
r
al
in
f
o
r
m
atio
n
o
f
s
p
ee
ch
s
ig
n
al.
Ma
n
y
d
ee
p
n
eu
r
al
n
etwo
r
k
m
o
d
els
s
u
ch
as
f
u
lly
co
n
n
ec
te
d
n
eu
r
al
n
etwo
r
k
s
[
1
4
]
,
d
ee
p
d
e
n
o
is
in
g
au
to
en
co
d
er
,
C
NN,
L
STM
h
av
e
b
ee
n
ef
f
ec
tiv
el
y
u
s
ed
f
o
r
s
p
ee
ch
d
e
n
o
is
in
g
ev
e
n
in
d
i
v
er
s
e
n
o
is
y
co
n
d
itio
n
s
[
1
5
]
,
[
1
6
]
.
Dee
p
lear
n
in
g
m
eth
o
d
s
in
d
if
f
er
en
t
ap
p
licatio
n
s
an
d
ex
p
er
im
en
tal
s
et
u
p
s
h
av
e
b
ee
n
im
p
lem
en
ted
[
1
7
]
,
[
1
8
]
f
o
r
n
o
r
m
al
s
p
ee
ch
en
h
a
n
ce
m
en
t
a
n
d
s
p
ee
ch
en
h
a
n
ce
m
en
t
i
n
co
c
k
tail
p
ar
ties
.
L
MS
alg
o
r
ith
m
in
wav
elet
d
o
m
ain
[
1
9
]
is
u
s
ed
to
m
in
im
ize
n
o
is
e
f
r
o
m
s
ig
n
als.
B
o
th
tr
ad
itio
n
al
m
eth
o
d
s
an
d
d
ee
p
lear
n
in
g
ap
p
r
o
ac
h
es
[
2
0
]
ar
e
co
m
p
ar
ed
f
o
r
s
p
ee
ch
en
h
a
n
ce
m
en
t a
n
d
d
ee
p
lear
n
in
g
ap
p
r
o
ac
h
es h
av
e
p
r
o
v
e
d
b
etter
th
an
tr
ad
itio
n
al
m
eth
o
d
s
.
Dee
p
lear
n
in
g
ap
p
r
o
ac
h
es
h
av
e
also
p
r
o
v
ed
b
etter
f
o
r
s
p
ee
ch
en
h
an
ce
m
e
n
t
[
2
1
]
,
im
p
r
o
v
em
en
t
in
in
tellig
ib
ilit
y
o
f
s
p
ee
ch
[
2
2
]
a
n
d
in
co
m
b
in
atio
n
with
d
is
cr
ete
tr
an
s
f
o
r
m
s
[
2
3
]
.
C
ascad
ed
C
NN
ar
e
u
s
ed
f
o
r
s
p
ee
ch
em
o
tio
n
r
ec
o
g
n
itio
n
i
n
n
o
is
y
co
n
d
itio
n
s
[
2
4
]
.
T
w
o
s
tag
e
s
p
ee
ch
en
h
a
n
ce
m
en
t
b
y
u
s
in
g
o
p
tim
u
m
v
alu
es
o
f
m
ag
n
itu
d
e
a
n
d
p
h
as
e
h
av
e
b
ee
n
v
er
y
ef
f
ec
tiv
e
in
n
o
is
e
m
in
im
izatio
n
o
r
n
o
is
e
r
ed
u
ctio
n
[
2
5
]
.
A
n
im
p
r
o
v
e
d
R
L
S
alg
o
r
ith
m
is
u
s
ed
to
d
en
o
is
e
elec
tr
o
ca
r
d
io
g
r
am
(
E
C
G
)
s
ig
n
als
f
o
r
f
o
u
r
t
y
p
es
o
f
r
ea
l
n
o
is
es
f
r
o
m
MI
T
-
B
I
T
d
ataset
[
2
6
]
.
Her
e
u
s
e
o
f
a
s
y
s
to
lic
ar
ch
itectu
r
e
en
ab
led
f
aster
p
r
o
ce
s
s
in
g
an
d
b
etter
n
o
is
e
r
ed
u
ctio
n
as
co
m
p
ar
ed
to
R
L
S.
T
h
e
p
r
o
p
o
s
ed
alg
o
r
ith
m
[
2
6
]
o
u
tp
e
r
f
o
r
m
ed
th
e
c
o
n
v
e
n
tio
n
al
R
L
S
alg
o
r
ith
m
with
an
d
with
o
u
t
s
y
s
to
lic
ar
ch
itectu
r
e
in
ter
m
s
o
f
co
n
v
er
g
en
ce
s
p
ee
d
,
s
ig
n
al
-
to
-
n
o
is
e
r
atio
(
SNR
)
,
an
d
m
ea
n
s
q
u
ar
ed
er
r
o
r
(
MSE
)
.
2.
P
RO
P
O
SE
D
M
E
T
H
O
D
T
h
e
s
ch
em
atic
o
f
L
MS
ad
a
p
tiv
e
f
ilter
-
b
ased
s
p
ee
ch
en
h
an
c
em
en
t
s
y
s
tem
is
g
i
v
en
in
Fig
u
r
e
1
.
T
h
e
co
r
r
u
p
te
d
s
p
ee
ch
(
)
co
r
r
u
p
ted
b
y
n
o
is
e
(
)
an
d
r
ef
er
e
n
ce
n
o
is
e
(
)
is
g
iv
en
as
in
p
u
t
to
s
y
s
tem
.
T
h
e
(
)
an
d
(
)
s
h
o
u
ld
b
e
c
o
r
r
elate
d
.
I
n
L
MS
alg
o
r
ith
m
,
th
e
co
s
t f
u
n
ct
io
n
wh
ich
is
least m
ea
n
s
q
u
ar
e
v
alu
e
o
f
er
r
o
r
s
ig
n
al
(
)
is
m
in
im
ized
h
av
i
n
g
m
u
ltip
le
iter
atio
n
s
i.e
.
Sin
ce
(
)
an
d
(
)
ar
e
co
r
r
elate
d
an
d
s
ig
n
al
p
o
wer
is
co
n
s
tan
t,
m
in
im
izin
g
th
e
co
s
t
f
u
n
ctio
n
will
m
in
im
ize
th
e
n
o
is
e
ter
m
in
(
1
)
.
I
n
L
MS
alg
o
r
ith
m
,
th
e
s
te
ep
est
d
escen
t
alg
o
r
ith
m
is
u
s
ed
f
o
r
u
p
d
atin
g
th
e
f
ilter
.
T
h
e
weig
h
t
u
p
d
ate
e
q
u
a
tio
n
f
o
r
L
MS
alg
o
r
ith
m
is
g
i
v
en
b
y
(
2
)
.
C
ost
fun
c
tio
n
=
{
[
2
(
)
]
}
=
{
[
(
(
)
−
(
)
)
2
]
}
=
{
[
(
(
)
+
(
)
−
(
)
)
2
]
}
=
{
[
(
(
)
)
2
]
+
[
(
(
)
−
(
)
)
2
]
+
2
[
(
)
(
(
)
−
(
)
)
]
}
(
1
)
(
+
1
)
=
(
)
+
(
)
(
)
(
2
)
T
h
e
s
tep
s
ize
.
p
la
y
a
n
im
p
o
r
t
an
t
r
o
le
in
co
n
v
er
g
en
ce
o
f
L
MS
alg
o
r
ith
m
.
C
o
n
v
e
r
g
en
ce
t
im
e
will
b
e
h
ig
h
if
s
tep
s
ize
is
s
m
all
o
r
v
ice
v
er
s
a
[
6
]
.
2
.
1
.
Wa
v
elet
do
m
a
in a
da
pti
v
e
f
il
t
er
ba
s
ed
o
n L
M
S a
lg
o
rit
hm
(
WDAF
-
L
M
S
)
I
n
t
h
i
s
a
l
g
o
r
it
h
m
,
t
h
e
i
n
p
u
t
s
i
g
n
a
l
(
)
i
s
a
p
p
l
i
e
d
t
o
N
th
o
r
d
e
r
w
a
v
ele
t
d
o
m
a
i
n
a
d
a
p
t
i
v
e
L
MS
f
i
lt
e
r
.
.
=
[
(
)
,
(
−
1
)
,
(
−
2
)
,
.
.
.
