T
E
L
KO
M
N
I
KA
T
e
lec
om
m
u
n
icat
ion
,
Com
p
u
t
i
n
g,
E
lec
t
r
on
ics
an
d
Cont
r
ol
Vol.
18
,
No.
4
,
Augus
t
2020
,
pp.
2193
~
2204
I
S
S
N:
1693
-
6930,
a
c
c
r
e
dit
e
d
F
ir
s
t
G
r
a
de
by
Ke
me
nr
is
tekdikti
,
De
c
r
e
e
No:
21/E
/KP
T
/2018
DO
I
:
10.
12928/
T
E
L
KO
M
NI
KA
.
v18i4.
14215
2193
Jou
r
n
al
h
omepage
:
ht
tp:
//
jour
nal.
uad
.
ac
.
id/
index
.
php/T
E
L
K
OM
N
I
K
A
A
r
ab
ic
d
ig
its sp
e
e
c
h
r
e
c
ogn
iti
on
an
d
sp
e
a
k
e
r
i
d
e
n
t
ific
at
io
n
in
n
oi
sy e
n
vi
r
o
n
m
e
n
t
u
si
n
g a
h
y
b
r
id
m
od
e
l
of
V
Q an
d
GM
M
Abd
e
lk
b
ir
Ou
is
aad
an
e
1
,
S
aid
S
af
i
2
,
M
il
ou
d
F
r
ik
e
l
3
1,
2
D
ep
ar
t
men
t
o
f
Mat
h
emat
i
cs
an
d
Co
m
p
u
t
er
Sci
e
n
ce,
Po
l
y
d
i
s
c
i
p
l
i
n
ary
Facu
l
t
y
,
Su
l
t
a
n
Mo
u
l
a
y
Sl
i
ma
n
e
U
n
i
v
ers
i
t
y
,
Mo
ro
cc
o
3
E
N
SICA
E
N
Sch
o
o
l
,
L
A
C
L
ab
o
ra
t
o
r
y
Caen
-
N
o
rman
d
i
e
U
n
i
v
ers
i
t
y
,
Fran
ce
Ar
t
icle
I
n
f
o
AB
S
T
RA
CT
A
r
ti
c
le
h
is
tor
y
:
R
e
c
e
ived
S
e
p
26
,
2019
R
e
vis
e
d
F
e
b
24
,
2020
Ac
c
e
pted
Apr
20
,
2020
T
h
i
s
p
a
p
er
p
re
s
en
t
s
a
n
au
t
o
ma
t
i
c
s
p
ea
k
er
i
d
en
t
i
f
i
cat
i
o
n
a
n
d
s
p
eech
rec
o
g
n
i
t
i
o
n
fo
r
A
rab
i
c
d
i
g
i
t
s
i
n
n
o
i
s
y
en
v
i
r
o
n
me
n
t
.
In
t
h
i
s
w
o
rk
,
t
h
e
p
r
o
p
o
s
e
d
s
y
s
t
em
i
s
ab
l
e
t
o
i
d
e
n
t
i
fy
t
h
e
s
p
eak
er
aft
er
s
av
i
n
g
h
i
s
v
o
i
ce
i
n
t
h
e
d
at
ab
a
s
e
an
d
ad
d
i
n
g
n
o
i
s
e.
T
h
e
me
l
freq
u
en
c
y
ce
p
s
t
ral
c
o
effi
ci
e
n
t
s
(MFCC)
i
s
t
h
e
b
e
s
t
a
p
p
r
o
ach
u
s
e
d
i
n
b
u
i
l
d
i
n
g
a
p
ro
g
ram
i
n
t
h
e
Mat
l
ab
p
l
at
f
o
rm;
a
l
s
o
,
t
h
e
q
u
an
t
i
za
t
i
o
n
i
s
u
s
e
d
f
o
r
g
e
n
e
r
a
t
i
n
g
t
h
e
c
o
d
e
b
o
o
k
s
.
T
h
e
G
a
u
s
s
i
a
n
m
i
x
t
u
r
e
m
o
d
e
l
l
i
n
g
(
G
M
M
)
a
l
g
o
r
i
t
h
m
s
a
r
e
u
s
e
d
t
o
g
e
n
e
r
a
t
e
t
e
m
p
l
a
t
e
,
f
e
a
t
u
r
e
-
m
a
t
c
h
i
n
g
p
u
r
p
o
s
e
.
I
n
t
h
i
s
p
a
p
e
r
,
w
e
h
a
v
e
p
r
o
p
o
s
e
d
a
s
y
s
t
e
m
b
a
s
e
d
o
n
M
F
C
C
-
G
M
M
a
n
d
M
F
C
C
-
V
Q
a
p
p
r
o
a
c
h
e
s
o
n
t
h
e
o
n
e
h
a
n
d
a
n
d
b
y
u
s
i
n
g
t
h
e
h
y
b
r
i
d
a
p
p
r
o
a
c
h
MFCC
-
VQ
-
G
M
M
o
n
t
h
e
o
t
h
e
r
h
an
d
fo
r
s
p
eak
er
mo
d
el
i
n
g
.
T
h
e
w
h
i
t
e
G
au
s
s
i
an
n
o
i
s
e
i
s
ad
d
e
d
t
o
t
h
e
cl
ean
s
p
eec
h
at
s
e
v
eral
s
i
g
n
a
l
-
to
-
n
o
i
s
e
ra
t
i
o
(SN
R)
l
e
v
el
s
t
o
t
e
s
t
t
h
e
s
y
s
t
em
i
n
a
n
o
i
s
y
en
v
i
ro
n
men
t
.
T
h
e
p
r
o
p
o
s
e
d
s
y
s
t
em
g
i
v
es
g
o
o
d
res
u
l
t
s
i
n
reco
g
n
i
t
i
o
n
rat
e.
K
e
y
w
o
r
d
s
:
Ar
a
bic
digi
ts
GM
M
M
F
C
C
S
NR
VQ
Th
i
s
i
s
a
n
o
p
en
a
c
ces
s
a
r
t
i
c
l
e
u
n
d
e
r
t
h
e
CC
B
Y
-
SA
l
i
ce
n
s
e
.
C
or
r
e
s
pon
din
g
A
u
th
or
:
Abde
lkbi
r
Ouis
a
a
da
ne
,
De
pa
r
tm
e
nt
of
M
a
thema
ti
c
s
a
nd
C
omput
e
r
S
c
ienc
e
,
P
olydi
s
c
ipl
inar
y
F
a
c
ult
y
,
S
ult
a
n
M
oulay
S
li
mane
Unive
r
s
it
y
,
M
or
oc
c
o
.
