Indonesi
an
Journa
l
of El
ect
ri
cal Engineer
ing
an
d
Comp
ut
er
Scie
nce
Vo
l.
1
3
,
No.
3
,
Ma
rch
201
9
, p
p.
866
~
8
75
IS
S
N: 25
02
-
4752, DO
I: 10
.11
591/ijeecs
.v1
3
.i
3
.pp
866
-
8
75
866
Journ
al h
om
e
page
:
http:
//
ia
es
core.c
om/j
ourn
als/i
ndex.
ph
p/ij
eecs
Sp
ee
ch i
ntelligibi
lity enh
an
ce
m
ent for
T
h
ai
-
speakin
g cochle
ar
imp
lant listeners
Siri
po
rn
D
achasil
aruk
1
, Nip
ha
t
Jan
thar
am
in
2
, Apichai
Ru
n
gruan
g
3
1,2
Depa
rtment
of
Elec
tr
ical and
C
om
pute
r
Engi
n
e
eri
ng,
Facu
lty
of
Engi
n
ee
ring
,
N
a
resua
n
Unive
rsit
y
,
Th
ai
l
and
3
Depa
rtment of
Engl
ish
La
nguag
e,
Fa
cult
y
of
Hum
ani
ti
es,
Nare
su
an
Univer
si
t
y
,
T
hai
l
and
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
A
ug
03, 201
8
Re
vised
N
ov
2
3
, 2
018
Accepte
d
Dec
12, 201
8
Cochl
e
ar
implan
t
(CI)
li
sten
ers
e
ncount
er
difficul
ti
es
in
comm
unic
ating
wi
t
h
othe
r
pe
rsons
in
nois
y
li
sten
ing
envi
ronm
ent
s.
How
eve
r,
m
ost
CI
rese
ar
ch
has
bee
n
ca
rr
ie
d
out
using
th
e
En
gli
sh
la
ngu
age.
I
n
thi
s
stud
y
,
sin
gle
-
ch
annel
spee
ch
enh
ance
m
ent
(SE)
strat
e
gie
s
as
a
pre
-
pro
ce
ss
ing
appr
oa
c
h
for
the
CI
s
y
stem
were
inv
esti
gated
in
t
erms
of
Tha
i
spee
ch
int
elligi
b
il
i
t
y
improvem
ent.
Two
SE
al
gorithm
s,
namel
y
m
ult
i
-
band
spe
ct
r
al
subtra
ct
ion
(
MBS
S)
and
W
ei
ner
fi
lt
er
(W
F)
al
gorit
hm
s,
w
ere
ev
al
u
at
ed
.
S
pee
ch
sign
al
s
co
nsisting
of
m
onosy
l
la
bi
c
an
d
bis
y
llabic
Th
ai
words
wer
e
degr
ade
d
b
y
sp
ee
ch
-
shap
ed
noise
and
babble
noise
at
SN
R
le
vel
s
of
0,
5,
and
10
dB.
Then
the
noi
s
y
words
were
enha
nce
d
using
SE
al
gorit
hm
s.
Th
e
enha
n
ce
d
wor
ds
were
fed
int
o
th
e
CI
s
y
s
t
em
to
s
y
n
the
siz
e
vocod
ed
spee
c
h.
Th
e
vocod
ed
spee
ch
was
pre
sente
d
to
tw
enty
norm
al
-
h
e
ari
ng
l
iste
n
ers.
The
r
esult
s
ind
ic
a
te
d
th
at
spee
ch
intelligib
il
ity
was
m
arg
ina
lly
improved
by
the
MBS
S
al
gorit
hm
and
signifi
c
ant
l
y
improved
b
y
the
W
F
al
gorit
hm
in
so
m
e
condi
ti
ons.
Th
e
enha
nc
ed
bis
y
l
l
abi
c
words
sh
owed
a
not
ice
ably
high
er
in
te
lligibi
lit
y
improvem
ent
th
an
the
enha
n
ced
m
onosy
l
la
bi
c
words
in
al
l
condi
ti
ons
,
par
ticula
r
l
y
in
spee
ch
-
shap
ed
n
oise.
Such
outcom
es
m
ay
be
b
ene
fi
c
ial
t
o
Tha
i
-
spe
aki
ng
CI
li
st
ene
rs
.
Ke
yw
or
ds:
Sound c
odin
g
Sp
ect
ral s
ubtra
ct
ion
Sp
eec
h
e
nh
a
nc
e
m
ent
Sp
eec
h
intel
li
gi
bili
ty
Weine
r
filt
er
Copyright
©
201
9
Instit
ut
e
o
f Ad
vanc
ed
Engi
n
ee
r
ing
and
S
cienc
e
.
Al
l
rights re
serv
ed.
Corres
pond
in
g
Aut
h
or
:
Sirip
orn Dacha
sil
aru
k
,
Dep
a
rtm
ent o
f El
ect
rical
an
d
Com
pu
te
r
E
ng
i
neer
i
ng,
Faculty
of E
ngineerin
g,
Nar
es
ua
n Un
i
ve
rsity
, P
hitsa
nulo
k 650
00, T
ha
il
and
.
Em
a
il
:
siripo
r
nd@
nu.ac.t
h,
dsi
riporn@
ho
tm
ai
l.com
1.
INTROD
U
CTION
Most
CI
li
ste
ne
rs
can
ac
hiev
e
high
spe
ech
intel
li
gib
il
it
y
wh
ic
h
is
cl
os
e
to
the
capa
bili
ty
of
no
rm
al
-
hear
i
ng
(
NH)
l
ist
ener
s
i
n
qu
ie
t
li
ste
nin
g
e
nviro
nm
ents.
This
is
beca
us
e
al
m
os
t
al
l
the
sound
c
odin
g
str
at
egies
us
e
d
by
m
od
e
rn
C
I
de
vices
perform
well
i
n
qu
ie
t
li
ste
nin
g
en
vir
on
m
ents
[
1
]
.
H
owe
ver,
m
os
t
CI
li
ste
ner
s
su
f
fer
f
ro
m
de
creased
s
peec
h
intel
li
gib
il
it
y
m
or
e
tha
n
NH
li
ste
ner
s
in
no
isy
li
ste
ni
ng
en
vir
onm
e
nts.
T
he
higher
t
he
no
is
e
le
vel,
the
l
ower
t
he
s
peech
intel
li
gib
il
it
y
perform
ance
[
2
]
.
O
ne
of
t
he
sp
eci
fic
li
m
it
ation
s
of
CI
de
vices
in
te
rm
s
of
fr
eq
ue
ncy,
te
m
po
ral
and
am
plit
ude
reso
l
ution
s
[
3
]
is
relat
ed
to
transm
itti
ng
sp
eech
inf
or
m
at
ion
to
the
au
ditor
y
ne
rv
es
.
A
nothe
r
lim
i
ta
ti
on
is
the
eff
ect
of
c
hannel
interact
io
n,
wh
ic
h
res
ul
ts
from
the
overla
p
of
el
ect
rical
fiel
ds
betwee
n
el
ec
tro
des
[
4
]
.
Ele
ct
ric
stim
ulati
o
n
of
one
el
ect
r
od
e
m
ay
be
distor
te
d
by
the
sti
m
ulatio
n
of
oth
e
r
el
ect
rodes.
S
uc
h
interact
ions
c
an
dec
rease
int
el
li
gib
i
li
ty
per
fo
rm
ance.
Th
er
efore,
CI
researc
he
rs
hav
e
in
creasi
ng
ly
at
tem
pted
to
im
pr
ov
e
sp
eec
h
intel
li
gib
il
it
y
per
form
ance,
par
ti
cula
rly
by
dev
el
op
i
ng s
pe
ech e
nh
a
ncem
ent (SE)
strategi
es for
u
se
in
a
dverse
noisy
e
nvir
on
m
ents.
Gen
e
rall
y,
sin
gle
-
c
hannel
S
E
strat
eg
ie
s
ar
e
us
e
d
i
n
m
os
t
tradit
ion
al
C
I
syst
em
s,
and
this
ca
n
be
exten
ded
to
ap
ply
to
m
ulti
-
ch
ann
el
SE
strat
e
gies.
The
refor
e
,
sin
gle
-
c
ha
nn
e
l
SE
strat
e
gies
we
re
em
plo
ye
d
i
n
this
stud
y.
S
e
ver
al
stu
dies
hav
e
in
dicat
e
d
that
these
sing
le
-
c
ha
nn
el
SE
al
go
rit
hm
s
i
m
pr
o
ved
sp
eec
h
intel
li
gib
il
it
y
s
ign
ific
a
ntly
for
hea
rin
g
-
im
paired
(
H
I)
li
ste
ner
s
.
SE
al
gor
it
h
m
s
su
ch
as
the
pre
-
proce
ssin
g
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Sp
eec
h
i
ntell
igibil
it
y enhan
ce
men
t f
or
t
ha
i
-
s
pea
ki
ng co
c
hle
ar
im
pl
an
t l
ist
eners
(
Siri
porn
Da
c
hasil
aruk
)
867
appr
oach
ha
ve
bee
n
a
pp
li
ed
to
CI
syst
em
s,
i
nclu
d
ing
s
ubspa
ce
-
base
d,
W
e
iner
filt
er,
an
d
sp
ect
ral
s
ubtra
ct
ive
al
gorithm
s.
Loizou
et
al
.
[
5
]
dem
on
strat
ed
that
the
subsp
a
ce
-
base
d
al
gori
thm
pr
op
os
ed
by
Hu
a
nd
L
oizou
[
6
]
sign
ific
a
ntly
i
m
pr
ov
e
d
sente
nce
rec
ogniti
on
in
sp
eec
h
-
sha
ped
no
ise
at
a
5
dB
sign
al
-
to
-
no
ise
rati
o
(S
NR
)
a
m
on
g
a
gro
up
of
f
ourteen
Cl
arion
C
I
li
s
te
ner
s,
with
a
n
a
ver
a
ge
im
pr
ovem
ent
of
44%.
H
oweve
r,
this
al
gorithm
can
al
so
pr
ov
i
de
r
e
cogniti
on
be
ne
fits
for
sta
ti
ona
ry
noise
(e
.g.
s
peech
-
s
hap
e
d
no
ise
(SSN
)),
bu
t
t
he
al
gorithm
do
e
s
not
gu
a
ra
nt
ee
i
m
pr
ovem
ent
f
or
non
-
s
ta
ti
on
ary
no
is
es
(e.
g.
babb
le
no
ise
(BB
N)).
Bolner
et
al
.
[
7
]
sh
owe
d
t
hat
a
W
ei
ner
filt
er
(
WF)
base
d
on
a
pri
ori
S
NR
e
stim
at
es
[
8
]
sig
nificantl
y
im
pr
ov
e
d
sentence
recog
niti
on
in
a
SS
N
co
nd
it
io
n
at
0
dB
SN
R
in
a
gr
ou
p
of
te
n
N
H
li
ste
ner
s
,
bu
t
the
re
wa
s
n
o
i
m
pr
ovem
ent
i
n
a
BB
N
co
ndit
ion
.
Additi
on
a
ll
y,
a
stud
y
by
Koning
et
al
.
[
9
]
showe
d
that
the
W
F
al
go
rithm
,
app
l
ie
d
as
a
n
env
el
ope
-
wei
gh
ti
ng
a
ppro
a
ch,
pr
ov
i
ded
bo
t
h
sp
eec
h
intel
li
gib
il
it
y
a
nd
s
peec
h
qu
al
it
y
i
m
pr
ovem
ent f
or a
group o
f
si
x Du
tc
h
-
s
pea
kin
g C
I
li
ste
ne
rs and si
x D
utch
-
sp
ea
king
NH li
ste
ner
s
.