.
.
.
.
.
.
.
(
+
−
1
)
(
3
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
1
6
,
No
.
2
,
Ap
r
il
20
2
6
:
8
0
6
-
817
808
T
h
en
th
is
in
p
u
t sig
n
al
is
tr
an
s
f
o
r
m
ed
to
wav
elet
d
o
m
ai
n
s
ig
n
al
(
,
)
as g
iv
en
in
(
4
)
.
(
,
)
=
{
}
(
4
)
wh
er
e
(
,
)
is
lev
el
3
ap
p
r
o
x
im
a
tio
n
co
ef
f
icien
t
o
f
DW
T
o
f
(
)
.
T
h
e
t
h
r
ee
le
v
el
wav
elet
tr
a
n
s
f
o
r
m
d
ec
o
m
p
o
s
itio
n
o
f
(
)
is
s
h
o
wn
in
Fig
u
r
e
2
[
2
7
]
.
T
h
is
wav
ele
t
d
o
m
ain
s
ig
n
al
(
,
)
is
u
s
ed
in
W
DA
F
s
y
s
tem
s
h
o
wn
in
Fig
u
r
e
3
.
T
h
e
o
u
tp
u
t
(
)
is
g
iv
en
b
y
(
5
)
.
(
)
=
∑
(
,
)
−
1
=
0
(
5
)
T
h
is
o
u
tp
u
t is co
m
p
ar
ed
with
t
h
e
d
esire
d
s
ig
n
al
(
)
an
d
er
r
o
r
s
ig
n
al
(
)
as g
iv
en
b
y
(
6
)
.
(
)
=
(
)
−
(
)
(
6
)
T
h
e
weig
h
t
o
f
W
DAF
is
ch
an
g
ed
ac
co
r
d
i
n
g
t
o
er
r
o
r
s
ig
n
al
(
)
th
r
o
u
g
h
m
u
ltip
le
iter
atio
n
s
.
T
h
e
weig
h
t
u
p
d
ate
eq
u
atio
n
is
g
iv
e
n
b
y
(
2
)
.
Fig
u
r
e
1
.
B
lo
ck
d
iag
r
am
o
f
a
d
ap
tiv
e
f
ilter
-
b
ased
s
p
ee
ch
en
h
an
ce
m
en
t sy
s
tem
Fig
u
r
e
2
.
T
h
e
th
r
ee
le
v
el
wav
e
let
tr
an
s
f
o
r
m
d
ec
o
m
p
o
s
itio
n
o
f
(
)
[
2
7
]
Fig
u
r
e
3
.
A
wav
elet
d
o
m
ain
a
d
ap
tiv
e
f
ilter
[
1
9
]
+
Pr
i
ma
r
y
i
nput
x
(
n
)
=s
(
n
)
+N
(
n
)
T
r
a
n
sv
e
r
s
a
l
f
i
l
t
e
r
w
(
n
)
Σ
R
e
f
e
r
e
n
c
e
n
o
i
s
e
p
(
n
)
y
(
n
)
We
i
g
h
t
c
o
n
t
r
o
l
me
c
h
a
n
i
s
m
w
(
n
)
x(
n
)
x
(
n
-
1)
x
(
n
-
N
+
1
)
z
-
1
z
-
1
z
-
1
y
(
n
)
+
d
(
n
)
e
(
n
)
w
0
w
1
w
N
-
1
Σ
Σ
Dis
cr
et
e
Wa
v
elet
T
ra
ns
f
o
rm
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:
2088
-
8
7
0
8
C
a
s
ca
d
ed
s
p
ee
ch
en
h
a
n
ce
me
n
t sys
tem
u
s
in
g
d
ee
p
lea
r
n
in
g
meth
o
d
(
K
a
vith
a
A
)
809
2
.
2
.
A
s
pee
c
h
e
n
ha
nc
em
e
nt
s
y
s
t
em
b
a
s
e
d
o
n
f
u
lly
co
n
ne
ct
e
d
de
e
p
ne
ura
l
n
et
wo
rk
u
s
i
ng
s
t
o
c
ha
s
t
i
c
g
ra
d
ie
nt
d
e
s
ce
nt
w
it
h m
o
m
e
nt
um
o
ptim
iz
er
A
f
u
lly
co
n
n
ec
ted
d
ee
p
n
eu
r
al
n
etwo
r
k
(
FC
DNN)
is
a
s
u
b
s
et
o
f
ar
tific
ial
n
eu
r
al
n
etwo
r
k
wh
er
e
n
eu
r
o
n
s
o
f
o
n
e
lay
er
ar
e
c
o
n
n
ec
ted
to
ev
er
y
n
e
u
r
o
n
o
f
th
e
n
ex
t
la
y
er
.
T
h
ese
n
etwo
r
k
s
ar
e
wid
ely
u
s
ed
i
n
v
ar
io
u
s
task
s
lik
e
r
eg
r
ess
io
n
,
class
if
icatio
n
,
an
d
p
atter
n
r
ec
o
g
n
itio
n
as
t
h
ey
ca
n
lear
n
co
m
p
lex
m
ap
p
in
g
s
f
r
o
m
in
p
u
ts
an
d
o
u
tp
u
ts
.
T
h
e
c
h
o
ic
e
o
f
ac
tiv
atio
n
f
u
n
ctio
n
is
v
er
y
im
p
o
r
ta
n
t
in
d
eter
m
in
i
n
g
th
e
in
f
o
r
m
atio
n
to
b
e
tr
an
s
m
itted
f
r
o
m
o
n
e
lay
e
r
to
th
e
n
e
x
t.
T
h
u
s
,
it
c
o
n
tr
o
ls
in
f
o
r
m
atio
n
ex
ch
a
n
g
e
an
d
lear
n
in
g
f
r
o
m
d
ata
f
o
r
a
n
eu
r
al
n
etwo
r
k
m
o
d
el.
I
n
th
is
p
ap
er
,
th
e
o
u
tp
u
t
o
f
f
u
lly
co
n
n
ec
ted
lay
er
is
p
ass
ed
th
r
o
u
g
h
a
r
ec
tifie
d
lin
ea
r
u
n
it
(
R
eL
U)
ac
tiv
atio
n
f
u
n
ctio
n
d
ef
in
e
d
b
y
(
7
)
.
A
r
eg
r
ess
io
n
lay
er
is
u
s
ed
in
n
eu
r
al
n
et
wo
r
k
s
f
o
r
r
eg
r
ess
io
n
task
s
.
I
t
co
m
p
u
tes
th
e
h
alf
-
m
e
an
-
s
q
u
ar
ed
-
er
r
o
r
lo
s
s
f
o
r
r
e
g
r
e
s
s
io
n
task
s
.
Fo
r
s
p
ee
ch
e
n
h
an
ce
m
en
t
p
r
o
b
lem
s
,
a
r
eg
r
ess
io
n
lay
er
m
u
s
t
f
o
llo
w
th
e
f
in
al
f
u
lly
co
n
n
ec
ted
lay
er
.
DNN
m
o
d
el
is
tr
ain
ed
u
s
in
g
s
to
ch
asti
c
g
r
ad
ien
t
d
escen
t w
ith
m
o
m
en
tu
m
(
SG
DM
)
o
p
tim
izer
,
p
r
o
g
r
ess
iv
ely
im
p
r
o
v
in
g
its
p
r
ed
ictio
n
o
v
e
r
m
u
ltip
le
ep
o
ch
s
.
(
)
=
ma
x
(
0
,
)
(
7
)
As
s
h
o
wn
in
Fig
u
r
e
4
,
f
u
lly
c
o
n
n
ec
ted
d
e
ep
n
eu
r
al
n
etwo
r
k
u
s
in
g
s
to
ch
asti
c
g
r
ad
ien
t
d
e
s
ce
n
t
with
m
o
m
en
tu
m
o
p
tim
izer
(
FC
DNN
-
SGDM
)
is
f
ir
s
t
tr
ain
ed
f
r
o
m
m
ag
n
itu
d
e
s
p
ec
tr
u
m
o
f
n
o
is
y
s
p
ee
ch
a
n
d
clea
n
s
p
ee
ch
to
d
ev
el
o
p
a
n
o
n
lin
e
ar
r
elatio
n
b
etwe
en
t
h
e
n
o
is
y
s
p
ec
tr
u
m
an
d
clea
n
s
p
ec
tr
u
m
.