E
mail:
Abde
lkbi
r
.
wis
s
@gmail.
c
om
1.
I
NT
RODU
C
T
I
ON
Automatic
s
pe
e
c
h
r
e
c
ognit
ion
(
ASR
)
is
a
n
im
por
ta
nt
topi
c
of
s
pe
e
c
h
pr
oc
e
s
s
ing.
ASR
is
a
tec
hnology
that
a
ll
ows
a
n
e
lec
tr
onic
platf
o
r
m
s
uc
h
a
s
a
s
mar
tp
hone
or
a
c
omput
e
r
to
identif
y
s
poke
n
wor
ds
by
hu
mans
[
1]
.
S
pe
e
c
h
is
a
powe
r
f
u
l
a
nd
na
tu
r
a
l
tool
f
or
c
omm
u
nica
ti
on.
F
or
thi
s
,
the
s
pe
e
c
h
r
e
c
ognit
ion
s
ys
tem
make
s
the
int
e
r
a
c
ti
on
be
twe
e
n
a
hu
man
a
nd
a
mac
hine
mor
e
f
lui
d
a
nd
s
im
pler
[
2]
.
I
n
r
e
c
e
nt
ye
a
r
s
,
the
r
e
s
e
a
r
c
he
r
s
ha
ve
de
ve
loped
mor
e
im
por
tant
r
e
s
e
a
r
c
h
in
biom
e
tr
ic
s
e
c
ur
it
y
tec
hnology
with
s
pe
a
ke
r
r
e
c
ogni
ti
on
to
make
the
c
omm
unica
ti
on
be
twe
e
n
humans
a
nd
mac
hines
to
be
mor
e
na
tur
a
l
[
3]
.
T
he
s
pe
a
ke
r
r
e
c
ognit
ion
s
y
s
tem
c
a
n
be
c
las
s
if
ied
int
o
identif
ica
ti
on
a
nd
ve
r
if
ica
ti
on
(
r
e
c
ognit
ion
)
.
S
pe
a
ke
r
identif
ica
ti
on
is
the
pr
oc
e
s
s
of
a
utom
a
ti
c
a
ll
y
r
e
c
ognizing
who
i
s
s
pe
a
king
ba
s
e
d
o
n
indi
vidual
in
f
or
mation
include
d
in
s
pe
e
c
h
wa
ve
s
.
S
pe
a
ke
r
ve
r
if
ica
ti
on,
on
the
o
ther
ha
nd
,
is
the
pr
oc
e
s
s
of
a
c
c
e
pti
ng
or
r
e
jec
ti
ng
the
identit
y
c
laim
of
a
s
pe
a
ke
r
.
T
his
tec
hnique
make
s
the
s
pe
a
ke
r
ve
r
if
y
thei
r
ident
it
y
a
nd
c
ontr
ol
s
uc
h
a
s
s
e
c
ur
it
y
c
ontr
ol,
tele
phone
s
hopping,
a
c
c
e
s
s
s
e
r
vice
s
to
the
voice
mail,
da
taba
s
e
a
c
c
e
s
s
s
e
r
vice
s
a
nd
r
e
mot
e
a
c
c
e
s
s
to
c
omput
e
r
s
[
4]
.
An
ASR
s
ys
tem
invol
ve
s
two
pha
s
e
s
:
the
t
r
a
ini
ng
pha
s
e
a
nd
the
tes
ti
ng
pha
s
e
.
At
the
tr
a
ini
ng
pha
s
e
,
the
pa
r
a
mete
r
s
of
the
c
las
s
if
ica
ti
on
model
a
r
e
e
s
ti
mate
d
us
ing
a
lar
ge
number
o
f
tr
a
ini
ng
da
ta.
T
he
e
xtr
a
c
ti
on
of
f
e
a
tur
e
s
is
done
f
r
o
m
a
ll
s
pe
e
c
h
s
ignals
us
ing
va
r
ious
f
e
a
tur
e
e
xtr
a
c
ti
on
tec
hniques
s
uc
h
a
s
M
F
C
C
,
L
P
C
,
L
DA
,
R
AST
A,
e
tc
[
1,
3
,
5]
.
T
he
s
e
f
e
a
tur
e
s
a
r
e
in
the
f
or
m
of
ve
c
tor
s
is
s
tor
e
d
in
r
e
f
e
r
e
nc
e
models
,
in
p
a
r
ti
c
ular
,
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
1693
-
6930
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
,
Vol.
18
,
No
.
4
,
Augus
t
2020
:
2193
-
2204
2194
the
a
c
ous
ti
c
model,
whic
h
is
us
e
d
to
c
ha
r
a
c
ter
ize
that
wor
d
us
ing
the
c
las
s
if
ica
ti
on
a
lgor
it
hm
in
th
e
tes
ti
ng
pha
s
e
[
5]
.
T
he
mos
t
tec
hnique
us
e
d
f
or
s
pe
e
c
h
f
e
a
tur
e
s
c
o
mpr
is
e
s
the
M
e
l
-
f
r
e
que
nc
y
c
e
p
s
tr
a
l
c
oe
f
f
icie
nts
(
M
F
C
C
)
,
pe
r
c
e
ptual
li
ne
a
r
p
r
e
diction
(
P
L
P
)
a
n
d
the
li
ne
a
r
p
r
e
dictive
c
oding
(
L
P
C
)
c
oe
f
f
icie
nts
.
[
6]
.
T
he
M
F
C
C
s
a
r
e
the
be
s
t
known,
f
or
that
r
e
a
s
on
w
e
us
e
it
in
thi
s
pa
pe
r
.
T
he
M
F
C
C
s
a
r
e
the
be
s
t
kno
wn,
they
a
r
e
les
s
s
us
c
e
pti
ble
to
s
pe
a
ke
r
-
de
pe
nde
nt
va
r
iations
a
nd
ther
e
f
or
e
we
us
e
it
in
th
is
s
tudy.
[
7
]
.
M
a
ny
matc
hing
tec
hniques
a
r
e
us
e
d
in
s
p
e
e
c
h
a
nd
s
p
e
a
ke
r
r
e
c
ognit
ion,
s
uc
h
a
s
dyna
mi
c
ti
me
wa
r
ping
(
DT
W
)
,
hidden
M
a
r
kov
model
s
(
HM
M
)
that
a
r
e
ve
r
y
f
r
e
que
ntl
y
us
e
d
in
s
pe
e
c
h
r
e
c
ognit
ion
[
8,
9]
,
a
r
ti
f
icia
l
ne
ur
a
l
ne
twor
k
(
AN
N)
,
ga
us
s
ian
mi
xtur
e
model
(
GM
M
)
a
nd
ve
c
tor
qu
a
nti
z
a
ti
on
(
VQ
)
a
r
e
ge
ne
r
a
ti
ve
models
us
e
d
f
or
c
r
e
a
ti
ng
a
s
pe
a
ke
r
model
.