Ov
e
r
al
m
os
t
fo
ur
deca
des
,
sp
ect
ral
s
ubtra
ct
ion
(
SS
)
al
gorithm
s
ha
ve
been
de
velo
pe
d
in
m
any
ver
si
ons,
an
d
s
om
e
of
the
se
ha
ve
bee
n
ap
plied
in
CI
syst
em
s.
An
SS
al
go
rithm
ref
err
e
d
to
as
the
IN
T
EL
S
E
al
gorithm
was
first
ap
plied
by
Ho
c
hb
e
r
g
et
al
.
[
10
]
.
Co
nso
nan
t
-
vo
wel
-
co
ns
ona
nt
w
ords
corrupted
by
SSN
at
SN
R
ra
ng
i
ng
f
ro
m
-
10
to
25
dB
for
N
H
li
st
ener
s
a
nd
f
ro
m
-
5
to
25
dB
for
CI
li
ste
ner
s
wer
e
pr
ocesse
d
by
the
IN
T
EL
SE
al
gorithm
s.
The
e
nh
a
nce
d
w
ords
wer
e
pr
e
sente
d
to
te
n
N
H
li
ste
ner
s
a
nd
te
n
Nu
cl
e
us
CI
li
st
ener
s
.
W
or
d
rec
ogniti
on
was
sig
nificantl
y
i
m
pr
ov
e
d
for
CI
li
ste
ne
r
s
but
not
f
or
N
H
li
ste
ner
s
.
A
stud
y
by
W
ei
ss
[
11
]
ind
ic
at
ed
that
wh
e
n
the
nois
y
sp
eech
sig
na
ls
wer
e
en
hanced
by
the
INTEL
SE
al
gorithm
,
the
err
or
of
th
e
seco
nd
f
or
m
ant
extracti
on
w
as
co
ns
ide
ra
bly
reduce
d
in
t
he
Nu
cl
e
us
im
plant
c
odin
g
st
rategy.
These
eff
ect
s
wer
e
u
se
d
t
o
i
m
pr
ov
e s
peec
h pe
rcep
ti
on i
n
t
he pr
i
or
stu
dy.
Yang
a
nd
F
u
[
12
]
fou
nd
sign
ific
a
nt
im
pr
ov
em
ents
with
the
S
S
al
gorithm
pro
po
se
d
by
Gu
sta
fss
on
et
a
l.
[
13
]
w
he
n
it
w
as a
pp
li
ed
t
o sentence
rec
ogniti
on
i
n
SS
N at
diff
e
ren
t S
N
Rs (i.e.
0,
3,
6,
and
9
dB)
in
a
gro
up
of
se
ven
C
I
li
ste
ner
s
w
ho
us
e
d
dif
fere
nt
CI
de
vices
(i.
e.
N
ucleus
,
Me
d
-
El,
a
nd
Cl
ario
n).
Ver
sc
huur
et
a
l.
[
14
]
in
dicat
ed
that
se
ntence
recog
niti
on
wi
th
the
nonlinea
r
SS
(NSS
)
al
gorithm
propose
d
by
Loc
kwo
od
a
nd
Boudy
[
15
]
was
sig
nifica
nt
ly
i
m
pr
ov
ed
i
n
SS
N
at
both
5
an
d
10
dB
SN
Rs
f
or
seve
nteen
Nu
cl
e
us
CI
li
s
te
ner
s.
H
owev
er,
s
uc
h
be
ne
f
it
s
m
ay
be
limit
ed
to
s
uppr
e
s
sing
non
-
sta
ti
onary
no
ise
.
A
la
te
r
stud
y
by
Kall
e
l
et
al
.
[
16
]
a
ppli
ed
the
N
SS
al
gorithm
pr
op
os
e
d
b
y
Be
r
outi
et
al
.
[
17
]
a
nd
the
m
ulti
-
ban
d
SS
(MBSS)
al
go
rithm
pr
opos
e
d
by
Kam
at
h
and
Loiz
ou
[
18
]
to
three
bilat
era
l
Neu
relec
CI
l
ist
ener
s
an
d
fi
f
ty
NH
l
ist
ener
s.
T
he
resu
lt
s
sho
w
ed
that
the
a
ver
a
ge
w
ord
r
ecognit
ion
im
pro
vem
ent
was
4
–
9%
f
or
bi
la
te
ra
l
Neurele
c
CI
li
s
te
ner
s
a
nd
7
–
13%
f
or
N
H
li
ste
ner
s
at
al
l
SNR
s
(i.e.
-
3,
0,
3,
an
d
6
dB
).
M
or
e
over,
the
re
su
lt
s
al
so
in
dicat
ed
that
the
MB
S
S
al
gorithm
enh
a
nce
d
s
peec
h
intel
li
gib
il
it
y
m
or
e
than
t
he
NS
S
al
gorit
hm
fo
r
sing
le
a
nd m
ulti
ple interfe
rin
g no
ise
sou
rces
.
Nev
e
rtheless
,
m
os
t
SE
strat
egies
for
C
I
li
ste
ner
s
hav
e
bee
n
e
valuate
d
us
i
ng
the
E
ng
li
sh
la
nguag
e
.
A
few
stu
dies
ha
ve
eval
uated
st
rategies
us
i
ng
Fr
e
nch,
H
e
br
e
w,
D
utch/Flem
ish
,
an
d
Chi
ne
se,
but
SE
stra
te
gies
hav
e
ne
ver
be
en
eval
uated
usi
ng
t
he
Th
ai
la
nguag
e
.
Di
fferent
la
ng
uag
e
s
ha
ve
dif
fere
nt
aco
us
ti
c
cu
es
an
d
phonem
ic
s,
whic
h
m
a
y
pr
od
uc
e
diff
e
re
nt
intel
li
gib
il
it
y
per
fo
rm
ances
us
i
ng
the
sam
e
SE
te
chn
iq
ues
.
E
ng
li
s
h
is
a
non
-
to
nal
la
nguag
e
,
wh
e
reas
T
hai
is
a
ton
al
la
ng
uag
e
wh
ic
h
is
sim
i
la
r
to
m
any
Asian
la
ngua
ges
(e.
g.
Chinese
a
nd
Viet
nam
ese).
A
tona
l
la
ngua
ge
us
e
s
a
tona
l
le
vel
wh
ic
h
is
disti
ng
uis
he
d
by
the
f
un
dam
ental
fr
e
qu
e
ncy
(
F0)
co
ntours
t
o
re
pr
ese
nt
le
xi
cal
m
eaning
.
Eac
h
to
ne
of
a
w
ord
re
pr
ese
nts
a
dif
fer
e
nt
m
ea
ning.
Norm
al
l
y
a
Th
ai
syl
la
ble
consi
sts
of
an
init
ia
l
conso
na
nt
(a
sing
le
/c
luste
re
d
c
on
s
ona
nt),
a
vowel
(s
hort/l
ong),
an
opti
onal
fin
al
con
s
on
a
nt
and
a
to
nal
le
ve
l.
Ther
e
are f
iv
e
disti
nctive
to
nes
in
T
hai
syllable
s:
the
m
idd
le
/¯/,
the
lo
w
/ˋ/
,
the
fall
ing
/
ˆ/,
the
high
/ˊ
/,
and
th
e
risin
g
/ˇ/
.
T
he
Thai
to
nes
of
m
on
os
yl
la
bic
words
a
re
c
omm
on
ly
avail
able.
E
xa
m
ples
of
m
on
os
yl
la
bic
wor
ds
with
five
to
ne
s
diff
e
re
ntiat
i
ng
th
ei
r
m
eani
ng
a
re
/pāa/
(t
hrow),
/pàa/
(for
e
st),
/
pâa/
(a
un
t
),
/
pá
a/
(d
a
d)
a
nd
/p
ǎa/
(d
a
d).
T
hes
e
com
po
ne
nts
are
im
po
rtant
t
o
the
pe
rfor
m
ance
of
sp
eec
h
intel
li
gib
il
it
y
a
m
on
g
T
hai
-
s
peak
i
ng C
I
li
ste
ner
s
.
No
st
ud
ie
s
ha
ve
sp
eci
fical
ly
evaluate
d
SE
m
et
ho
ds
with
Thai
-
s
pea
king
CI
li
ste
ner
s.
T
he
re
fore,
t
he
obj
ect
ive
of
the
present
st
ud
y
is
to
inve
sti
gate
the
sp
eech
intel
li
gi
bili
ty
per
form
ance
of
existi
ng
S
E
al
gorithm
s
and
to
assess
w
hethe
r
dif
fer
e
nt
SE
al
go
rit
hm
s
pr
ovide
diff
e
ren
t
intel
li
gi
bili
ty
per
form
ance
in
var
i
ou
s
noisy
env
i
ronm
ents
for
T
hai
-
s
peaki
ng
CI
li
ste
ne
r
s.
The
i
nvest
igati
on
of
the
i
m
pr
ov
em
ent
of
Tha
i
word
rec
ogniti
on
c
once
ntrate
s
on
SE
al
gorithm
s
as
the
pr
e
-
proce
ssin
g
a
ppr
oac
h,
nam
ely
m
ulti
-
ban
d
s
pectral
su
bt
racti
on
(MBSS)
an
d
W
ei
ner
filt
er
(
WF)
al
gorithm
s.
Both
achie
ve
a
trade
-
off
be
t
ween
ef
fecti
ve
no
ise
reducti
on
,
s
pe
ech
dist
or
ti
on,
and
l
ow
c
ompu
ta
ti
on
c
os
ts
for
real
-
ti
m
e
i
m
ple
m
entat
i
on
s
[
8
,
14
]
.
T
he
WF
al
gorithm
with
nonvoc
oded
s
peech
has
s
ho
wn
high
s
peec
h
intel
li
gib
il
it
y
scor
es
i
n
thre
e
la
nguag
e
s:
Eng
li
s
h,
Chinese
,
a
nd
J
apan
e
se
[
19
]
.
The
MB
SS
al
gorithm
is
on
e
ver
si
on
of
the
sp
ect
ral
subtra
ct
i
on
al
gorith
m
wh
ic
h
has
bee
n
s
how
n
to
yi
el
d
a
s
pe
ech
intel
li
gib
i
li
ty
i
m
pr
ov
em
ent
f
or
F
ren
c
h
-
sp
ea
king
a
nd
Ma
nd
a
rin
-
sp
ea
king
C
I
li
ste
ner
s.
S
om
e
p
re
vious
stu
di
es
su
pp
or
t
sel
e
ct
ing
both
S
E
al
gorithm
s.
A
sp
eec
h
intel
li
gib
il
it
y
evaluati
on
was
cond
ucted
on
NH
li
ste
ne
rs
i
n
a
CI
sim
ulatio
n
with
a
no
i
se
-
ba
nd
voc
oder.
T
he
pr
ese
nt
stud
y
was
ex
te
nd
e
d
from
a stud
y
by
D
acha
sil
aruk
et
al.
[
20
]
wi
th
a lar
ge
r n
um
ber
of
sub
j
ect
s.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
1
3
, N
o.
3
,
Ma
rc
h
201
9
:
866
–
8
7
5
868
2.
SPEE
CH E
N
HANCE
MEN
T FOR
THE
CI
S
YS
TE
M
2
.
1.
S
peech
Enhan
c
emen
t Alg
orithm
s
Figure
1
prese
nts
a
CI
sim
ul
at
ion
with
a
noise
-
band
vo
c
od
e
r
based
on
sp
eec
h
en
ha
nc
e
m
ent.
The
no
isy
s
peec
h
is
processe
d
with
a
SE
al
go
rith
m
to
ge
ner
at
e
enh
a
nce
d
s
pee
ch
.