T
h
e
s
h
o
r
t
tim
e
Fo
u
r
ier
tr
an
s
f
o
r
m
(
STFT
)
is
a
p
p
lied
to
b
o
th
s
p
ee
c
h
es
to
o
b
t
ain
its
m
ag
n
itu
d
e
s
p
ec
tr
u
m
an
d
p
h
ase
s
p
ec
tr
u
m
.
T
h
e
m
ag
n
itu
d
e
s
p
ec
tr
u
m
o
f
n
o
is
y
s
p
ee
ch
is
u
s
ed
as
f
in
al
p
r
ed
icto
r
.
W
h
er
ea
s
th
e
m
ag
n
itu
d
e
s
p
ec
tr
u
m
o
f
clea
n
s
p
ee
ch
s
ig
n
al
is
u
s
ed
as
f
in
al
tar
g
et
.
Fo
r
th
e
wh
o
le
s
p
ee
c
h
d
atab
ase
p
r
ed
icto
r
s
an
d
tar
g
et
s
ar
e
ca
lcu
lated
an
d
u
s
ed
to
tr
ain
DNN
m
o
d
els.
I
n
th
e
test
in
g
s
tag
e
p
er
f
o
r
m
an
ce
is
test
ed
u
s
in
g
th
e
n
o
is
y
test
d
atab
ase.
I
n
v
alid
atio
n
s
tag
e,
th
e
u
n
k
n
o
w
n
n
o
is
y
s
p
ee
ch
s
ig
n
al
is
ap
p
lied
to
tr
ain
ed
FC
C
DNN
m
o
d
el
.
First
m
ag
n
itu
d
e
s
p
ec
tr
u
m
an
d
p
h
ase
s
p
ec
tr
u
m
ar
e
s
ep
ar
ated
,
an
d
n
o
is
y
s
p
ee
ch
s
ig
n
al
is
ca
lcu
lated
b
y
STFT
.
On
ly
I
ts
m
ag
n
itu
d
e
s
p
ec
tr
u
m
is
g
iv
en
to
tr
ain
ed
DNN
m
o
d
el
.
T
h
e
tr
ain
ed
FC
C
DN
N
m
o
d
el
p
r
o
v
i
d
es
th
e
m
ag
n
itu
d
e
s
p
ec
tr
u
m
o
f
r
ec
o
v
er
ed
s
p
ee
ch
s
ig
n
al.
T
h
en
in
v
er
s
e
STFT
is
ap
p
lied
to
th
is
m
ag
n
itu
d
e
s
p
ec
tr
u
m
an
d
t
h
e
p
h
ase
s
p
ec
tr
u
m
o
f
n
o
is
y
s
p
ee
ch
s
ig
n
al
to
r
ec
o
v
e
r
s
p
ee
ch
s
ig
n
al
as
s
h
o
wn
in
Fig
u
r
e
5.
Fig
u
r
e
4
.
T
r
ain
in
g
s
tag
e
o
f
FC
DNN
-
SGDM
m
o
d
el
Fig
u
r
e
5
.
T
esti
n
g
a
n
d
v
alid
atio
n
s
tag
e
o
f
t
r
ain
ed
FC
DNN
-
SGDM
m
o
d
el
F
C
D
N
N
-
S
G
D
M
m
o
d
e
l
f
o
r
t
r
a
i
n
i
n
g
C
l
e
a
n
sp
e
e
c
h
si
g
n
a
l
W
i
n
d
o
w
i
n
g
a
n
d
o
v
e
r
l
a
p
p
i
n
g
D
i
scret
e
t
r
an
sf
o
r
m
M
a
g
n
i
t
u
d
e
sp
e
c
t
r
u
m
N
o
i
s
y
sp
e
e
c
h
si
g
n
a
l
W
i
n
d
o
w
i
n
g
a
n
d
o
v
e
r
l
a
p
p
i
n
g
D
i
scret
e
t
r
an
sf
o
r
m
M
a
g
n
i
t
u
d
e
sp
e
c
t
r
u
m
F
C
D
N
N
-
S
G
D
M
mo
d
e
l
f
o
r
t
r
a
i
n
i
n
g
a
n
d
t
e
st
i
n
g
N
o
i
s
y
sp
e
e
c
h
si
g
n
a
l
W
i
n
d
o
w
i
n
g
a
n
d
o
v
e
r
l
a
p
p
i
n
g
D
i
scret
e
t
r
a
n
sf
o
r
m
M
a
g
n
i
t
u
d
e
sp
e
c
t
r
u
m
P
h
a
se
sp
e
c
t
r
u
m
R
e
c
o
v
e
r
e
d
sp
e
e
c
h
si
g
n
a
l
M
a
g
n
i
t
u
d
e
sp
e
c
t
r
u
m
I
n
v
e
r
se
sh
o
r
t
t
i
me
F
o
u
r
i
e
r
t
r
a
n
sf
o
r
m
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
1
6
,
No
.
2
,
Ap
r
il
20
2
6
:
8
0
6
-
817
810
3.
E
XP
E
R
I
M
E
N
T
A
L
SE
T
UP
I
n
th
e
ex
p
er
im
en
tal
s
etu
p
s
h
o
wn
in
Fig
u
r
e
6
,
a
two
-
s
tag
e
c
ascad
ed
s
p
ee
ch
en
h
a
n
ce
m
en
t
m
eth
o
d
is
p
r
o
p
o
s
ed
wh
e
r
e
f
i
r
s
t
s
tag
e
is
W
DAF
-
L
MS
an
d
s
ec
o
n
d
s
ta
g
e
is
FC
DNN
-
SGDM.
I
n
th
e
s
ec
o
n
d
s
tag
e,
t
h
e
r
ec
o
v
er
e
d
s
ig
n
al
is
ag
ain
p
ass
ed
th
r
o
u
g
h
tr
ain
ed
FC
DNN
-
SGDM
b
ased
s
p
ee
ch
en
h
an
ce
m
en
t
s
y
s
tem
to
f
u
r
th
er
r
em
o
v
e
th
e
r
esid
u
al
b
ac
k
g
r
o
u
n
d
n
o
is
e.
Fo
r
e
v
alu
at
io
n
o
f
a
s
p
ee
ch
e
n
h
an
ce
m
e
n
t
s
y
s
tem
,
a
p
r
o
p
er
d
atab
ase
is
r
eq
u
ir
ed
.
Her
e
clea
n
s
en
ten
ce
d
atab
ase
is
tak
en
f
r
o
m
Hin
d
i
s
p
ee
ch
d
atab
ase
[
2
]
.
T
h
e
n
o
is
y
v
e
r
s
io
n
o
f
clea
n
s
en
ten
ce
s
is
p
r
e
p
ar
ed
b
y
ad
d
in
g
b
ab
b
le
n
o
is
e,
ca
r
n
o
is
e
an
d
m
ac
h
in
e
g
u
n
n
o
is
es
f
r
o
m
NOI
SEX
-
92
d
atab
ase
[
1
]
to
th
is
clea
n
s
en
ten
ce
at
0
dB
d
o
wn
to
–
13
dB
.
Fig
u
r
e
6
.
A
two
-
s
tag
e
s
p
ee
c
h
en
h
an
ce
m
e
n
t sy
s
tem
I
n
W
DAF
-
L
MS
b
ased
s
p
ee
ch
en
h
an
ce
m
e
n
t
s
y
s
tem
,
th
e
f
ilter
o
r
d
e
r
is
tak
en
as
2
0
a
n
d
s
tep
s
ize
µ
is
tak
en
as
0
.
0
0
8
.
Her
e
‘
DB
2
’
w
av
elet
is
u
s
ed
f
o
r
DW
T
f
o
r
r
e
f
er
en
ce
n
o
is
e
s
ig
n
al
tak
en
as
i
n
p
u
t
to
tr
an
s
v
er
s
al
f
ilter
Fo
r
FC
C
D
NN
-
SGDM,
t
o
tal
4
1
9
s
p
ee
ch
s
am
p
les
ar
e
t
ak
en
f
o
r
tr
ain
in
g
,
test
in
g
an
d
v
alid
atio
n
.