T
he
Ga
us
s
ian
mi
xtu
r
e
model
is
wide
ly
us
e
d
in
s
pe
a
ke
r
identi
f
ica
ti
on
modeling
s
ys
tem
[
10]
.
I
n
thi
s
pa
pe
r
,
the
VQ
method
is
e
mpl
oye
d
a
nd
it
wil
l
be
c
ompar
e
d
with
the
GM
M
model.
T
he
e
mpl
oye
d
method
ha
s
c
ha
r
a
c
ter
ize
d
by
it
s
e
a
s
ies
t
im
pleme
ntation
a
nd
it
s
highes
t
a
c
c
ur
a
c
y.
F
or
ve
c
tor
qua
nti
z
a
ti
on
(
V
Q)
t
he
L
B
G
(
L
inde,
B
uz
o,
a
nd
G
r
a
y)
a
lgo
r
it
hm
a
nd
the
k
-
mea
ns
a
lgor
it
hm
a
r
e
the
mos
t
f
a
mi
li
a
r
a
lgor
it
hms
[
11
-
13]
.
I
n
voice
a
ppli
c
a
ti
ons
,
s
pe
e
c
h
is
da
mage
d
due
to
i
nter
f
e
r
e
nc
e
with
ba
c
kgr
ound
nois
e
.
C
ons
e
que
ntl
y,
we
c
a
nnot
know
whe
ther
the
s
ignal
c
ontains
v
a
li
d
inf
or
mation
or
not
th
r
ough
di
r
e
c
t
obs
e
r
va
ti
on
[
14
]
.
I
n
thi
s
pa
pe
r
,
the
pe
r
f
o
r
manc
e
of
a
s
pe
a
ke
r
identif
ica
ti
on
s
ys
tem
pr
e
s
e
nted
f
o
r
the
c
lea
n
s
pe
e
c
h
ha
s
be
e
n
f
ur
ther
inves
ti
ga
ted
he
r
e
by
a
dding
nois
e
in
pa
r
t
icula
r
a
ddi
ti
ve
white
Ga
us
s
ian
nois
e
(
AW
GN
)
to
the
c
lea
n
‘
s
pe
a
ke
r
s
’
utt
e
r
a
nc
e
s
in
the
tr
a
ini
ng
a
nd
tes
ti
ng
pha
s
e
[
15
,
16
]
.
M
uc
h
wor
k
s
is
done
in
s
pe
e
c
h
p
r
oc
e
s
s
ing
f
or
many
langua
ge
s
in
s
pe
a
ke
r
s
ys
tem
indepe
nde
nt,
a
utom
a
ti
c
s
pe
e
c
h
r
e
c
ognit
ion
f
or
a
n
is
olate
d
wor
d
or
c
onti
nuous
s
pe
e
c
h
s
uc
h
a
s
[
17
-
19]
,
e
tc.
E
xis
ti
n
g
s
pe
e
c
h
r
e
c
ognit
ion
s
ys
tems
a
r
e
wor
king
we
ll
f
o
r
E
ur
o
pe
a
n
langua
ge
s
li
ke
E
ngli
s
h
[
20
-
24]
.
T
he
r
e
s
e
a
r
c
he
s
f
or
the
Ar
a
bic
langua
ge
s
pe
e
c
h
r
e
c
ognit
ion
is
s
ti
ll
we
a
k,
e
s
pe
c
ially
the
c
onti
nuous
s
pe
e
c
h
r
e
c
ognit
ion
i
n
a
nois
y
e
nvir
onment
[
25
-
29]
.
H
owe
ve
r
,
a
ll
a
ppli
c
a
ti
ons
i
n
s
pe
e
c
h
r
e
c
ognit
ion
a
r
e
mos
tl
y
a
va
il
a
ble
in
E
ng
li
s
h,
li
ke
the
wor
ks
pr
e
s
e
nted
in
[
30
-
34]
.
De
s
pit
e
Ar
a
bic
be
ing
the
f
our
th
mos
t
wide
ly
s
poke
n
langua
ge
of
th
e
wor
ld,
whic
h
is
why
our
r
e
s
e
a
r
c
h
wa
s
f
oc
us
e
d
on
Ar
a
bic
s
pe
e
c
h
r
e
c
ognit
i
on
in
nois
y
e
nvi
r
onments
a
nd
be
c
a
us
e
it
is
the
f
ir
s
t
wo
r
k
us
ing
the
methods
that
we
mentioned
e
a
r
li
e
r
.
T
he
Ar
a
bic
langua
ge
is
the
f
our
th
lar
ge
s
t
langua
ge
s
poke
n
by
ne
a
r
ly
1.
6
bil
li
on
M
us
li
ms
na
ti
ve
s
pe
a
ke
r
s
,
thi
s
langua
ge
s
poke
n
by
the
major
it
y
of
the
pe
ople
in
the
M
iddl
e
E
a
s
t
a
nd
Nor
th
Af
r
ica
;
note
that
Ar
a
bic
ha
s
many
di
f
f
e
r
e
nt
diale
c
ts
.
T
his
is
s
ome
li
t
t
le
wor
k
in
A
r
a
bic
s
pe
a
ke
r
a
nd
s
pe
e
c
h
r
e
c
ognit
ion
[
35
-
38]
.
W
e
pr
e
s
e
nted
in
T
a
ble
1
a
L
it
e
r
a
tur
e
r
e
view
f
or
s
pe
e
c
h
r
e
c
ognit
ion
r
e
s
e
a
r
c
h
us
ing
M
F
C
C
,
GM
M
a
nd
VQ
tec
hniques
r
e
ga
r
ding
Ar
a
bic
or
other
langua
ge
s
.
−
B
a
c
kgr
ound
:
s
pe
e
c
h
r
e
c
ognit
ion
wi
th
s
pe
a
ke
r
ident
if
ica
ti
on
s
ys
tems
ha
ve
wide
ly
e
xtens
ive
a
ppli
e
d
f
iel
ds
.
M
a
ny
wor
ks
ha
d
pe
r
f
o
r
med
in
thi
s
a
r
e
a
us
ing
mu
lt
ipl
e
tec
hniques
.
M
F
C
C
,
HM
M
,
GM
M
a
nd
VQ
a
r
e
the
mos
t
pr
omi
ne
nt
methods
.