Af
te
r
th
at
,
the
en
ha
nced
s
peech
is
fed
int
o
the
CI
syst
em
to
pro
duce
vo
c
ode
d
s
peec
h
.
Furt
her
desc
riptio
ns
of
t
he
MB
SS
an
d
WF
al
gor
it
h
m
s
are
giv
e
n
in
K
a
m
at
h
and
Loi
zou
[
18
]
and
Scal
art
and
Vi
ei
ra
[
8
]
resp
ec
ti
vely
.
B
oth
alg
ori
thm
s
are
br
ie
fly
descr
i
bed
i
n
th
is
sect
ion
.
Assum
e
that
no
isy
sp
eech
sig
nals
(
y
)
at
a
sam
pli
ng
rate
of
16
kHz
are
gen
e
rat
ed
by
add
i
ng
noise
(
n
)
t
o
cl
ean
s
pe
ech
sig
nals
(
x
)
.
The
n
,
t
he
po
wer
sp
ect
r
um
of
t
he
no
isy
s
peech
sig
nals
can
be
appr
ox
im
at
ely
est
i
m
at
ed
as
fol
lows
:
2
2
2
(
)
(
)
(
)
Y
k
X
k
N
k
(1)
T
he
MB
SS
a
l
gorithm
is
slig
htly
dif
fer
e
nt
to
the
N
SS
a
lgorit
hm
s
.
The
MB
SS
us
es
a
factor
of
su
bt
racti
on
e
stim
at
ed
in
each
f
reque
ncy
bi
n
an
d
eac
h
f
r
equ
e
ncy
ba
nd,
wh
e
reas
the
NS
S
us
es
t
his
facto
r
est
i
m
at
e
d
in
e
ach
fr
e
quency
bin
.
T
he
co
nce
pt
of
the
MB
S
S
is
that
the
c
har
act
erist
ic
s
of
the
no
ise
s
pe
ct
ru
m
m
ay
no
t
aff
ect
the
sp
eec
h
s
pe
ct
ru
m
equ
al
ly
acro
s
s
the
e
ntir
e
fr
e
quency
ba
nd.
T
he
noise
sp
ect
r
um
m
a
y
aff
ect
so
m
e
fr
eq
uen
c
y
bands
m
or
e
or
le
ss
t
han
ot
her
s
.
T
he
re
for
e,
sp
ect
r
al
sub
tract
ion
is
perf
or
m
ed
sepa
ratel
y
in
each
fr
e
qu
e
ncy
b
a
nd. In
the
i
th
ba
nd, th
e
s
pec
trum
o
f
the
cle
an
s
peec
h
is es
tim
a
te
d
as:
22
2
ˆˆ
(
)
(
)
(
)
,
i
i
i
i
i
i
i
X
k
Y
k
N
k
b
k
e
(2)
wh
e
re
2
ˆ
()
Nk
is
the
est
i
m
at
ed
sp
ect
ru
m
of
the n
ois
e
sign
al
,
b
i
a
nd
e
i
are
the
sta
rt
and
sto
p
bi
ns
of
the
i
th
ba
nd
,
and
α
i
a
nd
δ
i
a
r
e
the ove
r
-
s
ubtract
ion
an
d we
igh
t fact
or
s
of
t
he
i
th
ba
nd,
res
pecti
vely
.
T
he wei
ght fact
or
c
an be
ind
ivi
du
al
ly
se
t for eac
h band
.
Fo
r
t
he
WF
a
lgorit
hm
,
t
he
gain
f
unct
io
n
g(k)
is
exp
res
sed
with
a
pri
or
i
SN
R
k
.
The
k
is
est
i
m
at
ed
from
the
pa
st
a
nd
present
est
im
a
tes
of
k
as
a
wei
ghte
d
com
bin
at
ion.
The
ˆ
()
k
m
of
t
he
prese
nt
fr
am
e
m
is est
im
at
ed
as foll
ows:
()
1
k
k
gk
(3)
2
2
22
ˆ
(
1
)
()
ˆ
(
)
(
1
)
m
a
x
1
,
0
(
1
)
(
)
k
k
k
kk
Xm
Ym
m
N
m
N
m
(4)
wh
e
re
α
de
note
s
a
s
m
oo
thin
g
co
ns
ta
nt
(
α
=
0.98),
a
nd
ˆ
(
1
)
k
Xm
,
()
k
Ym
an
d
()
k
Nm
are
the
sp
ect
ru
m
of
the
enh
a
nce
d
s
peech
si
gn
al
at
the
past
fr
am
e
m
-
1,
the
sp
ect
r
um
of
the
no
is
y
sp
eech
an
d
the
noise
sign
al
at
the
pr
ese
nt
fr
am
e
m
, res
pecti
vely
.
Figure
1.
Bl
oc
k diag
ram
o
f
a
CI sim
ulati
on
b
ase
d on sp
eec
h
e
nh
a
ncem
ent
E
n
h
a
n
c
e
d
S
p
e
e
c
h
P
r
e
-
e
m
p
h
a
s
i
s
C
h
1
:
S
u
m
o
f
m
a
g
n
i
t
u
d
e
s
q
u
a
r
e
n
-
of
-
m
c
h
a
n
n
e
l
m
a
x
i
m
a
s
e
l
e
c
t
i
o
n
N
o
i
s
y
S
p
e
e
c
h
F
F
T
F
i
l
t
e
r
b
a
n
k
C
h
2
:
S
u
m
o
f
m
a
g
n
i
t
u
d
e
s
q
u
a
r
e
C
h
22
:
S
u
m
o
f
m
a
g
n
i
t
u
d
e
s
q
u
a
r
e
S
p
e
e
c
h
E
n
h
a
n
c
e
m
e
n
t
A
l
g
o
r
i
t
h
m
s
V
o
c
o
d
e
d
s
p
e
e
c
h
C
h
1
:
F
i
l
t
e
r
e
d
n
o
i
s
e
C
h
2
:
F
i
l
t
e
r
e
d
n
o
i
s
e
C
h
22
:
F
i
l
t
e
r
e
d
n
o
i
s
e
C
I
S
y
s
t
e
m
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Sp
eec
h
i
ntell
igibil
it
y enhan
ce
men
t f
or
t
ha
i
-
s
pea
ki
ng co
c
hle
ar
im
pl
an
t l
ist
eners
(
Siri
porn
Da
c
hasil
aruk
)
869
2.2
.
S
ound
C
od
in
g
S
tr
at
e
gies
C
omm
ercial
m
anu
fact
ur
e
rs
of
CI
dev
ic
e
s
pro
pose
m
a
ny
sou
nd
co
di
ng
st
rategies
su
c
h
as
t
he
Con
ti
nu
ou
s
I
nt
erleave
d
Sam
pling
(CIS)
stra
te
gy
and
the
Adva
nced
Co
m
bin
at
ion
Enc
od
e
r
(
ACE)
strat
egy.
Gen
e
rall
y,
the
CIS
strat
egy
is
pr
op
os
e
d
in
CI
de
vices
m
ade
by
al
l
m
anu
fact
ur
e
rs,
and
it
has
di
ff
e
ren
t
i
m
ple
m
entat
io
ns
de
pe
ndin
g
on
t
he
m
anu
fa
ct
ur
er
.
T
he
Co
chlear
C
om
pan
y
,
pro
duci
ng
Nu
cl
e
us
CI
de
vices
,
offer
s
both
the
CIS
an
d
AC
E
strat
eg
ie
s
.
T
he
diff
e
re
nce
be
tween
the
se
str
at
egies
li
es
in
the
cha
nn
el
m
axim
a
sel
ect
ion
sta
ge
.
Cha
nnel
m
axi
m
a
sel
ect
ion
is
pe
rfor
m
ed
in
the
ACE
but
no
t
i
n
t
he
CI
S.
A
fe
w
st
ud
ie
s
ha
ve
rev
eal
e
d
that
m
os
t
Nu
cl
eus
CI
li
ste
ner
s
pr
efer
red
th
e
AC
E
ov
e
r
the
CI
S
,
an
d
the
m
ean
scor
es
us
in
g
t
he
ACE
sh
ow
sig
nifica
ntly
hig
he
r
sp
e
ech
intel
li
gib
il
it
y
than
th
os
e
us
in
g
the
CIS
[
21
]
,
[
22
]
.
Mo
reov
er
,
the
pre
ferre
d
strat
egy co
rr
e
s
ponded
w
it
h t
he
sp
eec
h
i
ntell
i
gib
il
it
y
ou
tc
ome
.
The
ACE
st
rategy
can
be
de
scribe
d
as
an
n
-
of
-
m
strat
e
gy
[
23
]
.
T
he
s
pe
ech
si
gn
al
is
deco
m
po
se
d
into
m
c
hannel
s
relat
ed
to
the
nu
m
ber
of
el
ect
rodes,
but
on
ly
the
n
chann
el
s
with
m
axim
u
m
a
m
pli
tude
s
are
sel
ect
ed
f
or
si
m
ul
ta
ne
ous
sti
m
ula
ti
on
.
T
he
con
ce
pt
of
the
ACE
strat
e
gy
i
s
to
i
ncr
ease
te
m
po
ral
res
olu
ti
on
an
d
reduce
re
dund
ant
inform
at
io
n
in
sp
eec
h
.
The
m
os
t
i
m
po
rtant
cha
nn
el
s
co
ntaini
ng
im
po
rtant
sp
eech
inf
or
m
at
ion
ca
n
be
update
d
m
or
e
fr
e
qu
e
ntly
by
rem
ov
in
g
the
le
ss
sig
ni
fic
ant
c
hannel
s
[
24
]
.
T
his
st
rategy
m
ay
red
uce
th
e
overall
S
NR
le
vel
an
d
pr
es
um
ably
red
uce
s
cha
nnel
inter
act
ion
[
25
]
.
A
dd
it
io
nally
,
the
powe
r
consum
ption
of
el
ect
rical
sti
m
ula
ti
on
ca
n b
e d
ec
reased
, a
nd this
m
ay
leng
then batt
ery li
f
e f
or
C
I device
s
[
26
]
.
As
show
n
in
F
ig
ure
1
,
in
pa
rt
of
the
CI
syst
e
m
the
enh
anc
ed
sp
eec
h
was
filt
ered
by
a
pre
-
em
hasis
filt
er
to
am
plify
the
hi
gh
-
f
re
quency
c
om
po
ne
nts
of
sp
eec
h
inf
or
m
at
ion
.
T
hen,
the
f
ram
e
-
by
-
f
ram
e
pr
oc
essin
g
of
12
8
sam
ples
with
a
n
overla
p
of
75%
was
app
li
ed
to
the
pr
e
-
em
ph
asi
zed
s
peec
h
.
The
gr
eat
er
the
ove
rlap
of
the
fr
am
e
,
th
e
higher
t
he
channel
sti
m
u
la
ti
on
rates.
Each
f
r
am
e
of
the
pr
e
-
em
ph
asi
zed
s
peec
h
was
deco
m
po
se
d
usi
ng
the
fast
-
Four
ie
r
tra
ns
f
or
m
(F
FT
)
i
nto
un
i
form
fr
e
qu
e
ncy
ba
nd
s
(
128
bin
s
),
with
th
e
fr
e
qu
e
ncy
ba
nd
of
eac
h
bin
a
t
12
5
Hz
.
O
nly
the
first
64
bi
ns
we
re
us
e
d
,
to
ge
ner
at
e
a
frequ
e
ncy
res
olut
ion
of
22
c
hanne
ls
.
T
he
pow
er
s
of
con
sec
utive
bi
ns
wer
e
s
umm
ed
within
frequ
e
ncy
ra
ng
e
s
sp
eci
fied
in
the
CI
syst
e
m
.