Sp
ee
ch
d
atab
ase
[
2
]
o
f
2
1
s
p
ea
k
e
r
s
is
tak
en
f
o
r
tr
ai
n
in
g
,
test
in
g
an
d
v
alid
atio
n
.
W
h
er
e
2
0
s
en
ten
ce
s
ar
e
tak
en
f
r
o
m
ea
ch
s
p
ea
k
er
.
T
r
ain
i
n
g
d
atase
t
co
n
s
is
ts
o
f
to
tal
3
1
4
Hin
d
i
s
p
ee
ch
s
am
p
les
an
d
test
in
g
d
ataset
co
n
s
is
t
s
o
f
8
4
Hin
d
i
s
p
ee
ch
s
am
p
les.
7
5
%
a
n
d
2
0
%
o
f
th
e
d
atab
ase
a
r
e
u
s
ed
f
o
r
tr
ain
i
n
g
a
n
d
test
in
g
r
es
p
ec
tiv
ely
.
Als
o
,
2
1
s
am
p
les
i.e
.
5
%
s
am
p
les
a
r
e
u
s
ed
f
o
r
v
alid
atio
n
o
f
th
e
r
esu
lt
s
.
Fo
r
STFT
,
“
Han
n
in
g
win
d
o
w”
o
f
le
n
g
th
6
4
is
u
s
ed
f
o
r
f
r
am
in
g
.
Fiv
e
lay
er
s
ar
e
u
s
ed
f
o
r
tr
ain
in
g
th
e
m
o
d
el.
SGDM
is
u
s
ed
f
o
r
T
r
ain
in
g
o
p
tio
n
s
with
m
o
m
en
tu
m
eq
u
als to
0
.
9
0
0
0
,
l
ea
r
n
r
ate
d
r
o
p
f
ac
to
r
is
.
0
0
0
1
,
m
ax
im
u
m
e
p
o
ch
s
ar
e
2
0
,
m
in
i
b
atch
s
ize
is
6
4
.
T
h
e
p
er
f
o
r
m
a
n
ce
p
a
r
am
eter
s
u
s
ed
in
e
x
p
er
im
e
n
tal
s
etu
p
f
o
r
ev
alu
atio
n
o
f
s
y
s
tem
ar
e
SNR
im
p
r
o
v
em
e
n
t,
p
e
r
ce
p
tu
al
e
v
alu
atio
n
o
f
s
p
ee
ch
q
u
ality
(
PES
Q)
an
d
s
h
o
r
t
-
tim
e
o
b
jectiv
e
in
tellig
ib
ilit
y
(
STOI
)
.
PESQ
v
alu
e
r
an
g
es
f
r
o
m
-
0
.
5
to
4
.
5
an
d
th
e
h
ig
h
e
r
PESQ
m
ea
n
s
b
etter
p
er
ce
p
tu
al
q
u
ali
ty
.
STOI
v
alu
e
is
a
s
ca
lar
q
u
an
tity
,
a
n
d
r
an
g
es
f
r
o
m
-
1
to
1
.
I
t
m
ea
s
u
r
es
th
e
in
tellig
ib
ilit
y
o
f
th
e
r
ec
o
v
er
e
d
s
p
ee
ch
s
ig
n
al
b
y
co
m
p
ar
in
g
it
with
th
e
clea
n
s
p
ee
ch
s
ig
n
al.
A
h
ig
h
er
v
alu
e
f
o
r
STOI
co
r
r
esp
o
n
d
s
to
a
h
ig
h
er
in
tellig
ib
ilit
y
o
f
s
p
ee
ch
s
ig
n
al.
4.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
I
t
is
o
b
s
er
v
ed
f
r
o
m
T
a
b
le
1
t
o
T
ab
le
3
th
at
t
h
e
p
r
o
p
o
s
ed
s
y
s
tem
p
r
o
v
id
es
im
p
r
o
v
e
m
en
t
in
s
p
ee
ch
q
u
ality
an
d
i
n
tellig
ib
ly
in
ter
m
s
o
f
SNR
im
p
r
o
v
em
en
t,
PE
SQ
an
d
STOI
at
all
in
p
u
t
SNR
lev
el
f
o
r
all
th
r
ee
n
o
is
es.
I
n
th
is
ex
p
er
im
en
tal
s
etu
p
,
th
e
co
r
r
u
p
te
d
s
ig
n
al
was
te
s
ted
at
in
p
u
t
SNR
l
ev
els
r
an
g
in
g
f
r
o
m
0
dB
d
o
wn
t
o
–
13
dB
.
E
v
e
n
u
p
to
-
13
dB
in
p
u
t SNR
lev
el,
th
e
s
y
s
tem
p
r
o
v
i
d
es su
b
s
eq
u
en
t im
p
r
o
v
em
e
n
ts
.
T
ab
le
1
,
Fig
u
r
e
7
,
an
d
Fig
u
r
e
8
s
h
o
w
th
e
en
h
an
ce
m
en
t
in
n
o
is
y
s
p
ee
ch
s
ig
n
al
co
r
r
u
p
ted
b
y
b
ab
b
l
e
n
o
is
e
at
0
d
B
,
-
5
d
B
,
-
6
d
B
,
-
7
d
B
,
-
8
d
B
,
-
9
d
B
,
-
1
0
d
B
,
-
1
1
d
B
,
-
1
2
d
B
,
-
1
3
d
B
in
p
u
t
SN
R
lev
els
in
ter
m
s
o
f
SNR
,
PESQ
an
d
STOI
,
wh
en
it
is
p
ass
ed
th
r
o
u
g
h
th
e
s
y
s
tem
,
with
th
e
s
p
ec
if
ic
m
etr
ics
d
etailed
in
Fig
u
r
e
7
(
a)
th
r
o
u
g
h
Fig
u
r
e
7
(
d
)
.
T
ab
le
1
s
h
o
ws
th
e
co
m
p
ar
is
o
n
s
o
f
o
u
tp
u
t
at
b
o
t
h
s
tag
es.
W
h
er
e
s
tag
e
-
1
is
W
DAF
-
L
MS
an
d
s
tag
e
-
2
is
F
C
C
D
NN
-
SGDM.
T
h
e
W
DAF
-
L
MS
ac
h
iev
ed
a
m
ax
im
u
m
SNR
im
p
r
o
v
em
en
t o
f
1
5
.
2
3
8
d
B
at
–
1
0
d
B
in
p
u
t
SNR
wh
ile
th
e
s
u
b
s
eq
u
en
t
FC
C
DNN
-
SGDM
s
tag
e
f
u
r
th
er
en
h
a
n
ce
d
th
e
p
er
f
o
r
m
an
ce
,
ac
h
iev
in
g
th
e
o
v
er
all
m
ax
im
u
m
SNR
im
p
r
o
v
em
en
t
o
f
1
7
.
0
6
1
d
B
at
–
8
d
B
in
p
u
t
SNR
.
E
v
en
a
t
-
1
3
d
B
in
p
u
t
SNR
lev
el
,
th
e
p
r
o
p
o
s
ed
alg
o
r
ith
m
p
r
o
v
id
es
s
u
b
s
eq
u
en
t
im
p
r
o
v
em
e
n
t
with
im
p
r
o
v
em
en
ts
in
F
r
e
q
u
e
n
c
y
D
o
mai
n
A
d
a
p
t
i
v
e
f
i
l
t
e
r
u
s
i
n
g
LM
S
a
l
g
o
r
i
t
h
m
(WDAF
-
LM
S
)
N
o
i
s
y
S
i
g
n
a
l
R
e
f
e
r
e
n
c
e
N
o
i
se
F
C
C
N
D
D
-
S
G
D
M
S
t
a
g
e
1
S
t
a
g
e
2
Re
c
o
v
er
e
d
S
i
g
n
a
l
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:
2088
-
8
7
0
8
C
a
s
ca
d
ed
s
p
ee
ch
en
h
a
n
ce
me
n
t sys
tem
u
s
in
g
d
ee
p
lea
r
n
in
g
meth
o
d
(
K
a
vith
a
A
)
811
PESQ
an
d
STOI
v
alu
es
wh
ich
s
h
o
ws
th
at
at
-
1
3
d
B
in
p
u
t
SNR
lev
el,
s
p
ee
ch
q
u
ality
an
d
in
tellig
ib
ilit
y
i
s
r
etain
ed
in
d
en
o
is
ed
s
ig
n
al.