−
Obje
c
ti
ve
s
:
T
he
a
im
of
thi
s
pa
pe
r
is
to
e
xe
c
ute
a
s
mall
-
s
c
a
le
Ar
a
bic
digi
t’
s
s
pe
e
c
h
r
e
c
ognit
ion
s
y
s
te
m
in
a
nois
y
e
nvir
onment
ba
s
e
d
on
M
F
C
C
in
f
e
a
tur
e
s
e
xtr
a
c
ti
on
a
nd
hybr
id
GM
M
-
VQ
f
or
f
e
a
tur
e
s
t
r
a
in
ing
a
nd
c
las
s
if
ica
ti
on.
T
his
s
ys
tem
c
a
n
r
e
c
ognize
a
nd
r
e
s
pond
to
digi
ts
'
s
pe
e
c
h
input
s
a
nd
c
om
pa
r
e
a
n
unknown
s
pe
a
ke
r
'
s
s
pe
e
c
h
a
ga
ins
t
a
da
taba
s
e
of
N
known
s
pe
a
ke
r
s
.
T
he
be
s
t
matc
h
is
r
e
tu
r
ne
d
a
s
the
identif
ied
s
pe
a
ke
r
a
nd
the
digi
t
is
s
poke
n.
−
T
he
pr
oblem
:
T
he
A
r
a
bic
langua
ge
is
a
mong
the
mos
t
s
poke
n
langua
ge
s
in
the
wor
ld
with
a
r
ound
300
mi
ll
ion
na
ti
ve
s
pe
a
ke
r
s
.
How
e
ve
r
,
c
ompar
e
d
to
o
th
e
r
langua
ge
s
,
the
r
e
s
e
a
r
c
h
is
s
ti
ll
poo
r
in
Ar
a
bic
s
pe
e
c
h
r
e
c
ognit
ion,
e
s
pe
c
ially
in
a
nois
y
e
nvir
onment
.
T
he
pr
e
s
e
nc
e
of
ba
c
kgr
ound
nois
e
,
a
s
we
ll
a
s
the
diver
s
it
y
of
A
r
a
bic
diale
c
ts
,
a
r
e
c
ons
ider
e
d
c
ha
ll
e
nge
s
f
or
A
r
a
b
ic
s
pe
e
c
h
r
e
c
ognit
ion
.
S
tudyi
ng
M
or
oc
c
a
n
A
r
a
bic,
whic
h
is
ve
r
y
dif
fi
c
ult
,
whic
h
is
e
ve
n
c
ha
ll
e
ngin
g
in
the
or
thog
r
a
phic
r
ules
,
the
mul
ti
ple
a
c
c
e
nts
a
nd
voc
a
bular
y
a
c
c
or
ding
to
the
r
e
gions
of
M
or
oc
c
o.
−
T
he
pr
opos
e
d
s
olut
ion
:
T
o
c
ontr
ibut
e
to
de
ve
lopi
ng
Ar
a
bic
s
pe
e
c
h
r
e
c
ognit
ion
s
ys
tems
,
we
buil
t
t
wo
s
ys
tems
in
one
.
T
he
f
ir
s
t
f
or
s
pe
a
ke
r
identi
f
ica
ti
on
a
nd
the
s
e
c
ond
f
or
A
r
a
bic
s
poke
n
d
igi
ts
with
AW
GN
ba
c
kgr
ound
nois
e
.
W
he
r
e
f
or
e
,
we
pr
opos
e
the
hy
br
id
GM
M
-
VQ
model
a
long
with
M
F
C
C
a
s
a
f
e
a
tur
e
e
xtr
a
c
ti
on
tec
hnique.
He
r
e
VQ
wa
s
us
e
d
f
o
r
tr
a
ini
ng
da
ta
a
nd
f
or
s
pe
a
ke
r
identi
f
ica
ti
on
then
the
G
M
M
f
or
r
e
c
ognit
ion
.
T
he
e
f
f
ica
c
y
of
the
p
r
opos
e
d
meth
od
is
obs
e
r
ve
d
while
pe
r
f
or
mi
ng
di
f
f
e
r
e
nt
e
xpe
r
i
m
e
nts
a
nd
c
ompar
e
d
to
e
a
r
li
e
r
wor
k
.
T
he
da
taba
s
e
f
or
thi
s
wor
k
ha
s
be
e
n
buil
t
us
ing
100
male
a
nd
f
e
male
s
pe
a
ke
r
s
,
the
voc
a
bular
y
c
ons
is
ts
of
10
wor
ds
r
e
pr
e
s
e
nti
ng
the
Ar
a
bic
s
poke
n
digi
t
s
f
r
om
0
to
9.
M
F
C
C
tec
hnique
ha
s
be
e
n
us
e
d
t
o
e
xtr
a
c
t
the
f
e
a
tur
e
s
a
nd
GM
M
,
VQ
,
GM
M
-
VQ
models
ha
ve
be
e
n
us
e
d
f
or
r
e
c
ognit
ion.
T
he
s
ys
tem
is
tes
ted
by
us
ing
tes
t
da
ta
s
poke
n
by
15
s
pe
a
ke
r
s
a
nd
a
c
hieve
s
a
n
ove
r
a
ll
wor
d
-
a
c
c
ur
a
c
y
of
98.
33
%
in
c
lea
n
c
ondit
i
on
us
ing
GM
M
+
VQ
.
T
he
r
e
s
t
of
thi
s
pa
pe
r
is
or
ga
nize
d
a
s
f
oll
ows
:
I
n
s
e
c
ti
on
2,
we
c
lar
if
y
the
s
ys
tem
a
r
c
hit
e
c
tur
e
in
mor
e
de
pth.
T
he
e
xpe
r
im
e
ntal
r
e
s
ult
s
a
r
e
pr
e
s
e
nted
in
s
e
c
ti
on
3
f
oll
owe
d
by
dis
c
us
s
ion
in
s
e
c
ti
on
4.
F
inally,
we
indi
c
a
te
the
c
onc
lus
ion
a
nd
f
u
tur
e
wor
k
in
s
e
c
ti
on
5
.
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
A
r
abic
digi
ts
s
pe
e
c
h
r
e
c
ognit
ion
and
s
pe
ak
e
r
identif
ication
in
…
(
A
bde
lkbir
Ouis
aadane
)
2195
T
ab
l
e
1
.
Su
mmary
t
ab
l
e
o
f
l
i
mi
t
ed
l
i
t
era
t
u
r
e
rev
i
e
w
o
n
s
p
eech
reco
g
n
i
t
i
o
n
s
t
u
d
i
e
s
d
o
n
e
i
n
A
rab
i
c
an
d
o
t
h
er
l
a
n
g
u
ag
e
s
A
ut
hor
L
a
ngua
ge
Y
e
a
r
F
e
a
tu
r
e
e
xt
r
a
c
ti
on
M
e
th
od
R
e
c
ognt
io
n r
a
te
(
%
)
G
io
r
gi
o B
ia
ge
tt
i
e
t
a
l
.