The
c
utoff
fr
e
quenc
ie
s
of
the
22
channels
wer
e
187.5
,
312.5,
437.5
,
562.5,
687.5
,
812.5,
937.
5,
1062.
5,
1187.
5,
1312.
5,
15
62.
5,
1812.
5,
2062.
5,
23
12.5,
2687.
5,
30
62.
5,
3562.
5,
40
62.
5,
4687.
5,
5312.
5,
6062.
5,
69
37.
5,
an
d
79
37.5
Hz.
Af
te
r
th
at
,
the
1
2
en
ve
lop
e
cha
nnel
s
with
the
la
rgest
a
m
plit
ud
es
wer
e
m
od
ulate
d
by
wh
it
e
noise
with
the
sa
m
e
cutof
f
f
re
qu
e
ncie
s
as
the
FFT
filt
er
ban
k
.
Fina
ll
y,
v
ocoded
s
peech
was
sy
nth
esi
ze
d
by
su
m
m
ing
al
l
the
sel
ect
ed
cha
nn
el
s
of
th
e
m
od
ulate
d
si
gn
al
.
Th
e
voc
oded
sp
eec
h
wa
s
then
pr
ese
nted
to N
H
li
ste
ners
f
or
te
sti
ng
.
3.
PERF
O
R
MANC
E E
V
ALU
ATIO
N
All
the
Thai
w
ords
i
n
t
his
st
ud
y
we
re
from
a
c
orp
us
wh
ic
h
is
c
omm
on
ly
us
e
d
i
n
cl
ini
cal
pr
act
ic
e
with HI list
ene
rs.
In this st
udy t
he
Thai wor
d
te
st was d
i
vid
ed
i
nto
8
li
sts of
m
on
osy
ll
abic words an
d
8
l
ist
s o
f
bisyl
la
bic
w
ords
[
27
]
,
[
28
]
.
Each
li
st
had
25
wor
ds
a
nd
th
e
total
w
ords
a
re
400
w
ords
.
Af
te
r
the
w
ord
s
we
re
sel
ect
ed
from
t
he
co
rpus,
al
l
the
rec
orded
w
ords
we
re
co
rrup
te
d
by
sp
eec
h
-
s
ha
ped
no
ise
(S
SN)
at
SN
R
le
vels
of
0
an
d
5
dB,
and
babble
noise
(BBN
)
at
SN
R
le
vels
of
5
an
d
10
dB.
The
le
vels
of
SN
R
we
re
caref
ully
chosen
to
a
void
flo
or
an
d
cei
li
ng
ef
fects
[
7
]
,
[
29
]
,
as
well
as
par
ti
cula
rly
no
isy
an
d
e
nhanced
m
onos
yl
la
bic
words.
The
n,
t
he
noisy
w
ord
s
we
re
proces
s
ed
us
in
g
the
M
BSS
a
nd
WF
a
lgorit
hm
s.
The
en
ha
nced
an
d
no
isy
words
we
re
processe
d
usi
ng
the
ACE
st
rategy
to
pr
oduce
the
voco
de
d
s
peech
sig
nals.
The
vo
c
oded
s
peech
sign
al
s
we
re
presente
d
to
N
H
sub
j
ect
s
in
a
total
of
24
c
onditi
ons
([2
SE
al
gorithm
s
+
1
unproces
sed
S
E
al
gorithm
]
2
word ty
pes
2 n
oise ty
pes
2
S
NR le
vels).
T
we
nty
NH
subj
ect
s
(
14
m
a
les,
6
fem
al
es,
a
ge
ra
ng
e
f
ro
m
20
to 4
0
a
nd
m
ean
age
26
)
pa
rtic
ipate
d
i
n
this
ex
per
im
ent
.
All
sub
j
ect
s
wer
e
native
sp
eakers
of
T
hai
,
an
d
wer
e
un
dergr
a
duat
e
st
ud
e
nts
a
nd
sta
ff
at
a
Thai
public
unive
rsity
.
Otos
cop
y
was
unde
rtake
n
f
or
al
l
subj
ect
s
to
c
heck
f
or
any
abno
rm
aliti
es
in
thei
r
m
idd
le
ears
.
T
hen,
a
ll
s
ubj
ect
s
un
der
t
ook
a
pure
t
on
e
a
ud
i
ogram
te
st
to
c
onfirm
that
the
y
ha
d
NH
th
re
sh
ol
ds
(<
25
dB
H
L,
betwee
n
0.
25
and
8
k
Hz).
A
ll
su
bject
s
we
r
e
paid
f
or
t
heir
pa
rtic
ipati
on
and
they
al
l
sign
e
d
a
consent
form
.
This e
xp
e
rim
e
nt w
a
s a
pprove
d by the
Ethic
s
Com
m
it
te
e o
f t
he
unive
rsity
.
The
vo
c
oded
s
peech
sig
nals
wer
e
pr
e
sente
d
u
sin
g
a
la
pt
op,
un
il
at
erall
y
thr
ough
a
hea
dpho
ne.
T
he
su
bject
s
us
ed
on
ly
one
prefe
rr
e
d
ear,
t
he
one
that
was
m
os
t
com
fo
rtabl
e
fo
r
t
hem
,
to
li
ste
n
to
the
voco
de
d
sp
eec
h
sig
nals
in
al
l
te
ste
d
c
onditi
ons.
The
volum
es
of
t
he
vo
c
oded
s
peec
h
sig
nals
we
re
cal
ibrat
ed
t
o
be
at
a
com
fo
rtable
co
nv
e
rsional
le
ve
l.
Each
s
ubj
e
ct
was
assesse
d
in
a
total
of
24
c
onditi
ons
ov
e
r
tw
o
sessi
on
s
on
separ
at
e
days
(
12
co
ndit
ion
s
per
s
essio
n,
on
e
sessio
n
pe
r
da
y),
with
a
bre
ak
of
at
le
ast
one
w
eek
betwe
en
th
e
two
s
essio
ns
t
o av
oid
lea
rn
i
ng eff
ect
s
. Testin
g
la
ste
d ap
pro
xim
at
ely o
ne h
our
in
eac
h
se
ssion.
Be
fore
the
act
ual
te
sts
wer
e
carrie
d
out,
t
he
resea
rc
her
s
offer
e
d
trai
nin
g
te
sts
t
o
en
su
re
t
hat
th
e
su
bject
s
cl
earl
y
under
st
ood
how
t
o
do
the
t
est
s.
I
n
the
trai
ning
te
sts
the
s
ubj
ect
s
we
re
a
sk
e
d
to
li
ste
n
t
o
both
no
isy
a
nd
en
ha
nced
w
ords
i
n
al
l
co
nd
it
io
ns
durin
g
a
5
-
m
inu
te
te
st,
to
fam
il
ia
rise
t
hem
sel
ves
with
the
vo
c
oded
s
peec
h
sig
nals.
The
y
wer
e
trai
ne
d
in
bo
th
ses
sio
ns
be
fore
they
under
t
ook
the
act
ual
te
sts.
In
the
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
1
3
, N
o.
3
,
Ma
rc
h
201
9
:
866
–
8
7
5
870
act
ual
te
sts,
the
su
bject
s
li
ste
ned
t
o
the
w
ords
an
d
wrote
them
do
w
n
on
their
pa
per
s
.
T
hey
co
uld
gue
ss
the
words
if
they
wer
e
uncertai
n.
Sub
j
ect
s
we
r
e
scor
e
d
base
d
on
how
m
any
wor
ds
they
id
entifi
ed
c
orrec
tl
y
in
each
li
st.
O
ne
word
li
st
wa
s
a
dm
inist
ered
pe
r
co
ndit
ion
,
a
nd
each
li
st
co
nt
ai
ned
on
e
hundre
d
word
s
.
N
o
w
ord
li
st
was
rep
eat
ed
acr
os
s
the
c
onditi
ons
in
ea
ch
sessio
n
f
or
each
sub
j
ect
.
T
he
li
st
-
to
-
c
ondi
ti
on
m
app
ing
a
nd
t
he
order o
f
te
ste
d
conditi
o
ns i
n e
ach sessi
on w
e
re r
a
ndom
iz
ed
for
eac
h
s
ubj
ec
t. Th
e
sub
j
ect
s
did
not k
now wh
ic
h
conditi
on
w
ould
be
te
ste
d
a
nd
w
hat
the
te
ste
d
c
onditi
on
s
w
ou
l
d
be.
In
ord
er
to
a
void
li
ste
ning
fati
gu
e
wh
i
c
h
m
ay
aff
ect
per
f
or
m
ance,
the
s
ubj
ect
s
we
re
gi
ven
a b
rea
k
e
ve
ry
20
m
inu
te
s
durin
g
the
act
ual
te
st,
or
w
he
nev
e
r
they
r
e
qu
ire
d a
r
est
.
The
sc
or
e
s
of
correct
w
ords
wer
e
a
ve
rag
e
d
base
d
on
t
he
per
ce
ntage
of
words
ide
ntifie
d
c
orrectl
y.
The
n
t
he
sco
r
es
wer
e
sta
ti
sti
cal
ly
analy
ze
d
us
i
ng
S
PSS
so
ft
war
e
.
A
n
analy
sis
of
va
ri
ance
(
ANO
V
A)
with
rep
eat
e
d
m
eas
ur
es
was
c
ar
ried
ou
t
to
in
ve
sti
gate
the
di
ff
e
ren
ces
bet
w
een
m
ean
sco
res
in
te
rm
s
of
SE
al
gorithm
s,
S
NR
le
vels,
noi
se
ty
pes,
and
word
ty
pes
.
A
po
st
hoc
Bo
nfer
roni
-
c
orrect
ed
te
st
with
a
m
ul
ti
ple
paire
d
com
par
ison
was
em
pl
oyed
to
ex
am
i
ne
the
in
div
id
ual
relat
ion
s
hi
ps
bet
ween
th
e
m
ean
scor
es
in
each
conditi
on.
4.
RESU
LT
S
AND DI
SCUS
S
ION
The
m
ean
per
c
entage
c
orrect
scor
e
s
of
20
N
H
sub
j
ect
s
in
24
c
onditi
ons
are
show
n
in
F
igure
2.
All
the
co
ndit
ion
s
co
ns
ist
ed
of
no
isy
words
a
nd
en
ha
nced
words
due
t
o
diff
e
re
nt
SE
al
gorithm
s,
SN
R
le
vels,
no
ise
ty
pes
,
a
nd
w
ord
ty
pes
.
The
re
su
lt
s
of
no
isy
a
nd
e
nha
nced
w
ords
s
howe
d
c
onsider
ably
increase
d
m
ean
scor
e
s as S
NR
le
vels incr
ease
d
at
the sam
e n
oise ty
pe
an
d
word
ty
pe
.
F
or the n
oisy w
or
ds
, th
e m
ean sco
res o
f
the
bisyl
la
bic
w
ords
with
BB
N
inc
rease
d
sli
ghtl
y
as
SN
R
le
vels
in
creased
.
T
he
m
ean
sco
res
s
howe
d
extrem
el
y
hig
h
intel
li
gib
il
ity
at
5
and
10
dB
SN
R.
At
5
dB
SN
R
the
m
ea
n
sco
res
f
or
no
isy
wo
r
ds
with
SSN
wer
e
sli
ghtl
y
hi
gh
e
r
tha
n
th
ose
for
noisy
w
ords
with
BB
N
for
m
on
os
yl
la
bi
c
word
s
,
an
d
alm
os
t
the
sam
e
for
bisyl
la
bic
w
ords.
The
e
nh
a
nc
ed
wor
ds
rev
ea
le
d
hi
gher
m
ean
sc
or
e
s
tha
n
t
he
no
isy
w
ords
in
m
os
t
co
nd
it
ion
s
.