Fig
u
r
e
8
illu
s
tr
ates
th
e
wav
ef
o
r
m
co
m
p
ar
is
o
n
s
p
ec
if
ically
at
-
5
d
B
in
p
u
t
SNR
,
d
is
p
lay
in
g
t
h
e
clea
n
s
p
ee
ch
in
Fig
u
r
e
8
(
a
)
,
th
e
n
o
is
y
s
p
ee
ch
in
Fig
u
r
e
8
(
b
)
,
t
h
e
d
en
o
is
ed
s
ig
n
al
af
ter
s
tag
e
-
1
in
Fig
u
r
e
8
(
c)
,
an
d
th
e
d
e
n
o
i
s
ed
s
ig
n
al
af
ter
s
tag
e
-
2
in
Fi
g
u
r
e
8
(
d
)
.
As
o
b
s
er
v
ed
in
th
ese
wav
ef
o
r
m
s
,
th
e
s
ig
n
al
r
ec
o
v
er
e
d
af
ter
s
tag
e
-
2
is
clo
s
er
to
o
r
ig
in
al
s
ig
n
al
as c
o
m
p
ar
ed
to
th
e
s
ig
n
al
r
ec
o
v
e
r
ed
af
ter
s
tag
e
-
1.
T
ab
le
2
,
Fig
u
r
e
9
,
an
d
Fig
u
r
e
1
0
s
h
o
w
th
e
en
h
an
ce
m
en
t
i
n
n
o
is
y
s
p
ee
ch
s
ig
n
al
co
r
r
u
p
t
ed
b
y
ca
r
n
o
is
e
at
0
d
B
,
-
5
d
B
,
-
6
d
B
,
-
7
d
B
,
-
8
d
B
,
-
9
d
B
,
-
1
0
d
B
,
-
1
1
d
B
,
-
1
2
d
B
,
-
1
3
d
B
in
p
u
t
SN
R
lev
els
in
ter
m
s
o
f
SNR
,
PESQ
an
d
STOI
,
wh
en
it
is
p
ass
ed
th
r
o
u
g
h
th
e
s
y
s
tem
,
with
th
e
s
p
ec
if
ic
m
etr
ics
d
etailed
in
Fig
u
r
e
9
(
a)
th
r
o
u
g
h
Fig
u
r
e
9
(
d
)
.
T
a
b
le
2
s
h
o
ws
th
e
c
o
m
p
a
r
is
o
n
s
o
f
o
u
tp
u
t
at
b
o
t
h
s
tag
es.
T
h
e
W
DAF
-
L
M
S
s
h
o
ws
m
ax
im
u
m
SNR
im
p
r
o
v
em
en
t
o
f
2
1
.
2
4
0
6
d
B
at
-
1
3
d
B
in
p
u
t
SNR
lev
el
wh
ile
th
e
FC
C
DN
N
-
SGDM
f
u
r
th
er
im
p
r
o
v
es
in
SNR
to
2
2
.
1
7
0
5
d
B
at
-
1
3
d
B
in
p
u
t
SNR
lev
el.
I
t
is
o
b
s
er
v
ed
th
at
at
-
1
3
d
B
in
p
u
t
SNR
lev
el,
th
e
p
r
o
p
o
s
ed
alg
o
r
ith
m
p
r
o
v
id
es
s
u
b
s
eq
u
en
t
im
p
r
o
v
em
en
t
with
im
p
r
o
v
em
en
ts
in
PESQ
an
d
STO
I
v
alu
es.
W
h
ich
s
h
o
ws
th
at
at
-
1
3
d
B
in
p
u
t
SNR
lev
el
s
p
ee
ch
q
u
ality
an
d
in
tellig
ib
ilit
y
is
r
etain
ed
in
d
en
o
is
ed
s
ig
n
al.
Fig
u
r
e
1
0
illu
s
tr
ates
th
e
wav
ef
o
r
m
co
m
p
ar
is
o
n
s
p
ec
if
ically
at
-
5
d
B
in
p
u
t
SNR
,
d
i
s
p
lay
in
g
th
e
clea
n
s
p
ee
ch
in
Fig
u
r
e
1
0
(
a)
,
th
e
n
o
is
y
s
p
ee
ch
in
Fig
u
r
e
1
0
(
b
)
,
t
h
e
d
en
o
is
ed
s
ig
n
al
af
ter
s
tag
e
-
1
in
Fig
u
r
e
1
0
(
c)
,
an
d
th
e
d
en
o
is
ed
s
ig
n
al
af
ter
s
tag
e
-
2
in
Fig
u
r
e
1
0
(
d
)
.
As
o
b
s
er
v
ed
in
th
ese
wav
ef
o
r
m
s
,
th
e
s
ig
n
al
r
ec
o
v
er
e
d
af
ter
s
tag
e
-
2
is
clo
s
er
to
o
r
ig
in
al
s
ig
n
al
as c
o
m
p
ar
ed
t
o
th
e
s
i
g
n
al
r
ec
o
v
er
ed
a
f
ter
s
tag
e
-
1.
T
ab
le
1
.
Per
f
o
r
m
an
ce
co
m
p
a
r
is
o
n
o
f
s
p
ee
ch
en
h
a
n
ce
m
en
t sy
s
tem
at
s
tag
e
-
1
an
d
s
tag
e
-
2
(
f
i
n
al
s
tag
e)
f
o
r
b
a
b
b
le
n
o
is
e
I
n
p
u
t
S
N
R
S
N
R
i
m
p
r
o
v
e
me
n
t
P
ESQ
S
TO
I
W
D
A
F
-
LM
S
F
C
C
D
N
N
-
S
G
D
M
W
D
A
F
-
LM
S
F
C
C
D
N
N
-
S
G
D
M
W
D
A
F
-
LM
S
F
C
C
D
N
N
-
S
G
D
M
0
dB
1
3
.
2
4
5
7
1
3
.
7
4
2
1
2
.
6
7
1
6
2
.
9
9
6
3
0
.
9
6
3
4
0
.
9
6
7
8
-
5
dB
1
4
.
7
4
7
8
1
6
.
3
6
4
1
2
.
3
8
4
7
2
.
7
7
8
4
0
.
9
4
8
6
0
.
9
5
0
8
-
6
dB
1
4
.
9
1
8
4
1
6
.
6
9
3
5
2
.
3
2
7
1
2
.
7
0
0
2
0
.
9
4
0
9
0
.
9
4
5
8
-
7
dB
1
5
.
0
5
5
2
1
6
.
9
2
8
3
2
.
2
6
9
1
2
.
6
2
0
.
9
3
2
0
.
9
3
9
9
-
8
dB
1
5
.
1
5
5
2
1
7
.
0
6
1
2
.
2
1
3
6
2
.
5
1
6
1
0
.
9
2
1
4
0
.
9
3
2
9
-
9
dB
1
5
.
2
2
0
1
1
7
.
0
4
8
8
2
.
1
5
7
5
2
.
4
0
1
7
0
.
9
0
9
0
.
9
2
3
3
-
10
dB
1
5
.
2
3
8
1
6
.
8
8
3
1
2
.
1
0
3
2
2
.
2
8
3
8
0
.
8
9
4
7
0
.
9
1
0
8
-
11
dB
1
5
.
1
7
0
5
1
6
.
5
0
1
7
2
.
0
4
7
5
2
.
1
5
7
3
0
.
8
7
7
4
0
.
8
9
3
-
12
dB
1
4
.
9
6
5
4
1
5
.
7
8
8
5
1
.
9
8
2
3
2
.
0
1
6
9
0
.
8
5
4
2
0
.
8
6
3
-
13
dB
1
4
.
5
8
5
3
1
4
.
8
6
2
1
1
.
9
1
2
4
1
.
8
6
8
0
.
8
2
3
4
0
.
8
2
2
3
T
ab
le
2
.
Per
f
o
r
m
an
ce
co
m
p
a
r
is
o
n
o
f
s
p
ee
ch
en
h
a
n
ce
m
en
t sy
s
tem
at
s
tag
e
-
1
an
d
s
tag
e
-
2
(
f
i
n
al
s
tag
e)
f
o
r
ca
r
n
o
is
e
I
N
P
U
T
S
N
R
S
N
R
i
m
p
r
o
v
e
me
n
t
P
ESQ
S
TO
I
W
D
A
F
-
LM
S
F
C
C
D
N
N
-
S
G
D
M
W
D
A
F
-
LM
S
F
C
C
D
N
N
-
S
G
D
M
W
D
A
F
-
LM
S
F
C
C
D
N
N
-
S
G
D
M
0
d
B
9
.