[
4]
E
ngl
is
h
T
im
it
2017
K
a
r
hune
n
-
L
oè
ve
tr
a
ns
f
or
m (
D
K
L
T
)
E
M
,
G
M
M
97.70%
(
noi
s
y c
ondi
ti
ons
)
M
ohi
t
D
ua
e
t
a
l
.
[
5]
H
in
di
2018
M
F
C
C
, G
F
C
C
,
BF
CC
H
M
M
-
G
M
M
M
F
C
C
65.25 %
,
G
F
C
C
75.02%
,
B
F
C
C
75.56 %
B
ha
dr
a
gi
r
i
J
a
ga
n M
oh
a
n a
nd
R
a
me
s
h B
a
bu N
.
[
7]
E
ngl
is
h
2014
M
F
C
C
D
T
W
s
a
ti
s
f
yi
ng
T
. K
. D
a
s
e
t
a
l
.
[
8]
E
ngl
is
h
2016
M
F
C
C
V
Q
, H
M
M
90%
D
. N
a
ga
jy
ot
hi
a
nd
P
. S
id
da
ia
h
[
11]
S
pe
a
ke
r
voi
c
e
E
ngl
is
h
2017
M
F
C
C
V
Q
, L
B
G
hi
gh a
c
c
ur
a
c
y
A
r
na
v G
upt
a
a
nd H
a
r
s
hi
t
G
upt
a
[
12]
S
pe
a
ke
r
voi
c
e
E
ngl
is
h
2013
M
F
C
C
VQ
89%
A
nkur
M
a
ur
ya
e
t
a
l
.
[
13]
H
in
di
2017
M
F
C
C
V
Q
, G
M
M
85.49 %
us
in
g M
F
C
C
-
V
Q
94.12
%
us
in
g M
F
C
C
-
G
M
M
V
e
e
na
a
nd
M
a
th
e
w
[
14]
E
ngl
is
h
T
im
it
2015
M
F
C
C
S
V
M
-
G
M
M
95%
i
n c
le
a
n, 90%
i
n nois
y
M
us
a
b A
l
-
K
a
lt
a
kc
hi
e
t
a
l
.
[
15]
E
ngl
is
h
T
im
it
2017
P
N
C
C
,
M
F
C
C
G
M
M
-
U
B
M
95%
i
n c
le
a
n,
75.83%
S
N
R
(
0
-
30)
dB
S
. B
. D
honde
a
nd S
. M
. J
a
ga
de
[
17]
E
ngl
is
h
T
im
it
2016
M
F
C
C
V
Q
, G
M
M
98.4 %
w
it
h M
F
C
C
–
VQ
99. 2 %
w
it
h M
F
C
C
-
G
M
M
U
. G
. P
a
ti
l
e
t
a
l
.
[
18]
H
in
di
2
016
M
F
C
C
VQ
-
G
M
M
94.31 %
S
. K
a
r
pa
ga
va
ll
i
e
t
a
l
.
[
21]
T
a
mi
l
2012
M
F
C
C
H
M
M
92%
R
a
f
ik
D
je
mi
li
e
t
a
l
.
[
22]
E
ngl
is
h
I
V
iE
c
or
pus
2012
M
F
C
C
G
M
M
,
M
L
P
,
VQ,
L
V
Q
96.4%
w
it
h G
M
M
,
M
L
P
,
VQ
94.6%
w
it
h L
V
Q
B
id
ha
n B
a
r
a
i
e
t
a
l
[
23]
E
ngl
is
h
2017
M
F
C
C
, G
F
C
C
V
Q
/G
M
M
100%
i
n c
le
a
n, 90%
i
n nois
y
C
he
n W
a
ng e
t
a
l
[
24]
E
ngl
is
h
2008
M
F
C
C
VQ
-
G
M
M
93.1%
.
N
H
a
mm
a
mi
a
nd M
B
e
dda
[
25]
A
r
a
bi
c
2010
M
F
C
C
VQ
-
M
W
S
T
93.12%
A
w
a
is
M
a
hmood
e
t
a
l
[
26]
A
r
a
bi
c
2014
M
F
C
C
, M
D
L
F
G
M
M
96.89%
M
A
ls
ul
a
im
a
n e
t
a
l
[
28]
A
r
a
bi
c
2016
M
F
C
C
, M
D
L
F
, a
nd
M
D
L
F
-
MA
G
M
M
94%
M
oha
me
d K
he
li
f
a
e
t
a
l
[
29]
A
r
a
bi
c
2017
M
F
C
C
H
M
M
/G
M
M
B
e
twe
e
n 94%
a
nd 97%
A
z
z
e
d
in
e
T
oua
z
i
a
nd
M
oha
me
d
D
e
bye
c
he
[
35]
A
r
a
bi
c
2017
M
F
C
C
H
M
M
99.89 %
in
c
le
a
n,
95.94%
i
n mul
ti
-
c
ondi
ti
on
A
ni
s
s
a
I
me
n A
mr
ous
e
t
a
l
[
38]
A
r
a
bi
c
2011
M
F
C
C
H
M
M
93.91 %
in
c
le
a
n,
32.33%
in
P
in
k
noi
s
e
5dB
2.
T
HE
S
YST
E
M
AR
CHI
T
E
CT
UR
E
2.
1.
Ar
ab
ic
d
igi
t
s
s
p
e
e
c
h
r
e
c
ogn
it
ion
s
ys
t
e
m
T
he
f
ir
s
t
ten
Ar
a
bic
dig
it
s
a
r
e
:
“
S
if
f
e
r
”
,
“
W
a
he
d”
,
“
I
thnani”
,
“
T
ha
latha
”
,
“
Ar
ba
a
”
,
“
Kha
ms
a
”
,
“
S
it
ta”
,
“
S
a
ba
a
”
,
“
thama
nya
”
a
nd
“
ti
s
a
a
”
.
T
he
A
r
a
bic
s
po
ke
n
digi
ts
would
be
he
lp
f
ul
in
many
a
ppli
c
a
ti
ons
s
uc
h
a
s
tele
phone
dialing
s
ys
tems
,
ba
nking
s
ys
tems
,
a
ir
li
n
e
r
e
s
e
r
va
ti
ons
,
e
tc.