The
WF
re
flec
te
d
co
ns
ide
ra
bl
y
bette
r
perform
ance
i
m
pr
ovem
ent
than
the
MB
SS
in
a
l
m
os
t
al
l
con
di
ti
on
s,
excep
t
f
or
m
onos
yl
la
bic wor
ds at
5 dB
SN
R
of BB
N.
Fo
r
the
m
on
osy
ll
abic
word
s
,
the
WF
ref
le
ct
ed
a
great
er
i
ntell
igibil
it
y
i
m
p
rovem
ent
than
the
MB
SS
,
especial
ly
f
or
t
he
co
ndit
ion
of
SS
N
at
0
a
nd
5
dB SNR. Bot
h
al
gorithm
s yielded
alm
os
t t
he
sa
m
e
m
ean s
cor
es
for
BB
N
at 5
a
nd 10 dB S
NR.
How
e
ver, th
e
enh
a
nce
d wor
ds wit
h b
oth
al
gorithm
s show
e
d
lo
we
r
m
ean sco
res
than
the
no
isy
word
s
for
BB
N
at
10
dB
SN
R.
F
or
th
e
bisyl
la
bic
wo
r
ds
,
the
WF
al
go
rithm
sh
ow
e
d
a
no
ti
ceably
hi
gher
im
pr
ov
em
ent
than
the
MB
SS,
esp
eci
al
ly
in
the
conditi
on
of
SSN
at
0
and
5
dB
S
NR.
Both
al
gorithm
s
il
lustrate
d
a
sli
gh
t
i
m
pr
ov
em
ent
for
BB
N.
The
ov
e
rall
m
ean
scor
es
f
or
the
bi
syl
la
bic
wo
rds
wer
e
higher
tha
n
th
ose
for t
he
m
onos
yl
la
bic wor
ds.
A
tw
o
-
way
A
NOV
A
with
re
peated
m
easure
s
was
us
e
d
to
exp
l
or
e
t
he
two
facto
rs
of
SE
al
gorithm
and
S
NR
le
vel
.
F
or
t
he
m
onosy
ll
abic
words
with
S
SN,
the
sta
ti
sti
cal
anal
ysi
s
rev
eal
e
d
a
sig
nificant
e
ffec
t
of
SE
al
gorit
hm
[
F
(2,38
)=17.
60,
p
<
0.000
5]
an
d
S
NR
le
vel
[
F
(1,19
)=68.
73,
p
<
0.000
5].
P
os
t
ho
c
te
sts
f
or
S
E
al
gorith
m
s ind
ic
at
ed
that t
he WF
produce
d
s
ign
ific
a
ntly
h
ig
her
i
ntell
igibil
i
ty
sco
res
tha
n
the noisy
words
, and
sh
owe
d
sig
nifi
cantl
y
bette
r
pe
rfor
m
ance
than
the
MB
SS
at
0
dB
SN
R
.
Post
ho
c
te
sts
of
SN
R
le
vels
re
ve
al
ed
that
an
inc
reas
ed
S
NR
le
vel
pro
vid
e
d
sig
ni
fican
tl
y
i
m
pr
oved
i
ntell
igibil
it
y
scor
es
f
or
no
isy
words
a
nd
t
he
MB
SS
(
p
<0
.0005
),
but
this
diff
e
re
nce
was
not
sta
ti
sti
cally
sign
ific
ant
for
the
WF.
F
or
t
he
m
onos
yl
la
bic
words
with
B
BN,
the
re
were
sign
i
ficant
e
f
fects
of
S
NR
l
evel
[
F
(
1,19)=
105.7
3,
p
<
0.0
005]
an
d
a
sig
nificant
interact
ion
ef
fe
ct
betwee
n
S
E
al
gorithm
and
SN
R
le
ve
l
[
F
(
2,38)=
7.95,
p
<
0.05
]
.
Post
hoc
te
sts
of
S
NR
l
evels
ind
ic
at
ed
that
the
noisy
words
at
10
dB
SN
R
sh
owe
d
sign
i
f
ic
antly
hig
her
i
ntell
igibil
it
y
sc
or
es
tha
n
th
os
e
at
5
dB
SN
R.
F
or
the
bisyl
la
bi
c
words
with
S
SN
,
the
sta
ti
sti
cal
analy
sis
i
nd
ic
at
ed
a
sig
nificant
ef
fect
of
SE
al
gorithm
[
F
(2,38)
=
50.
28,
p
<
0.000
5],
a
sig
nificant
ef
fect
of
S
NR
le
vel
[
F
(
1,1
9)
=
268.57,
p
<
0.0
005]
and
a
sign
ific
a
nt
interact
ion
e
ff
ect
be
tween
SE
al
gorithm
and
S
N
R
le
vel
[
F
(2,38)
=
12.
27,
p
<
0.000
5].
Po
st
hoc
te
sts
of
SE
al
gorith
m
s
and
SN
R
l
evels
ind
ic
at
e
d
that
the
m
ultip
le
paire
d
com
par
is
on
yi
el
de
d
the
sam
e
resu
lt
s
as
the m
on
os
yl
la
bi
c w
or
ds
with
SSN.
The
relat
ive
diff
e
ren
ce
of
m
ean
scor
es
betwee
n
no
isy
wor
ds
and
en
ha
nc
ed
w
ords
is
sh
ow
n
in
Fig
ure
3.
The
re
was
a
co
ns
ide
rab
le
increase
i
n
int
el
li
gib
il
it
y
i
m
pr
ovem
ent
fo
r
e
nh
a
nce
d
words
with
SE
al
gorit
hm
s
f
or
both
w
ord
ty
pes
w
hich
wer
e
eval
ua
te
d
in
the
s
a
m
e
te
ste
d
co
nd
it
io
ns
.
E
xce
pt
f
or
m
on
os
yl
la
bic
words
in
BB
N
at
10
dB
SN
R
,
intel
li
gib
il
it
y
perform
ance
de
creased
.
T
he
WF
ex
hib
it
ed
a
hi
gh
i
m
pr
ovem
ent
bu
t
the
MB
SS
s
howe
d
a
low
i
m
pr
ov
em
ent
in
bo
t
h
w
ord
ty
pe
s.
I
n
te
rm
s
of
the
overall
res
ult
of
intel
li
gib
il
it
y
i
m
pr
ov
em
ent
i
n
the
w
ho
le
c
onditi
on,
SE
a
lgorit
hm
s
i
m
pr
ov
e
d
by
ap
pro
xim
a
te
ly
3%
f
or
t
he
MB
SS
and
by
12%
f
or
the
W
F.
A
tre
nd
of
decr
ease
d
intel
li
gib
il
it
y
per
form
ance
with
increase
d
S
NR
le
vels
was
f
ound
in
bo
t
h
SE
al
gori
thm
s.
The
WF
yi
el
ded
bette
r
intel
li
gib
il
ity
than
the
MB
S
S
at
the
sa
m
e
SN
R
le
vels.
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Sp
eec
h
i
ntell
igibil
it
y enhan
ce
men
t f
or
t
ha
i
-
s
pea
ki
ng co
c
hle
ar
im
pl
an
t l
ist
eners
(
Siri
porn
Da
c
hasil
aruk
)
871
(a) M
on
os
yl
la
bic wo
rd
s
(b)
Bi
syl
la
bic wor
ds
Figure
2
.
The
m
ean p
e
rcen
ta
ge
c
orrect sc
ores of
20 no
rm
a
l
-
hear
i
ng li
ste
ne
rs for the
nois
y words (
N
W)
and
the enha
nce
d words
with the
MB
SS
a
nd
WF
al
gorithm
s are
shown
f
or spee
ch
-
s
ha
ped noise (
SS
N)
at
0, 5 dB
SN
R a
nd
babbl
e
noise
(
BB
N)
at
5
a
nd 10 dB
SN
R.
The
p
l
us (+)
d
e
note
s th
at
the m
ean p
er
centage
co
rr
ect
scor
e
s ar
e
sig
ni
ficantl
y hig
her than
th
os
e at t
he
lo
we
r
S
NR
le
vel (
p
<0
.00
05)
. An ast
erisk
(*) den
otes t
ha
t t
he
m
ean p
e
rcen
ta
ge
c
orrect sc
ores are
sig
nifica
ntly
h
ig
her tha
n
th
os
e
at the s
a
m
e SN
R l
evel
; **
p
<
0.05,
***
p
<
0.000
5.
The
er
ror bars
in
dicat
e the s
ta
nd
a
rd
dev
ia
ti
on
of th
e sco
res
(a) M
on
os
yl
la
bic wo
rd
s
(b)
Bi
syl
la
bic wo
rd
s
Figure
3
.
Re
la
ti
ve
di
ff
e
ren
ce
s
in
intel
li
gib
il
it
y scor
e
s
betwe
en no
isy
a
nd e
nh
a
nce
d wor
ds wit
h
t
he
MB
S
S and
WF
al
gorithm
s
are
s
how
n f
or
sp
eec
h
-
s
ha
ped
no
ise
(
S
SN) at
0,
5 dB
SN
R,
a
nd for ba
bble
noise (BB
N) at
5,
10 d
B
SN
R
. Posit
ive
nu
m
bers in
dicat
e an
i
nc
reased
intel
li
gi
bili
ty
p
erfor
m
ance
wh
e
reas
ne
gative
nu
m
be
rs
represe
nt a
d
ec
reased
intel
li
gib
il
it
y perfor
m
ance.
The
er
ror bars
r
e
fer t
o
t
he
s
ta
ndar
d dev
i
at
ion
of the s
co
res
Exam
ples
of
t
he
el
ect
rod
ogr
a
m
s
for
of
the
m
on
osy
ll
abic
wor
d
“/
yā
am
/
”
an
d
the
bisyl
la
bic
wor
d
“/
nâe
-
nōn/”
ar
e
pr
ese
nted
in
Figure
4.
As
c
an
be
cl
early
s
een
f
ro
m
the
el
ect
ro
do
gr
am
s,
the
MB
SS
an
d
W
F
al
gorithm
s
can
reduce
no
ise
.
Howe
ver,
the
con
te
nt
of
s
pe
ech
in
form
at
i
on
was
noti
ceably
reduce
d
f
or
t
he
enh
a
nce
d
wor
ds
with
the
M
BSS,
a
nd
resi
dual
noise
sti
ll
r
e
m
ai
ned
in
s
om
e
segm
ents
of
the
e
nhance
d
wor
ds
with
the
WF.
The
WF
preser
ved
m
or
e
sp
ee
ch
in
f
or
m
at
ion
at
low
f
req
ue
nc
ie
s
rangin
g
from
18
7
to
563
H
z
or
in
the
el
ect
ro
de
channels
bet
ween
2
0
a
nd
22.
S
uch
p
rese
r
vation
of
in
for
m
at
ion
by
the
WF
in
noisy
conditi
ons
echo
e
d
a
sig
nif
ic
ant intel
li
gib
il
it
y im
pr
ov
em
e
nt.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
1
3
, N
o.
3
,
Ma
rc
h
201
9
:
866
–
8
7
5
872
Figure
4
.
E
le
ct
rod
ogram
s o
f
t
he
m
on
osy
ll
abic w
ord
“/
yā
am
/”
(guard
)
in
th
e left
co
l
um
n
and the
bisyl
la
bi
c
word “/
nâe
-
nō
n/” (
ce
rtai
n) in
the r
i
gh
t
c
olu
m
n.