0
5
9
0
1
0
.
1
6
9
3
3
.
4
5
5
8
3
.
4
6
5
4
0
.
9
8
4
4
0
.
9
8
7
6
-
5
d
B
1
3
.
9
9
7
2
1
5
.
0
9
2
6
3
.
4
0
7
5
3
.
4
1
8
0
.
9
8
4
0
.
9
8
7
5
-
6
d
B
1
4
.
9
7
4
1
1
6
.
0
6
5
9
3
.
3
9
0
6
3
.
4
1
0
9
0
.
9
8
3
8
0
.
9
8
7
5
-
7
d
B
1
5
.
9
4
5
4
1
7
.
0
3
0
4
3
.
3
7
1
6
3
.
4
0
2
3
0
.
9
8
3
7
0
.
9
8
7
4
-
8
d
B
1
6
.
9
0
9
5
1
7
.
9
9
6
6
3
.
3
4
8
3
.
3
8
9
1
0
.
9
8
3
5
0
.
9
8
7
3
-
9
d
B
1
7
.
8
6
4
1
1
8
.
9
2
9
7
3
.
3
2
2
1
3
.
3
7
1
0
.
9
8
3
2
0
.
9
8
7
2
-
1
0
d
B
1
8
.
8
0
2
3
1
9
.
8
6
5
9
3
.
2
8
1
7
3
.
3
4
6
3
0
.
9
8
2
8
0
.
9
8
7
-
1
1
d
B
1
9
.
7
0
8
8
2
0
.
7
6
2
5
3
.
2
2
6
1
3
.
3
0
6
9
0
.
9
8
2
2
0
.
9
8
6
7
-
1
2
d
B
2
0
.
5
4
0
9
2
1
.
5
6
2
7
3
.
1
2
1
2
3
.
2
0
9
8
0
.
9
8
0
7
0
.
9
8
5
6
-
1
3
d
B
2
1
.
2
4
0
6
2
2
.
1
7
0
5
2
.
9
6
9
4
3
.
0
2
1
9
0
.
9
7
7
4
0
.
9
8
2
4
T
ab
le
3
.
Per
f
o
r
m
an
ce
co
m
p
a
r
is
o
n
o
f
s
p
ee
ch
en
h
a
n
ce
m
en
t sy
s
tem
at
s
tag
e
-
1
an
d
s
tag
e
-
2
(
f
i
n
al
s
tag
e)
f
o
r
m
ac
h
in
e
g
u
n
n
o
is
e
I
N
P
U
T
S
N
R
S
N
R
i
m
p
r
o
v
e
me
n
t
P
ESQ
S
TO
I
W
D
A
F
-
LM
S
F
C
C
D
N
N
-
S
G
D
M
W
D
A
F
-
LM
S
F
C
C
D
N
N
-
S
G
D
M
W
D
A
F
-
LM
S
F
C
C
D
N
N
-
S
G
D
M
0
d
B
1
1
.
1
0
8
4
1
7
.
1
9
1
8
3
.
1
0
3
1
3
.
1
6
8
7
0
.
9
8
7
9
0
.
9
8
8
4
-
5
d
B
1
1
.
1
8
8
1
1
8
.
2
5
3
3
2
.
6
9
2
2
.
7
3
9
3
0
.
9
6
4
9
0
.
9
7
2
7
-
6
d
B
1
1
.
1
3
1
2
1
7
.
9
7
1
8
2
.
5
2
2
8
2
.
6
0
4
6
0
.
9
5
3
3
0
.
9
6
5
3
-
7
d
B
1
0
.
9
9
6
5
1
7
.
7
7
2
1
2
.
3
7
8
5
2
.
5
0
3
1
0
.
9
3
9
7
0
.
9
5
7
9
-
8
d
B
1
0
.
8
0
8
6
1
7
.
5
8
4
1
2
.
2
3
7
2
.
2
9
6
8
0
.
9
2
2
4
0
.
9
4
9
4
-
9
d
B
1
0
.
5
9
1
8
1
7
.
4
1
9
9
2
.
1
2
8
2
.
3
0
8
5
0
.
9
0
2
2
0
.
9
3
9
5
-
1
0
d
B
1
0
.
3
5
0
5
1
7
.
3
5
1
7
2
.
0
1
9
8
2
.
2
2
7
7
0
.
8
7
7
5
0
.
9
2
7
7
-
1
1
d
B
1
0
.
1
1
5
3
1
7
.
3
1
9
8
1
.
9
4
2
1
2
.
1
5
0
2
0
.
8
5
1
7
0
.
9
1
5
-
1
2
d
B
9
.
9
2
8
3
1
7
.
3
1
4
8
1
.
8
7
1
6
2
.
0
8
1
5
0
.
8
2
6
5
0
.
9
0
2
-
1
3
d
B
9
.
7
9
3
4
1
7
.
2
8
6
7
1
.
8
2
4
1
2
.
0
0
1
9
0
.
8
0
2
8
0
.
8
8
7
3
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
1
6
,
No
.
2
,
Ap
r
il
20
2
6
:
8
0
6
-
817
812
(
a)
(
b
)
(
c)
(
d
)
Fig
u
r
e
7
.
Per
f
o
r
m
an
c
e
co
m
p
ar
is
o
n
o
f
s
p
ee
ch
e
n
h
an
ce
m
en
t s
y
s
tem
f
o
r
b
a
b
b
le
n
o
is
e
at
d
if
f
e
r
en
t in
p
u
t SNR
lev
els:
(
a)
o
u
tp
u
t SNR
f
o
r
b
ab
b
le
n
o
is
e
at
all
in
p
u
t SNR
lev
e
l,
(
b
)
SNR
im
p
r
o
v
e
m
en
t f
o
r
F
C
C
D
NN
-
SGDM
as c
o
m
p
ar
ed
t
o
W
DAF
-
L
MS,
(
c)
PESQ
im
p
r
o
v
em
e
n
t f
o
r
FC
C
DNN
-
SG
DM
as c
o
m
p
ar
ed
t
o
W
DAF
-
L
MS,
an
d
(
d
)
STOI
im
p
r
o
v
em
e
n
t f
o
r
FC
C
DN
N
-
SGDM
a
s
co
m
p
ar
ed
to
W
DAF
-
L
MS
f
o
r
b
ab
b
le
n
o
is
e
at
all
I
n
p
u
t
SNR
lev
el
(
a)
(
b
)
(
c)
(
d
)
Fig
u
r
e
8
.
W
av
ef
o
r
m
co
m
p
ar
is
o
n
o
f
s
p
ee
ch
s
ig
n
als co
r
r
u
p
ted
b
y
b
a
b
b
le
n
o
is
e:
(
a)
clea
n
s
p
e
ec
h
,
(
b
)
n
o
is
y
s
p
ee
ch
,
(
c)
d
en
o
is
ed
s
p
ee
ch
s
ig
n
al
af
ter
s
tag
e
-
1
,
an
d
(
d
)
d
en
o
is
ed
s
p
ee
ch
s
ig
n
al
af
ter
s
tag
e
-
2
at
-
5
d
B
in
p
u
t
s
n
r
f
o
r
b
ab
b
le
n
o
is
e
-20
0
20
O
UT
P
UT
SN
R
I
N
PUT
SN
R
I
N
d
B
OU
T
P
U
T
SN
R
W
D
AF-
LM
S
OU
T
P
U
T
SN
R
FC
C
D
N
N
-
SG
D
M
0
0.5
1
1.5
2
2.5
SN
R I
m
p
ro
v
e
m
e
n
t
0
1
2
3
0d
B
-5d
B
-6d
B
-7d
B
-8d
B
-9d
B
-10d
B
-11d
B
-12d
B
-13d
B
PE
SQ
P
E
S
Q
W
DA
F
-L
M
S
P
E
S
Q
F
C
C
DNN
-SG
D
M
0.7
0.8
0.9
1
0d
B
-5d
B
-6d
B
-7d
B
-8d
B
-9d
B
-10d
B
-11d
B
-12d
B
-13d
B
ST
OI
S
T
O
I W
DA
F
-L
MS
S
T
O
I F
CCDN
N
-S
G
DM
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:
2088
-
8
7
0
8
C
a
s
ca
d
ed
s
p
ee
ch
en
h
a
n
ce
me
n
t sys
tem
u
s
in
g
d
ee
p
lea
r
n
in
g
meth
o
d
(
K
a
vith
a
A
)
813
(
a)
(
b
)
(
c)
(
d
)
F
i
g
u
r
e
9
.