T
he
s
e
ten
digi
ts
a
r
e
po
lys
yll
a
bic
wor
ds
e
xc
e
pt
“
z
e
r
o/s
if
f
e
r
”
whic
h
is
a
monos
yll
a
ble
wor
d
a
s
s
hown
in
T
a
ble
2.
T
he
s
yll
a
bles
in
the
Ar
a
bic
l
a
ngua
ge
a
r
e
C
V,
C
VC
,
a
nd
C
VC
C
inf
or
mation
the
de
c
ode
r
ne
e
ds
to
do
it
s
job.
V
indi
c
a
tes
a
(
long
or
s
hor
t)
vo
we
l
while
C
indi
c
a
tes
a
c
ons
ona
nt.
Ar
a
bic
utt
e
r
a
nc
e
s
c
a
n
on
l
y
s
tar
t
with
a
c
ons
ona
nt
[
37]
.
T
a
ble
2
.
Ar
a
bic
digi
ts
D
ig
it
s
A
r
a
bi
c
w
r
it
in
g
P
r
onunc
ia
ti
on
S
yl
la
bl
e
s
N
umbe
r
of
s
yl
la
bl
e
s
0
ر
ف
ص
s
e
fr
C
V
C
C
1
1
د
ح
ا
و
w
a
-
h
e
d
CV
-
C
V
C
2
2
ني
ن
ث
ا
a
a
th
-
n
a
yn
C
V
C
-
C
V
C
2
3
ةث
ث
th
a
-
l
a
-
th
a
h
CV
-
CV
-
C
V
C
3
4
ةع
ب
ر
أ
aa
r
-
b
aa
h
C
V
C
-
CV
-
C
V
C
3
5
ةس
م
خ
kh
a
m
-
s
a
h
C
V
C
-
C
V
C
2
6
ةت
س
s
e
t
-
t
a
h
C
V
C
-
C
V
C
2
7
ةع
ب
س
s
u
b
-
aa
h
C
V
C
-
C
V
C
2
8
ةي
ن
ا
م
ث
th
a
-
m
a
-
ny
e
h
CV
-
CV
-
C
V
C
3
9
ةع
س
ت
t
e
s
-
ah
C
V
C
-
C
V
C
2
2.
2
.
T
h
e
s
ys
t
e
m
ar
c
h
it
e
c
t
u
r
e
Automatic
s
pe
a
ke
r
r
e
c
ognit
ion
s
ys
tem
(
ASR
)
wa
s
de
f
ined
a
s
the
pr
oc
e
s
s
of
identif
ying
a
s
pe
a
ke
r
by
a
na
lyzing
s
pe
c
tr
a
l
s
ha
pe
of
his
voice
s
ign
a
l.
T
he
S
pe
a
ke
r
r
e
c
ognit
ion
a
s
il
lus
tr
a
ted
in
F
igur
e
1
r
e
pr
e
s
e
nts
the
pr
oc
e
s
s
of
identif
ying
a
pe
r
s
on
f
r
om
h
is
voice
a
f
ter
r
e
c
or
ding
it
with
a
mi
c
r
ophone
a
n
d
c
ompa
r
e
s
it
with
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
1693
-
6930
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
,
Vol.
18
,
No
.
4
,
Augus
t
2020
:
2193
-
2204
2196
a
nother
s
tor
e
d
a
s
tr
a
ini
ng.
T
his
block
s
ys
tem
is
s
pli
t
in
to
two
pha
s
e
s
:
the
f
i
r
s
t
one
r
e
pr
e
s
e
nts
the
tr
a
ini
ng
pha
s
e
a
nd
the
s
e
c
ond
one
is
tes
ti
ng.
Dur
ing
thes
e
two
pha
s
e
s
,
s
pe
a
k
e
r
identif
ica
ti
on
c
ons
is
ts
of
f
our
s
te
ps
:
voice
r
e
c
or
ding,
f
e
a
tur
e
e
xtr
a
c
ti
on
,
pa
tt
e
r
n
matc
hing
a
nd
de
c
is
ion
(
r
e
c
ognize
d/not
r
e
c
ognize
d)
.
F
igur
e
1.
Ge
ne
r
a
l
s
tr
uc
tur
e
of
s
pe
e
c
h
a
nd
s
pe
a
ke
r
r
e
c
ognit
ion
s
ys
tem
T
he
s
ys
tem
is
s
e
pa
r
a
ted
int
o
two
por
ti
ons
:
s
pe
a
ke
r
i
de
nti
f
ica
ti
on
a
nd
s
poke
n
digi
t
r
e
c
ognit
ion
.
W
he
r
e
:
-
S
poke
n
digi
t
r
e
c
ognit
ion
:
to
r
e
c
ognize
the
wo
r
d
a
mong
10
Ar
a
b
digi
ts
wor
ds
.
-
S
pe
a
ke
r
identif
ica
ti
on:
to
r
e
c
ognize
the
s
pe
a
ke
r
f
or
a
pa
r
ti
c
ular
s
poke
n
wor
d.
2.
2
.
1.
T
r
ain
in
g
p
h
as
e
T
he
s
pe
a
ke
r
’
s
r
e
f
e
r
e
nc
e
da
taba
s
e
a
long
with
the
s
pe
a
ke
r
I
Ds
a
nd
th
e
ir
a
udio
r
e
c
or
dings
a
r
e
s
tor
e
d;
th
e
s
ys
tem
c
a
n
buil
d
a
r
e
f
e
r
e
nc
e
model
f
or
that
s
pe
a
ke
r
.
R
e
ga
r
ding
in
tr
a
ini
ng
pha
s
e
f
or
s
poke
n
digi
t
r
e
c
ogn
it
ion,
we
pr
e
vious
ly
r
e
c
or
de
d
a
da
taba
s
e
a
nd
c
onve
r
ti
ng
it
int
o
a
c
ous
ti
c
ve
c
tor
s
us
ing
M
el
-
f
r
e
que
nc
y
c
e
ps
tr
um
c
oe
f
f
icie
nts
(
M
F
C
C
)
.
2.
2.
2
.
T
e
s
t
in
g
p
h
as
e
I
n
the
tes
ti
ng
pha
s
e
,
the
s
ys
tem
c
he
c
ks
that
the
s
pe
a
ke
r
's
input
s
pe
e
c
h
is
s
im
il
a
r
a
t
that
whic
h
is
s
tor
e
d
in
the
r
e
f
e
r
e
nc
e
.
T
he
r
e
f
o
r
e
,
the
s
ys
tem
c
a
n
ident
if
y
the
pe
r
s
on
who
is
s
pe
a
king
a
nd
the
d
igi
t
that
s
a
ying.