The
f
irst
r
ow s
hows
the cl
ean
wor
ds
. T
he
seco
nd ro
w
represe
nts the no
isy
wor
d
s at
5 dB S
NR s
pee
ch
-
s
ha
ped nois
e f
or
t
he
m
on
osy
ll
abic word a
nd 5 dB
SN
R
babble
no
ise
fo
r
the
b
isy
ll
abic
word. The
thir
d
a
nd fo
ur
th
ro
ws
s
how
the e
nhance
d
w
ords
with the
MB
S
S
an
d
WF
al
gorithm
s
, r
es
pecti
vely
The
outc
om
es
of
the
prese
nt
stud
y
ha
ve
de
m
on
strat
ed
that
the
re
is
po
te
ntial
for
sin
gle
-
cha
nnel
SE
al
gorithm
s
(i.e.
MB
SS
a
nd
W
F
)
un
der
var
i
ou
s
noisy
conditi
ons
to
be
a
ble
to
i
m
pr
ov
e
intel
li
gib
il
it
y
perform
ance
f
or
Thai
CI
li
st
ener
s
.
S
om
e
of
the
res
ults
of
the
prese
nt
s
tud
y
wer
e
co
nsi
ste
nt
with
th
os
e
of
Chen
et
al
.
[
29
]
.
This
stud
y
fo
un
d
that
the
best
intel
li
gib
il
it
y
per
fo
rm
anc
e
was
achieve
d
by
the
W
F
f
or
both
sta
ti
on
ary
nois
e
(e.g.
SSN)
a
nd
non
-
sta
ti
onary
no
ise
(e
.g.
BB
N)
.
Be
cau
s
e
of
the
fact
th
at
the
W
F
perf
or
m
s
a
trade
-
off
betw
een
ef
fecti
ve
noise
re
duct
ion
and
sp
eec
h
dis
tortio
n
[
30
]
,
th
e
W
F
en
ha
nce
d
s
peech
m
ay
con
ta
in
m
or
e
residu
al
no
ise
but
al
so
m
or
e
pr
eser
ved
le
xical
in
form
ation
.
T
he
m
ajo
rity
of
the
pr
ese
rv
e
d
le
xica
l
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Sp
eec
h
i
ntell
igibil
it
y enhan
ce
men
t f
or
t
ha
i
-
s
pea
ki
ng co
c
hle
ar
im
pl
an
t l
ist
eners
(
Siri
porn
Da
c
hasil
aruk
)
873
inf
or
m
at
ion
ha
s
lo
w
-
fr
e
quenc
y
com
po
ne
n
ts,
w
hich
are
im
portant
i
nfor
m
at
ion
for
sp
eec
h
unde
rstan
ding
[
4
]
.
No
t
s
urpr
isi
ngl
y,
the
W
F
en
ha
nced
s
peec
h
e
xh
i
bits
higher
word
recog
niti
on
sc
or
e
s.
F
or
the
MB
SS,
the
no
ise
sp
ect
r
um
natural
ly
aff
ect
s
the
sp
eec
h
s
pectr
um
acro
ss
var
i
ou
s
f
reque
ncy
bands
.
T
he
no
i
se
sp
ect
ru
m
m
ay
be
ov
e
r
-
e
stim
at
ed
and
over
-
s
ubtract
ed
f
ro
m
t
he
s
pectru
m
of
noisy
sp
eec
h
in
eac
h
f
re
qu
e
ncy
ba
nd.
So
m
e
segm
ents
of
le
xical
inform
at
i
on
m
ay
be
agg
ressi
vely
rem
ov
e
d
or
dist
or
t
ed
by
proce
ss
ing
from
the
MB
SS.
The
re
sid
ual
inf
or
m
at
ion
in
enh
a
nce
d
s
pee
ch
is
ins
uffici
ent
to
un
der
sta
nd
t
he
le
xical
m
eaning
.
Th
us,
w
ord
recog
niti
on
f
or
the
MB
SS
en
han
ce
d
s
peech
is
alm
os
t
equ
al
to
or
w
orse
than
that
of
noisy
sp
eec
h
in
so
m
e
conditi
ons.
B
isy
ll
abic
wo
r
ds
rev
eal
sig
ni
ficantl
y
highe
r
intel
li
gib
il
it
y
sco
res
t
han
m
on
os
yl
la
bic
words
in
al
l
te
ste
d
conditi
ons.
Ge
ne
rall
y,
the
bisyl
la
bic
words
ha
d
longer
s
peech
du
r
at
ion
s
an
d
co
nt
ai
ned
m
or
e
lex
ic
al
inf
or
m
at
ion
th
an
th
e
m
on
os
y
ll
abic
w
ords.
Wh
e
n
t
hey
we
re
m
asked
by
n
oise
or
s
om
e
segm
ents
of
l
exical
inf
or
m
at
ion
w
ere
rem
ov
e
d
or
de
grade
d
by
the
SE
al
gorith
m
s
and
CI
c
oding
st
rategy,
t
he
li
ste
ner
s
were
able
to un
der
sta
nd the c
on
te
nt of t
he wor
ds
a
nd guess t
he bisy
ll
abic w
ords
b
et
te
r
tha
n
t
he
m
onos
yl
la
bic wor
ds.
Disguised
wor
ds
w
hich
resu
l
t
fr
om
no
ise
,
SE
al
gorithm
s
and
CI
c
od
i
ng
strat
egies
c
ha
ng
e
s
ounds
and
a
ff
e
ct
intel
li
gib
il
it
y.
The
sound
c
hange
s
of
m
onos
yl
la
bic
w
ords
ca
n
be
cl
assifi
ed
i
nto
si
x
m
ajo
r
t
ypes
.
These
are
c
ons
on
a
nt
inserti
on,
conso
na
nt
repl
ace
m
ent,
con
s
on
a
nt
delet
io
n,
vowel
cha
ng
e
s,
ton
e
c
hanges
,
an
d
oth
e
rs.
Con
s
onant
inse
rtio
n
w
as
f
ound
only
in
t
he
final
posit
ion
,
as
i
n
se
(
sp
oil)
se
ng
(s
ound)
.
H
oweve
r
,
conso
na
nt
re
place
m
ent
and
c
on
s
ona
nt
delet
ion
wer
e
f
ound
in
both
t
he
i
niti
al
and
final
posit
ion
s
,
as
i
n
tam
(foll
ow)
ta
(eye);
pla
(
fish)
la
(do
nk
e
y);
hm
(f
ra
grance,
onio
n)
phm
(thin)
;
fan
(t
oo
t
h)
f
an
g
(list
en)
.
V
owel
cha
ng
es
we
re
rar
el
y
f
ound,
a
s
in
dang
(f
am
ou
s
)
dam
(b
la
ck)
.
To
ne
c
ha
ng
es
we
re
fou
nd
a
s
well
,
as
i
n
klng
(cam
era)
klon
g
(
drum
).
In
te
resti
ngly
,
so
m
e
changes
cannot
be
cat
a
gorized
.
F
or
e
xam
ple,
m
(h
ave
e
nough
f
ood)
kin
(eat
)
s
hows
that
a
n
init
ia
l
con
s
ona
nt
wa
s
in
sert
ed
a
nd
a
c
on
so
na
nt
rep
la
ncem
ent
was
m
ade
(/n/
/
m
/).
Anothe
r
exam
ple
is
m
âk
(m
any)
m
a
(co
m
e),
wh
ic
h
al
so
revea
ls
two
changes
.
On
e
was
a
to
ne
c
ha
ng
e
f
r
om
a
fall
ing
to
ne
to
a
m
id
to
ne.
A
noth
er
was
a
c
on
s
onant
delet
io
n;
/
k/
w
as
delet
ed.
Fo
r
bisyl
la
bic
words,
the
re
pl
ace
m
ent
of
pre
cedin
g
syl
la
bles
and
f
ollo
wi
ng
syl
la
bles
res
ults
in
so
un
d
changes
.
Eit
he
r
pr
ece
di
ng
sy
ll
ables
or
f
ollow
i
ng
syl
la
ble
s
wer
e
re
place
d
by
ne
w
syl
la
bles
w
hich
r
eflect
conso
na
nt
repl
ace
m
ent
an
d
ton
e
c
hanges
,
su
c
h
as
r
o
-
r
t
(
wait
for
the
bus)
l
-
r
t
(
w
heel);
fai
-
fá
(elect
rici
ty
)
fai
-
pà
(for
est
fi
re)
;
l
k
-
ni
(lit
tl
e)
d
k
-
n
i (b
a
by). Anothe
r
c
ha
ng
e
in
the
f
ollo
wing
syl
la
bles
was
a
c
ons
on
a
nt
delet
ion
a
nd
a
t
on
e
c
ha
nge,
as
in
nâ
-
tà
ng
(
window)
nâ
-
ta
(fac
e).
No
ti
ce
that
no
m
at
te
r
wh
at
ty
pe
s
of
s
ound c
hanges
wer
e
foun
d,
t
he
sou
nd
cha
ng
es led t
o new
m
eaning
s
.
In
pri
nci
ple,
Thai
syl
la
ble
sign
al
s
are
com
po
se
d
of
the
init
ia
l
con
sona
nt,
vowel
an
d
fina
l
con
sona
nt
sign
al
s
[
31
]
.
T
he
init
ia
l
cons
on
a
nt
a
nd
vow
el
sign
al
s
c
on
t
ai
n
ei
ther
l
ow
or
high
fr
e
qu
e
ncy
com
ponent
s.
Th
e
final
co
n
s
on
a
nt
sign
al
s
co
ntain
only
low
f
re
qu
e
ncy
com
ponen
ts
.
The
to
ne
is
the
change
of
F
0
co
ntou
rs
acro
ss
the
syl
la
ble.
T
he
si
gn
al
s
f
or
the
th
ree
phone
m
es
are
c
om
bi
ned
by
te
m
po
ral
co
nd
it
io
ns
,
and
the
vo
wel
sign
a
l
act
s
as
the
m
ai
n
syl
la
ble
si
gn
al
[
31
]
.
The
refor
e
,
the
vo
wel
sig
nal
is
the
la
r
gest
c
om
po
nen
t
a
nd
al
ways
do
m
inate
s
the
oth
e
rs.
I
n
oth
e
r
words,
w
hen
com
par
ed
to
the
vo
wels
the
init
ia
l
and
fi
na
l
conso
na
nts
play
m
ini
m
al
r
oles.
The
fi
nd
i
ng
s
r
arely
found
co
nfusa
ble
vowel
s.
This
is
not
s
urpr
isi
ng
si
nce
vowels,
in
general,
are
the
m
os
t
sal
ie
nt
c
om
po
ne
nts
an
d
they
are
the
key
facto
rs
for
identify
in
g
th
e
nu
m
ber
of
s
yl
la
bles
in
each
w
ord
.
On
the
co
ntra
r
y
,
the
stud
y
frequ
e
ntly
fo
un
d
co
nfusa
ble
con
s
ona
nts,
esp
eci
al
ly
con
s
onant
rep
la
cem
e
nt.
A
conso
na
nt
in
t
he
init
ia
l
po
si
ti
on
is
m
or
e
conf
us
able
tha
n
a
con
c
on
a
nt
in
the
final
posit
ion.
This
i
s
no
t
su
r
pr
isi
ng
;
init
ia
l
con
s
on
a
nts
are
al
ways
the
first
ph
on
em
e,
an
d
li
ste
ner
s
need
t
o
pe
rcei
ve
them
earli
er
tha
n
oth
e
r
phonem
es.
Th
us
,
it
is
not
easy
to
rec
ognize
the
se
so
unds
,
ei
t
her
in
processe
d
s
pe
ech
or
noisy
s
peech.
More
ov
e
r,
t
he
com
po
ne
nts
of
phonem
es
m
ay
be
ov
e
rla
pp
i
ng.