P
e
r
f
o
r
m
a
n
c
e
c
o
m
p
a
r
is
o
n
o
f
s
p
e
e
c
h
e
n
h
a
n
c
e
m
e
n
t
s
y
s
t
e
m
f
o
r
c
a
r
n
o
i
s
e
at
d
i
f
f
e
r
e
n
t
i
n
p
u
t
S
NR
l
e
v
e
ls
:
(
a
)
o
u
t
p
u
t
S
NR
f
o
r
c
a
r
n
o
i
s
e
at
a
l
l
i
n
p
u
t
SNR
l
e
v
e
l
,
(
b
)
S
NR
im
p
r
o
v
e
m
e
n
t
f
o
r
F
C
C
D
NN
-
S
GD
M
a
s
c
o
m
p
a
r
e
d
t
o
W
DA
F
-
L
MS
,
(
c
)
P
E
S
Q
i
m
p
r
o
v
e
m
e
n
t
f
o
r
F
C
C
DN
N
-
SG
D
M
a
s
c
o
m
p
a
r
e
d
t
o
W
D
AF
-
L
MS
f
o
r
,
a
n
d
(
d
)
S
T
O
I
i
m
p
r
o
v
e
m
e
n
t
f
o
r
F
C
C
DN
N
-
SG
D
M
as
c
o
m
p
a
r
e
d
t
o
W
D
AF
-
L
M
S
f
o
r
c
a
r
n
o
i
s
e
a
t
a
ll
i
n
p
u
t
S
NR
l
e
v
e
l
(
a)
(
b
)
(
c)
(
d
)
Fig
u
r
e
1
0
.
W
av
ef
o
r
m
c
o
m
p
a
r
is
o
n
o
f
s
p
ee
ch
s
ig
n
als co
r
r
u
p
te
d
b
y
ca
r
n
o
is
e:
(
a)
clea
n
s
p
ee
c
h
,
(
b
)
n
o
is
y
s
p
ee
ch
,
(
c)
d
en
o
is
ed
s
p
ee
ch
s
ig
n
al
af
ter
s
tag
e
-
1
,
(
d
)
an
d
d
en
o
is
ed
s
p
ee
ch
s
ig
n
al
af
ter
s
tag
e
-
2
at
-
5
dB
in
p
u
t
SNR
f
o
r
ca
r
n
o
is
e
-20
-10
0
10
20
1
2
3
4
5
6
7
8
9
10
O
UT
P
UT
SN
R
I
N
PUT
SN
R
I
N
D
B
OU
T
P
U
T
SNR
W
D
A
F-
LM
S
O
U
T
P
U
T
SN
R
FC
C
DN
N
-
SG
DM
0.8000
0.9000
1.0000
1.1000
1.2000
0DB
-5DB
-6DB
-7DB
-8DB
-9DB
-10DB
-11
D
B
-12DB
-13DB
SN
R I
m
p
ro
v
e
m
e
n
t
2.6
2.8
3
3.2
3.4
3.6
1
2
3
4
5
6
7
8
9
10
PE
SQ
PE
S
Q
W
DA
F
-L
M
S
PE
S
Q
F
CCDN
N
-S
G
D
M
0.97
0.975
0.98
0.985
0.99
0DB
-5DB
-6DB
-7DB
-8DB
-9DB
-10DB
-11DB
-12DB
-13DB
ST
OI
S
T
O
I W
DA
F
-L
MS
S
T
O
I F
CCDN
N
-S
G
DM
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
1
6
,
No
.
2
,
Ap
r
il
20
2
6
:
8
0
6
-
817
814
T
a
b
l
e
3
an
d
F
i
g
u
r
e
1
1
s
h
o
w
t
h
e
en
h
an
c
e
m
e
n
t
in
n
o
i
s
y
s
p
e
e
ch
s
i
g
n
a
l
c
o
r
r
u
p
t
ed
b
y
m
a
c
h
i
n
e
g
u
n
n
o
i
s
e
a
t
0
d
B
,
-
5
d
B
,
-
6
d
B
,
-
7
d
B
,
-
8
d
B
,
-
9
d
B
,
-
1
0
d
B
,
-
1
1
d
B
,
-
1
2
d
B
,
-
1
3
d
B
i
n
p
u
t
S
N
R
L
e
v
e
l
s
i
n
t
e
r
m
s
o
f
S
N
R
,
P
E
S
Q
a
n
d
S
T
O
I
,
wh
e
n
i
t
i
s
p
a
s
s
e
d
th
r
o
u
g
h
t
h
e
s
y
s
t
e
m
,
w
i
th
th
e
s
p
e
c
if
i
c
m
e
tr
i
c
s
d
e
t
a
i
l
ed
in
F
i
g
u
r
e
1
1
(
a)
t
h
r
o
u
g
h
F
i
g
u
r
e
1
1
(
d
)
.
T
a
b
l
e
3
s
h
o
w
s
t
h
e
c
o
m
p
a
r
i
s
o
n
s
o
f
o
u
tp
u
t
a
t
b
o
th
s
t
a
g
e
s
.
T
h
e
W
D
A
F
-
L
M
S
s
h
o
w
s
m
ax
i
m
u
m
S
N
R
i
m
p
r
o
v
e
m
e
n
t
o
f
1
1
.
1
8
8
1
d
B
a
t
-
5
d
B
in
p
u
t
S
N
R
l
e
v
e
l
wh
i
l
e
F
C
C
D
N
N
-
S
G
D
M
s
h
o
w
s
f
u
r
t
h
e
r
S
N
R
i
m
p
r
o
v
e
m
en
t
o
f
1
8
.
2
5
3
3
d
B
a
t
-
5
d
B
in
p
u
t
S
N
R
l
e
v
e
l
.
E
v
e
n
a
t
-
1
3
d
B
i
n
p
u
t
S
N
R
l
ev
e
l
,
t
h
e
p
r
o
p
o
s
ed
a
l
g
o
r
i
th
m
p
r
o
v
id
e
s
s
u
b
s
e
q
u
e
n
t
i
m
p
r
o
v
em
e
n
t
w
i
th
i
m
p
r
o
v
e
m
en
t
s
i
n
P
E
S
Q
a
n
d
S
T
O
I
v
a
lu
e
s
.
W
h
ic
h
s
h
o
w
s
th
a
t
a
t
-
1
3
d
B
i
n
p
u
t
S
N
R
l
e
v
e
l
s
p
e
e
c
h
q
u
a
l
i
t
y
an
d
i
n
te
l
l
i
g
ib
i
l
i
t
y
i
s
r
e
t
a
i
n
ed
in
d
e
n
o
i
s
e
d
s
i
g
n
a
l.
(
a)
(
b
)
(
c)
(
d
)
Fig
u
r
e
1
1
.
Per
f
o
r
m
a
n
ce
co
m
p
ar
is
o
n
o
f
s
p
ee
ch
en
h
a
n
ce
m
en
t
s
y
s
tem
f
o
r
m
ac
h
in
e
g
u
n
n
o
is
e
at
d
if
f
er
en
t
in
p
u
t
SNR
lev
els
:
(
a)
Ou
tp
u
t SNR
f
o
r
m
ac
h
in
e
g
u
n
n
o
is
e
at
all
in
p
u
t SNR
lev
el
,
(
b
)
SNR
im
p
r
o
v
em
en
t f
o
r
FC
C
DNN
-
SG
DM
as c
o
m
p
ar
ed
to
W
DAF
-
LMS
,
(
c)
PESQ
im
p
r
o
v
e
m
en
t f
o
r
FC
C
DNN
-
S
GDM
as c
o
m
p
ar
ed
to
W
DAF
-
LMS
,
an
d
(
d
)
STOI
im
p
r
o
v
e
m
en
t f
o
r
FC
C
DNN
-
S
GDM
as c
o
m
p
ar
ed
to
W
DAF
-
L
MS
f
o
r
m
ac
h
in
e
g
u
n
n
o
is
e
at
all
in
p
u
t SNR
lev
el
Fig
u
r
e
1
2
s
h
o
ws
d
en
o
is
ed
o
u
tp
u
t
s
ig
n
al
o
f
s
tag
e
-
1
an
d
s
ta
g
e
-
2
at
-
5
d
B
in
p
u
t
SNR
lev
el
an
d
th
e
q
u
ality
o
f
s
ig
n
als
r
ec
eiv
ed
a
f
ter
s
tag
e
-
2
is
b
etter
th
an
th
at
o
f
s
tag
e
-
1
.