Dur
ing
thi
s
pha
s
e
,
450
voice
s
a
mpl
e
s
a
r
e
r
e
c
or
de
d
by
10
male
voice
s
a
nd
5
f
e
male
voice
s
c
hos
e
n
f
r
om
our
da
taba
s
e
.
T
he
s
e
c
lea
n
tes
t
da
ta
is
then
mi
xe
d
with
white
Ga
us
s
ian
nois
e
(
AW
GN
)
with
di
f
f
e
r
e
nt
leve
l
s
of
S
NR
s
(
5,
10,
15
,
20
dB
)
.
T
he
c
ode
book
ve
c
tor
s
a
r
e
de
ve
lo
pe
d
us
ing
the
pr
opos
e
d
VQ
-
GM
M
a
pp
r
oa
c
h
f
r
om
a
s
pe
c
if
ic
s
pe
a
ke
r
's
voice
.
T
he
n,
they
will
be
c
ompar
e
d
with
the
r
e
f
e
r
e
nc
e
models
obtaine
d
in
the
tr
a
ini
ng
pha
s
e
.
2.
3
.
Mel
-
f
r
e
q
u
e
n
c
y
c
e
p
s
t
r
u
m
c
oe
f
f
icie
n
t
s
(
M
F
C
C)
M
e
l
-
f
r
e
que
nc
y
c
e
ps
tr
um
c
oe
f
f
icie
nts
(
M
F
C
C
)
a
r
e
popular
f
e
a
tur
e
s
e
xtr
a
c
ted
f
r
om
s
pe
e
c
h
s
ignals
f
or
us
e
in
r
e
c
ognit
ion
tas
ks
.
M
F
C
C
s
a
r
e
ba
s
e
d
on
a
pe
r
c
e
ptually
s
c
a
led
f
r
e
que
nc
y
a
xis
.
T
his
a
ls
o
a
ll
ow
s
f
or
be
tt
e
r
r
e
pr
e
s
e
ntation
of
the
s
pe
e
c
h.
T
he
f
o
ll
owing
r
e
latio
n
is
us
e
d
to
c
a
lcula
te
the
M
e
ls
s
c
a
le
f
or
a
given
f
r
e
que
nc
y
f
(
Hz
)
of
s
igna
l
is
given
by
(
1)
.
(
)
=
2595
10
(
1
+
700
)
(
1)
F
igur
e
2
s
hows
the
ge
ne
r
a
l
block
diag
r
a
m
f
or
e
xtr
a
c
ti
on
of
M
F
C
C
f
e
a
tur
e
s
ve
c
tor
s
.
T
he
ba
s
ic
f
ive
o
pe
r
a
ti
ons
a
r
e
c
a
r
r
ied
on
s
pe
e
c
h
s
ignal
t
o
ge
t
the
c
e
ps
tr
a
l
c
o
e
f
f
icie
nts
.
T
he
a
c
ous
ti
c
f
e
a
tur
e
s
us
e
d
in
thi
s
e
va
lu
a
ti
on
a
r
e
c
ompos
e
d
of
39
pa
r
a
mete
r
s
with
13
M
F
C
C
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
A
r
abic
digi
ts
s
pe
e
c
h
r
e
c
ognit
ion
and
s
pe
ak
e
r
identif
ication
in
…
(
A
bde
lkbir
Ouis
aadane
)
2197
F
igur
e
2
.
De
tailed
M
F
C
C
pr
oc
e
s
s
2.
4
.
Ve
c
t
or
q
u
an
t
izat
ion
Ve
c
tor
Qua
nti
z
a
ti
on
(
VQ
)
is
a
c
las
s
ica
l
a
nd
the
mo
s
t
f
r
e
que
ntl
y
us
e
d
pa
tt
e
r
n
-
matc
hing
tec
hnique
[
2
]
.
We
will
us
e
the
VQ
a
ppr
oa
c
h,
in
thi
s
pa
pe
r
,
d
ue
to
it
s
e
a
s
ies
t
im
pleme
ntation
a
nd
it
s
higher
a
c
c
ur
a
c
y
.
T
his
tec
hnique
c
ons
is
ts
of
e
xtr
a
c
ti
ng
a
s
mall
numb
e
r
of
r
e
pr
e
s
e
ntative
f
e
a
tur
e
ve
c
tor
s
a
s
a
n
e
f
f
icie
nt
mea
ns
of
c
ha
r
a
c
ter
izing
the
s
pe
a
ke
r
-
s
pe
c
if
ic
f
e
a
tur
e
s
.
T
he
t
r
a
ini
ng
da
ta
f
e
a
tur
e
s
c
r
e
a
ted
by
VQ
method
a
r
e
c
o
mbi
ne
d
to
c
r
e
a
te
a
c
ode
book
f
o
r
e
a
c
h
s
pe
a
ke
r
.
I
n
the
r
e
c
ogn
it
ion
pha
s
e
,
the
s
ys
tem
c
ompar
e
s
the
di
f
f
e
r
e
nc
e
b
e
twe
e
n
a
s
pe
a
ke
r
's
tes
t
da
ta
a
nd
the
c
ode
book
of
e
a
c
h
s
pe
a
ke
r
.
Ac
c
or
dingl
y
,
it
c
onc
ludes
t
he
r
e
c
ognit
ion
r
e
s
ult
[
39
-
41]
.
F
igur
e
3
s
hows
a
n
il
lus
tr
a
ti
ve
diagr
a
m
o
f
thi
s
r
e
c
ognit
ion
p
r
oc
e
s
s
.
One
s
pe
a
ke
r
c
a
n
be
dis
c
r
im
inate
d
a
ga
ins
t
f
r
om
a
nother
ba
s
e
of
the
loca
ti
on
of
c
e
ntr
oids
.
I
n
the
t
r
a
ini
ng
pha
s
e
,
us
ing
the
c
lus
ter
ing
L
B
G
a
lgor
it
hm
[
42]
.
I
n
F
igu
r
e
3
we
a
r
e
li
mi
ted
to
pr
e
s
e
nt
two
s
pe
a
ke
r
s
a
nd
he
r
a
c
ous
ti
c
ve
c
tor
s
.
T
he
ye
ll
o
w
c
ir
c
les
r
e
f
e
r
to
the
a
c
ous
ti
c
ve
c
tor
s
f
r
om
the
s
pe
a
ke
r
1
wh
il
e
the
blue
c
ir
c
les
a
r
e
f
r
om
the
s
pe
a
ke
r
2
.
A
s
pe
a
ke
r
-
s
pe
c
if
ic
VQ
c
ode
book
is
ge
ne
r
a
ted
f
or
e
a
c
h
known
s
pe
a
ke
r
by
c
lus
ter
ing
his
/
he
r
tr
a
ini
ng
a
c
ous
ti
c
ve
c
tor
s
.