F
or
ins
ta
nce,
the
co
m
po
nen
ts
of
a
n
init
ia
l
conso
na
nt
an
d
a
vo
wel
m
ay
be
c
om
bin
ed.
Con
se
quently
,
if
these
com
ponen
ts
are
disto
rted
by
no
ise
or
any
pro
cessi
ng stag
e, this a
ff
ect
s t
he
s
ound a
nd c
auses
t
he
m
eaning
t
o
c
hange
.
In
t
he
pr
ese
nt
stud
y,
t
he
set
s
of
te
ste
d
w
ord
s
ha
d
s
om
e
lim
it
a
ti
on
s
wh
e
n
deali
ng
with
relat
ed
fact
or
var
ia
ti
ons acc
ordi
ng
t
o
SE al
gorithm
s o
r no
is
y env
ir
onm
ent
s,
as in t
he
stu
dies b
y L
i
et
al.
[
19
]
a
nd
C
he
n
et
al.
[
29
]
.
The
pres
ent
corp
us
has
been
em
plo
ye
d
in
a
ud
i
ology
cl
inics
fo
r
ov
er
f
or
ty
ye
ars,
par
ti
cularly
w
it
h
HI
li
ste
ner
s.
As
a
resu
lt
,
f
or
the
Thai
la
nguage
corpu
s
,
exist
ing
s
peech
m
a
te
rial
s
are
inad
equ
at
e
a
nd
no
t
ver
y
pr
act
ic
al
for
ei
ther
cl
inics
or
researc
h
stu
dies.
This
div
e
rsit
y
in
sp
eech
m
at
erial
s
is
an
i
m
po
rtant
issue
.
The
te
st
m
at
eria
ls
of
c
onsona
nts,
vowels,
to
nes
or
word
s
are
m
or
e
su
it
able
for
analy
zi
ng
s
peech
i
nfor
m
ation
to
represe
nt
a
HI
li
ste
ner
’s
pe
r
cepti
on
a
bili
ty
and
re
veal
in
form
ative
resul
ts
abo
ut
the
eff
ect
s
of
fact
or
s
or
par
am
et
ric
var
ia
ti
on
.
T
he
se
nt
ence
te
st
m
at
er
ia
ls
inv
ol
ved
m
any
factor
s
(
e.g
.
m
eanin
g,
con
te
xt,
r
hyth
m
,
e
tc
.)
and
m
ay
be
mo
re
a
pprop
riat
e
fo
r
r
eal
-
li
fe
com
m
un
ic
at
ion
.
T
hu
s
,
the
de
velo
pm
ent
of
Thai
sp
eech
m
at
erial
s
sh
oul
d
be
unde
rtake
n
to
s
upport
m
or
e
eff
ect
ive
evaluati
ons
and
treat
m
ents.
I
n
oth
e
r
w
or
ds
,
a
la
r
ge
co
r
pu
s
of
div
e
rse
s
peec
h
m
at
erial
s
is
nee
ded,
wh
i
ch
will
be
nef
i
t
aud
i
ologica
l
evaluati
on
s
to
in
vestigat
e
sp
ee
c
h
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
1
3
, N
o.
3
,
Ma
rc
h
201
9
:
866
–
8
7
5
874
intel
li
gib
il
it
y no
t o
nly am
on
g HI
li
ste
ne
rs b
ut
also fo
r
CI
li
ste
ner
s.
No
neth
el
ess, ev
e
n wit
h
the
se li
m
it
ati
on
s
i
n
relat
ion
t
o
the
corp
us
, t
he pre
sent stu
dy
is
a
ste
pp
i
ng sto
ne t
o
f
uture st
ud
ie
s w
it
h C
I
li
ste
ne
rs.
Ther
e
rem
ai
ns
an
en
or
m
ou
s
gap
in
knowl
edg
e
a
nd
unde
rstan
ding
of
Thai
sp
eec
h
i
ntell
igibil
ity
a
m
on
g
T
hai
-
s
pe
akin
g
C
I
li
ste
ner
s
.
As
a
res
ul
t,
the
intel
li
gi
bili
ty
of
T
hai
s
peech
s
ounds
s
hould
be
inte
nsi
vely
inv
est
igate
d
in
te
rm
s o
f
Thai l
angua
ge
f
eat
ures (
e.
g.
co
nsona
nts, vowel
s
,
a
nd to
nes),
s
pee
ch
c
od
i
ng strat
egies,
SE
al
gorit
hm
s,
obj
ect
i
ve
a
nd
sub
j
ect
ive
m
easur
em
ents,
a
dv
e
rse
e
nvir
onm
ents
(
e.
g.
no
ise
ty
pes
a
nd
SNR
le
vels),
an
d
li
ste
ner
s
(
both
NH
a
nd
C
I
li
ste
ner
s
).
N
e
w
SE
al
go
rithm
s
sh
ould
be
de
velo
ped
us
in
g
oth
e
r
te
chn
iq
ues
s
uc
h
as
dee
p
m
ac
hin
e
le
ar
ning
[
32
]
,
c
om
pr
essive
sensi
ng
[
33
]
,
[
34
]
an
d
s
o
on.
Thes
e
al
gorithm
s
sh
oul
d
be
s
pec
ific
al
ly
adap
te
d
to
the
a
ud
it
ory
per
ce
ptio
n
of
CI
li
ste
ner
s
t
o
opti
m
a
ll
y
i
m
pro
ve
pe
rfor
m
ance
of
ei
ther
s
peec
h
i
ntell
igibil
it
y
or
qu
al
it
y
.
Ii
n
t
urn,
s
im
il
arit
ies
or
differe
nce
s
in
pe
rcep
ti
on
patte
r
ns
a
nd
cro
ss
-
li
ng
uisti
c
ob
s
er
vations m
ay
then
be pr
operly
app
li
ed
in
m
an
y sy
stem
s an
d rela
te
d
a
reas
for
f
uture
researc
h.
5.
CONCL
US
I
O
N
In
the
pr
ese
nt
stud
y
ab
ou
t
Thai
intel
li
gibi
li
t
y
per
f
or
m
a
nce,
tw
o
sin
gle
-
cha
nnel
SE
al
gorithm
s,
nam
ely
MB
SS
and
WF
al
gori
thm
s,
wer
e
eva
luate
d
by
twe
nt
y
NH
li
ste
ner
s
us
in
g
CI
sim
ulati
on
with
a
noise
-
band
voc
od
e
r.
The
resu
lt
s
r
eveale
d
that
sp
eech
intel
li
gib
il
it
y
per
form
ance
was
im
p
rove
d
for
both
S
E
al
gorithm
s
in
m
os
t
te
ste
d
con
diti
ons,
an
d
was
sta
ti
sti
cal
l
y
sign
ific
antly
i
m
pr
ov
e
d
by
the
W
F
al
gorithm
in
so
m
e
con
diti
ons.
T
he
WF
al
gorithm
per
f
orm
ed
bette
r
tha
n
the
MB
SS
al
gorithm
in
near
ly
al
l
con
diti
o
ns
.
T
his
tren
d
in
ou
tc
om
es
will
be
use
fu
l
inf
orm
ati
on
f
or
Thai
-
s
pe
akin
g
CI
li
ste
ner
s
in
f
uture
inv
est
igati
ons
of
the
sp
eec
h
intel
li
gib
il
it
y
per
form
ance
of
e
xisti
ng
SE
al
gorith
m
s,
and
in
de
ve
lop
in
g
ei
the
r
new
SE
al
gorit
hm
s
or
new sou
nd c
oding
strat
egies
f
or au
ditor
y
pro
stheses i
n
f
utur
e stu
dies.
ACKN
OWLE
DGE
MENTS
This
re
searc
h
was
s
upporte
d
by
gra
nts
f
rom
Nar
esuan
U
niv
e
rsity
.
The
auth
or
s
w
ould
li
ke
to
th
a
nk
al
l
the
su
bj
ect
s
who
par
ti
ci
pated
in
this
stud
y.
F
ur
t
herm
or
e,
we
would
al
so
li
ke
to
than
k
Cha
dar
t
han
Lua
ng
sa
wa
ng,
Sira
pr
a
bh
a
K
aewsr
i,
an
d
P
at
a
m
awad
ee
D
oungta
f
ro
m
the
Dep
a
rtm
ent
of
Otolary
ngology,
Nar
es
ua
n
U
nive
rsity
fo
r
subs
ta
ntial
help
in
colle
ct
ing
the
data,
an
d
Dr.
Ra
tt
inan
Tirav
anitc
hakul
a
nd
Ra
da
Dar
a
from
the
De
par
tm
ent
of
Com
m
un
ic
at
ion
Scie
nce
an
d
Dis
orde
r
s,
Fac
ulty
of
Me
dicine,
M
ahid
ol
Un
i
ver
sit
y, T
ha
il
and
for
th
ei
r
u
se
fu
l
sug
gestion
s
.
REFERE
NCE
S
[1]
Kokkinaki
s
K,
Azimi
B,
Hu
Y
,
Friedl
and
DR.
Single
and
Multi
pl
e
Micropho
ne
Noise
Reduction
Strategie
s
i
n
Cochl
e
ar
Im
pla
n
ts.
Tr
ends
in
Am
pli
ficati
on
.
2012
;16(2):
102
-
16.
[2]
Hu
Y,
Loi
zou
PC
,
Li
N,
Kasturi
K.
Us
e
of
a
sigm
oida
l
-
shape
d
f
unct
ion
for
nois
e
at
t
enuation
in
coc
hl
ea
r
implan
t
s.
Journal
of
the A
cousti
cal Society
of Americ
a
.
200
7;122(4):
128
-
34.
[3]
Qin
MK
,
Oxen
ham
AJ
.
Eff
e
cts
of
sim
ula
te
d
coc
hl
ea
r
-
impl
ant
proc
essing
on
spee
ch
re
ce
pt
i
on
in
f
luc
tu
at
in
g
m
aske
rs.
Journa
l
of
the A
cousti
c
al
Soc
ie
t
y
o
f
Am
eric
a
.
2003;114(
1):446
-
54.
[4]
Loi
zou
PC
.
Spee
ch
proc
essing
in
vocode
r
-
c
ent
r
ic
coc
hl
ea
r
implan
t
s.
In:
Molle
r
AR,
edi
tor
.
Cochlea
r
a
nd
Brai
nstem
Im
pla
nts.
64
.
Ba
sel,
Sw
itzer
la
nd:
Karge
r;
2006.
p.
109
-
43.
[5]
Loi
zou
PC
,
Lob
o
A,
Hu
Y.
Subs
pac
e
al
gor
it
hm
s
for
noise
r
educ
t
i
on
in
co
chlea
r
i
m
pla
nts.
Journal
of
th
e
A
coustica
l
Soci
e
ty
o
f Ameri
ca
.
2005
;118(5):2791
-
3.
[6]
Hu
Y,
Lo
iz
ou
P
C.
A
subs
pace
a
pproa
ch
fo
r
enh
anc
ing
spee
ch
c
orrupte
d
b
y
co
lo
red
noise
.
I
EEE
Signal
Proce
ss
in
g
Lett
ers
.
2002
;9(
7):204
-
6.
[7]
Bolne
r
F,
Goehri
ng
T,
Monagha
n
J,
van
Dijk
B
,
W
oute
rs
J,
Bleec
k
S,
ed
it
ors.
Sp
e
ec
h
enhancemen
t
based
on
n
eura
l
net
works
applie
d
to
coc
h
le
ar
im
plant
cod
ing
strategies
.
Int
ern
ati
onal
Confer
ence
on
Acoustic
s,
S
pee
ch
and
Signa
l
Proce
ss
ing
(ICAS
SP
);
2016;
Shanghai
,
China.
[8]
Scal
ar
t
P,
Viei
r
a
J.