Dis
p
lay
in
g
th
e
clea
n
s
p
ee
ch
in
Fig
u
r
e
1
2
(
a)
,
th
e
n
o
is
y
s
p
ee
c
h
in
Fig
u
r
e
1
2
(
b
)
,
th
e
d
e
n
o
is
ed
s
ig
n
al
af
ter
s
tag
e
-
1
in
Fig
u
r
e
1
2
(
c
)
,
an
d
th
e
d
en
o
is
ed
s
ig
n
al
af
ter
s
tag
e
-
2
i
n
Fig
u
r
e
1
2
(
d
)
.
-20
-10
0
10
20
1
2
3
4
5
6
7
8
9
10
Ou
tp
u
t
SN
R
INP
UT
S
N
R
I
N
dB
O
UT
P
UT
SNR
W
D
A
F
-
LM
S
O
UT
P
UT
SNR
FC
C
D
N
N
-
S
G
D
M
0.0000
1.0000
2.00
00
3.0000
4.0000
5.0000
6.0000
7.00
00
8.0000
0d
B
-5d
B
-6d
B
-7d
B
-8d
B
-9d
B
-10d
B
-11d
B
-12d
B
-13d
B
SN
R I
m
p
ro
v
e
m
e
n
t
0
0.5
1
1.5
2
2.5
3
3.5
0d
B
-5d
B
-6d
B
-7d
B
-8d
B
-9d
B
-10d
B
-11d
B
-12d
B
-13d
B
PE
SQ
P
E
S
Q
W
D
A
F
-
LM
S
P
E
S
Q
F
C
C
D
N
N
-
S
G
D
M
0
0.2
0.4
0.6
0.8
1
1.2
ST
OI
S
T
OI W
DA
F
-
LMS
S
T
O
I F
C
C
D
N
N
-
S
G
D
M
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:
2088
-
8
7
0
8
C
a
s
ca
d
ed
s
p
ee
ch
en
h
a
n
ce
me
n
t sys
tem
u
s
in
g
d
ee
p
lea
r
n
in
g
meth
o
d
(
K
a
vith
a
A
)
815
(
a)
(
b
)
(
c)
(
d
)
Fig
u
r
e
1
2
.
W
av
ef
o
r
m
c
o
m
p
a
r
is
o
n
o
f
s
p
ee
ch
s
ig
n
als
co
r
r
u
p
te
d
b
y
m
ac
h
in
e
g
u
n
n
o
is
e
at
-
5
d
B
in
p
u
t
SNR
:
(
a)
clea
n
s
p
ee
ch
,
(
b
)
n
o
is
y
s
p
e
ec
h
,
(
c)
d
en
o
is
ed
s
p
ee
c
h
s
ig
n
a
l a
f
ter
s
tag
e
-
1
,
a
n
d
(
d
)
d
e
n
o
is
ed
s
p
ee
ch
s
ig
n
al
af
ter
s
tag
e
-
2
at
-
5
d
B
in
p
u
t SNR
f
o
r
m
ac
h
in
e
g
u
n
n
o
is
e
5.
CO
NCLU
SI
O
N
Her
e,
a
two
-
s
tag
e
ca
s
ca
d
ed
s
p
ee
ch
e
n
h
an
ce
m
e
n
t
m
eth
o
d
is
p
r
o
p
o
s
ed
f
o
r
n
o
is
e
ca
n
ce
lla
tio
n
f
r
o
m
h
ig
h
ly
co
r
r
u
p
ted
s
p
ee
ch
s
ig
n
als.
W
h
er
e
s
tag
e
-
1
is
W
DA
F
-
L
MS
b
ased
s
y
s
tem
an
d
s
t
ag
e
-
2
is
FC
C
DNN
-
SGDM
b
ased
s
y
s
tem
.
T
h
e
r
esu
lt
is
p
r
esen
ted
f
o
r
s
p
ee
ch
c
o
r
r
u
p
te
d
b
y
b
ab
b
le
n
o
is
e,
ca
r
n
o
is
e
an
d
m
ac
h
i
n
e
gun
n
o
is
es
at
0
dB
an
d
-
5
dB
to
-
13
dB
in
p
u
t
SNR
lev
els
an
d
th
e
r
esu
lts
ar
e
co
m
p
a
r
e
d
at
b
o
th
s
tag
es
i.e
.
s
tag
e
-
1
an
d
s
tag
e
-
2
.
Sp
ee
c
h
q
u
ality
an
d
in
tellig
ib
ly
is
co
m
p
ar
ed
in
ter
m
s
o
f
SNR
im
p
r
o
v
em
en
t,
PESQ
a
n
d
STOI
at
all
in
p
u
t
SNR
lev
els
f
o
r
all
th
r
ee
n
o
is
es.
I
t
is
p
r
o
v
ed
th
at
th
e
p
r
o
p
o
s
ed
s
y
s
tem
s
h
o
ws
im
p
r
o
v
em
e
n
t
in
SNR
,
PES
Q
,
an
d
STOI
u
p
t
o
-
13
dB
in
p
u
t SNR
lev
el
co
r
r
u
p
t
ed
s
p
ee
ch
s
ig
n
al.
F
UNDING
I
NF
O
R
M
A
T
I
O
N
Au
th
o
r
s
s
tate
n
o
f
u
n
d
in
g
in
v
o
lv
ed
.
AUTHO
R
CO
NT
RI
B
UT
I
O
NS ST
A
T
E
M
E
N
T
First
a
u
th
o
r
o
n
e
h
as
co
n
tr
ib
u
ted
to
co
n
ce
p
tu
aliza
tio
n
,
m
eth
o
d
o
lo
g
y
an
d
f
o
r
m
al
an
al
y
s
is
wh
il
e
s
ec
o
n
d
au
th
o
r
h
as
p
r
e
p
ar
ed
th
e
o
r
ig
in
al
d
r
af
t
o
f
th
e
p
a
p
er
.
T
h
ir
d
a
u
th
o
r
is
in
v
o
lv
e
d
in
r
e
v
iew
an
d
e
d
itin
g
o
f
th
e
p
ap
er
.
Na
m
e
o
f
Aut
ho
r
C
M
So
Va
Fo
I
R
D
O
E
Vi
Su
P
Fu
Kav
ith
a
A
✓
✓
✓
Ma
h
esh
C
h
an
d
r
a
✓
Vijay
Ku
m
ar
Gu
p
ta
✓
C
:
C
o
n
c
e
p
t
u
a
l
i
z
a
t
i
o
n
M
:
M
e
t
h
o
d
o
l
o
g
y
So
:
So
f
t
w
a
r
e
Va
:
Va
l
i
d
a
t
i
o
n
Fo
:
Fo
r
mal
a
n
a
l
y
s
i
s
I
:
I
n
v
e
s
t
i
g
a
t
i
o
n
R
:
R
e
so
u
r
c
e
s
D
:
D
a
t
a
C
u
r
a
t
i
o
n
O
:
W
r
i
t
i
n
g
-
O
r
i
g
i
n
a
l
D
r
a
f
t
E
:
W
r
i
t
i
n
g
-
R
e
v
i
e
w
&
E
d
i
t
i
n
g
Vi
:
Vi
su
a
l
i
z
a
t
i
o
n
Su
:
Su
p
e
r
v
i
s
i
o
n
P
:
P
r
o
j
e
c
t
a
d
mi
n
i
st
r
a
t
i
o
n
Fu
:
Fu
n
d
i
n
g
a
c
q
u
i
si
t
i
o
n
Evaluation Warning : The document was created with Spire.PDF for Python.