Speec
h
enh
anc
ement
b
ase
d
on
a
priori
signal
to
noise
est
i
m
at
ion.
1996
IE
EE
Inte
rnat
iona
l
Confe
renc
e
on
A
cousti
cs,
Spe
ec
h
,
and
Signa
l Proc
essing,
Conf
ere
n
ce
Proceedi
ngs,
Vol
s 1
-
6
.
1996:6
29
-
32.
[9]
Koning
R,
Madhu
N,
W
oute
rs
J.
Ide
al
Ti
m
e
-
Freque
nc
y
M
asking
Algorit
h
m
s
Le
ad
to
D
iffe
ren
t
Speec
h
Inte
lligibi
li
t
y
an
d
Quali
t
y
in
No
rm
al
-
Hea
ring
an
d
Cochl
ea
r
Im
pla
nt
L
iste
n
ers.
IE
EE
Tr
ansacti
ons
on
Bi
omedic
al
Engi
ne
ering
.
20
15;62(1):
331
-
41.
[10]
Hochbe
rg
I,
Bo
othro
y
d
A
,
W
eis
s
M,
Hell
m
an
S.
Eff
ects
of
noi
se
and
noise
sup
pre
ss
ion
on
spee
ch
per
c
ept
ion
b
y
coc
hl
ea
r
implant users.
Ear
Hear
.
1992;13(4)
:263
-
71.
[11]
W
ei
ss
MR.
Eff
e
ct
s
of
noise
and
noise
red
uc
ti
on
proc
essing
on
t
he
oper
a
ti
on
of
the
Nucl
eus
-
22
coc
hl
ea
r
impl
an
t
proc
essor.
Journal
of
Re
hab
il
i
tat
i
on
Re
search
and
Dev
e
lopment
.
1
993;30(1):
117
-
2
8.
[12]
Yang
LP,
Fu Q
J.
Spect
ral
subtra
c
ti
on
-
base
d
spee
c
h
enha
nce
m
ent
f
or
coc
hlear
implant
pat
i
ent
s in
ba
ckgr
ound
noise.
Journal
of
the A
cousti
cal Society
of Americ
a
.
200
5;117(3):
1001
-
4.
[13]
Gus
ta
fss
on
H,
Nordholm
SE,
Cla
esson
I.
Spect
ra
l
subtrac
t
ion
using
red
uce
d
del
a
y
convol
u
t
ion
and
ada
pt
iv
e
ave
rag
ing.
IEEE
Tr
ansacti
ons on
Speech
and
Aud
io
Proc
essing
.
2
001;9(8):
799
-
80
7.
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Sp
eec
h
i
ntell
igibil
it
y enhan
ce
men
t f
or
t
ha
i
-
s
pea
ki
ng co
c
hle
ar
im
pl
an
t l
ist
eners
(
Siri
porn
Da
c
hasil
aruk
)
875
[14]
Verschuur
C,
L
utman
M,
W
ahat
NH
.
Ev
al
ua
ti
o
n
of
a
non
-
li
ne
a
r
spec
tr
al
subtr
a
ct
ion
no
ise
supp
ression
sche
m
e
i
n
coc
hl
ea
r
implant users.
Co
chl
ear
Implants
Inte
rna
ti
onal
.
2006;7(4)
:193
-
6.
[15]
Loc
kwood
P,
Boud
y
J.
Exp
eri
m
ent
s
with
a
Non
li
ne
ar
Spect
r
al
S
ubtra
c
tor
(Ns
s),
Hidden
Markov
-
Models
and
the
Projec
ti
on
,
for
R
obust
Speec
h
Re
cogni
ti
on
in
Car
s.
Speech
Comm
unic
ati
on
.
1992
;
11(2
-
3):215
-
28.
[16]
K
al
lel
F,
Frikha
M,
Ghorbel
M,
Ben
Ham
ida
A
,
Berger
-
Vac
hon
C.
Dual
-
cha
nne
l
spec
tr
al
subtr
a
ct
ion
al
gori
thms
base
d
spee
ch en
hanc
ement
ded
icate
d
to
a
b
ilate
r
a
l
co
chl
e
ar im
plant.
Appl
i
ed Ac
ou
stic
s
.
2012
;73(1)
:12
-
20.
[17]
Berout
i
M,
Sch
wart
z
R,
Makh
oul
J.
Enha
nc
e
m
ent
of
spee
ch
cor
rupte
d
b
y
a
cousti
c
noise
.
I
CASSP
79
IEEE
Inte
rnational
Co
nfe
renc
e
on
Ac
o
ustic
s,
Sp
eech
a
nd
Signal P
roc
e
ss
ing
.
1979:208
-
11.
[18]
Kam
at
h
S,
Loi
z
ou
P.
A
m
ult
i
-
band
spec
tral
subtraction
m
et
hod
for
enha
nci
ng
s
pee
ch
cor
rupte
d
b
y
col
or
ed
noise
.
2002
IEE
E
In
t
ernati
onal
Conf
ere
nce
on
Ac
o
ustic
s,
Spe
ec
h
,
and
Signal
Pro
ce
ss
ing,
Vo
ls
I
-
IV,
Proc
ee
d
ings
.
2002:4164
-
7.
[19]
Li
JF
,
Yang
L,
Z
hang
JP
,
Yan
Y
H,
Hu Y,
Akagi
M,
et
al.
Com
par
at
iv
e
int
e
ll
ig
ibi
l
i
t
y
in
vesti
g
at
ion of singl
e
-
ch
annel
noise
-
red
uc
ti
on
al
gorit
hm
s
for
Chine
se,
Jap
anes
e,
and
English.
Journal
of
the
Ac
oustic
a
l
Soc
ie
t
y
of
Ame
rica
.
2011;129(5):
329
1
-
301.
[20]
Dac
hasilaruk
S,
Lua
ngsawang
C,
Jantha
r
amin
N,
Kae
ws
ri
S,
Doungta
P,
edit
ors.
Ev
alua
ti
on
of
Thai
spee
c
h
int
elligi
b
il
i
ty
ba
sed
on
noise
re
du
ct
ion
techniqu
es
for
co
chlear
implant
s
.
The
1
0th
Im
age
and
Signal
Proc
essing,
BioMedi
c
al
Eng
ine
er
ing
and
Inf
orm
at
ic
s; 201
7
; S
hangha
i, Chi
na
.
[21]
Kief
er
J,
Hohl
S
,
Sturz
ebeche
r
E
,
Pfennigdorff
T
,
Gs
toe
t
tne
r
W
.
Com
par
ison
of
spee
ch
re
cogni
t
io
n
with
dif
fer
en
t
spee
ch
codi
ng
st
rat
eg
ie
s
(SP
EA
K,
CIS,
and
AC
E)
and
the
ir
relat
ionship
to
te
le
m
et
ri
c
m
ea
sures
of
compound
ac
ti
on
pote
ntials
in
th
e Nucle
us CI 24M
coc
h
lear
impla
n
t
s
y
st
em.
Audi
ol
ogy
.
2001
;40(1):32
-
42.
[22]
Skinner
MW
,
H
olde
n
LK,
W
hit
f
ord
LA,
Plant
K
L,
Ps
ar
ros
C,
Holden
TA.
Spe
ec
h
rec
ognition
wit
h
the
Nucle
us
2
4
SP
EAK,
ACE,
a
nd
CIS spee
ch
c
oding
stra
te
gi
es
in
newl
y
implan
t
ed
adu
lt
s.
Ear
a
nd
Hearing
.
200
2;23(3):
207
-
23.
[23]
Cochl
e
ar.
ACE
™
and
CIS
DS
P Stra
te
gi
es.
La
n
e Cove,
New
Sout
h
W
al
es,
Aus
tra
l
ia
:
2002.
[2
4]
Nogueira
W
,
Bu
chne
r
A,
Le
n
arz
T,
Ed
le
r
B
.
A
ps
y
choacousti
c
"N
ofM"
-
ty
p
e
spee
ch
codi
ng
str
a
te
g
y
for
coc
hl
ear
implant
s.
Eurasip
Journal
on
Ap
pli
ed
Signal P
ro
ce
ss
ing
.
2005
;20
05(18):3044
-
59.
[25]
W
il
son
BS
,
Dor
m
an
MF
.
Cochl
ea
r
implan
ts:
Curre
nt
designs
a
nd
f
uture
poss
ibi
li
t
ie
s.
Journal
of
Re
habi
li
ta
ti
o
n
Re
search
and
D
ev
e
lopment
.
200
8;45(5):
695
-
730.
[26]
Hu
Y,
Loizou
PC
.
A
new
sound
codi
ng
str
ateg
y
for
suppres
sing
noise
in
c
ochl
e
ar
implan
t
s.
Journal
of
th
e
Ac
ousti
cal
So
cie
ty
o
f
Am
eric
a
.
2
008;124(1):
498
-
509.
[2
7]
Kom
al
ara
jun
S
.
Deve
lopment
o
f Thai Spee
ch
Dis
cri
m
ina
t
ion
Mat
eri
a
ls:
Mahido
l U
nive
rsit
y
;
1979
.
[28]
Kangsana
rak
B.
Deve
lopment
o
f Thai Spondee
W
ords for
Spee
ch Audiom
et
r
y
:
Ma
hidol
Univ
ersity
;
1980.
[29]
Chen
F,
Hu
Y
,
Yuan
M.
Ev
al
u
a
ti
on
of
Noise
Re
duct
ion
Methods
for
Sent
ence
Re
cogni
ti
on
b
y
Ma
ndar
in
-
Speak
ing
Cochl
e
ar
Im
pla
n
t
L
iste
ner
s
.
Ear
and
Hearing
.
20
15;36(1):
61
-
71.
[30]
Chen
JD
,
Bene
sty
J,
Huang
Y,
Doclo
S.
New
insight
s
int
o
the
noi
se
red
uct
ion
W
i
ene
r
filter
.
IEE
E
Tr
ansacti
ons
on
Audi
o
Sp
ee
ch
an
d
Lang
uage
Pro
ce
ss
ing
.
2006
;14
(4):1218
-
34.
[31]
The
er
a
-
Um
pon
N,
Chansar
e
ewi
tt
a
y
a
S,
Aueph
anwir
i
y
aku
l
S.
Phonem
e
and
t
onal
ac
c
ent
r
ecogniti
on
for
Th
a
i
spee
ch.
E
xpe
rt
S
yste
ms
wit
h
Applicat
ions
.
2011;3
8(10):13254
-
9.
[32]
Mishra
C,
D.L
.
Gupta.
Dee
p
M
ac
hin
e
Learni
ng
a
nd
Neura
l
Ne
tworks:
An
Ov
erv
ie
w.
In
te
rnat
ional
Journal
o
f
Arti
ficial Int
el
l
ig
enc
e
(
IJA
I)
.
201
7;6:
66
-
73.
[33]
Sulong
A,
Guna
wan
TS,
Kha
li
f
a
OO
,
Kart
iwi
M,
Am
bika
irajah
Eb.
Spe
ec
h
enh
anc
ement
base
d
on
W
ie
ner
f
il
t
er
and
compress
ive sensing.
Indone
sian J
ournal
o
f El
e
ct
rica
l
Eng
in
ee
ring a
nd
Computer
Sc
ie
nc
e
.
2
016;2:
367
-
79.
[34]
Ning
K,
Qin
G
.
Com
pre
ss
ed
s
ensing
spee
ch
s
igna
l
enh
ancem
ent
rese
arc
h
.
Indone
sian
Journal
of
El
e
ct
ri
ca
l
Engi
ne
ering
and
Computer
Sc
ie
n
ce
.
2017;6(1)
:26
-
35.
Evaluation Warning : The document was created with Spire.PDF for Python.