T
he
r
e
s
ult
c
ode
wor
ds
(
c
e
ntr
oids
)
a
r
e
s
hown
in
F
igur
e
3
by
blac
k
c
ir
c
les
f
or
s
pe
a
ke
r
1
a
nd
r
e
d
c
ir
c
les
f
or
s
p
e
a
ke
r
2.
T
he
dis
tanc
e
f
r
om
a
ve
c
tor
to
the
c
los
e
s
t
c
ode
wor
d
of
a
c
ode
book
is
c
a
ll
e
d
a
VQ
-
dis
tor
ti
on.
T
he
VQ
d
is
tor
ti
on
il
lus
tr
a
tes
the
dis
tanc
e
f
r
om
the
ne
a
r
e
s
t
c
ode
book,
c
a
lcula
ted
in
the
tes
ti
ng
pha
s
e
of
s
pe
a
ke
r
r
e
c
ognit
ion
s
ys
tem.
T
he
‘
a
de
qua
te’
s
pe
a
ke
r
c
or
r
e
s
ponds
to
mi
nim
um
VQ
dis
tor
ti
on,
s
o
it
is
s
e
lec
ted
a
nd
ve
r
if
ied
[
42]
.
W
e
us
e
d
the
L
B
G
a
lgor
it
hm
to
buil
d
a
c
ode
book
f
r
o
m
a
s
e
t
of
tr
a
ini
ng
ve
c
tor
f
or
thi
s
pur
pos
e
[
42
]
.
F
igur
e
3
.
Ve
c
tor
qua
nti
z
a
ti
on
c
ode
book
2.
5
.
Gau
s
s
ian
m
ixt
u
r
e
m
od
e
l
Ga
us
s
ian
mi
xtur
e
model
(
GM
M
)
is
one
of
the
non
-
pa
r
a
metr
ic
methods
,
it
is
a
pa
r
a
metr
ic
p
r
oba
bil
it
y
de
ns
it
y
f
unc
ti
on
r
e
pr
e
s
e
nted
a
s
a
we
ight
e
d
s
um
o
f
Ga
us
s
ian
c
omponent
de
ns
it
ies
.
T
he
r
e
is
a
gr
e
a
t
s
im
il
a
r
it
y
be
twe
e
n
Ga
us
s
ian
mi
xtur
e
model
a
nd
Ve
c
tor
qua
nti
z
a
t
ion
model
in
ter
ms
of
ove
r
lapping
c
lus
ter
s
.
T
he
s
ymbol
na
med
λ
r
e
pr
e
s
e
nts
c
oll
e
c
ti
ve
ly
thes
e
pa
r
a
mete
r
s
;
it
is
given
in
f
o
r
mul
a
2
.
E
a
c
h
s
pe
a
ke
r
is
r
e
pr
e
s
e
nted
by
a
GM
M
a
nd
is
r
e
f
e
r
r
e
d
to
by
his
/her
model
λ
.
=
{
,
̅
,
Σ
}
=
1
,
…
,
(
2)
whe
r
e
̅
is
the
mea
n
ve
c
tor
a
nd
the
c
ova
r
ianc
e
matr
i
x
of
the
nor
mally
dis
tr
ibut
e
d
r
a
ndom
va
r
iable
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
1693
-
6930
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
,
Vol.
18
,
No
.
4
,
Augus
t
2020
:
2193
-
2204
2198
I
n
thi
s
method,
the
dis
tr
ibut
ion
of
the
f
e
a
tur
e
ve
c
tor
x
is
modele
d
c
lea
r
ly
us
ing
a
mi
xtur
e
of
M
Ga
us
s
ian
s
.
GM
M
pa
r
a
mete
r
s
a
r
e
e
s
ti
mate
d
f
r
om
tr
a
ini
ng
da
ta
us
ing
the
it
e
r
a
ti
ve
e
xpe
c
tation
-
maximi
z
a
ti
on
(
E
M
)
a
lgor
i
thm
.
T
he
s
e
pa
r
a
mete
r
s
of
GM
M
a
r
e
c
omput
e
d
in
tr
a
ini
ng
pha
s
e
to
c
r
e
a
te
a
s
pe
a
ke
r
model.
I
n
tes
ti
ng
pha
s
e
,
the
s
p
e
a
ke
r
model
ha
ving
highes
t
a
pos
ter
ior
i
pr
oba
bil
it
y
f
or
the
f
e
a
tur
e
s
o
f
a
n
unknown
vo
ice
is
s
e
lec
ted
a
s
identit
y
of
that
unknown
s
pe
a
ke
r
[
14
,
17
,
43
,
44
]
.
A
Ga
us
s
ian
mi
xtur
e
model
is
a
we
ight
e
d
s
um
o
f
M
c
omponent
Ga
us
s
ian
de
ns
it
ies
a
s
given
by
the
e
qua
ti
on:
1
(
|
)
(
|
,
)
M
i
i
i
i
p
x
w
g
x
=
=
(
3)
T
he
f
oll
owing
d
iagr
a
m
a
s
s
hown
in
F
igu
r
e
4
il
lus
tr
a
tes
the
GM
M
model
li
ng
pr
oc
e
s
s
of
s
pe
a
ke
r
da
ta.
I
t
s
hows
the
il
lus
tr
a
ti
ve
s
teps
of
Ga
us
s
ian
mi
xtur
e
modelli
n
g
of
s
pe
a
ke
r
’
s
da
taba
s
e
.
F
igur
e
4
.
GM
M
B
lock
diagr
a
m
o
f
S
pe
a
ke
r
identif
i
c
a
ti
on
2.
6.
S
ign
a
l
t
o
n
ois
e
r
at
io
T
he
S
ignal
to
Nois
e
R
a
ti
o
(
S
NR
)
is
a
method
to
mea
s
ur
e
the
s
ignal
s
tr
e
ngth
r
e
lative
to
ba
c
kgr
ound
nois
e
leve
ls
.
T
he
S
NR
is
e
xp
r
e
s
s
e
d
by
de
c
ibels
(
dB
)
us
ing
thi
s
f
o
r
mul
a
:
10
(
(
)
1
0
)
s
p
e
e
c
h
n
o
is
e
P
S
NR
d
B
log
P
=
(
4)
whe
r
e
,
()
s
p
e
e
c
h
xt
P
P
=
,
()
n
o
i
s
e
n
t
P
P
=
)
de
note
the
powe
r
o
f
s
pe
e
c
h
s
ignal
a
nd
no
is
e
,
r
e
s
pe
c
ti
ve
ly.
T
he
c
lea
n
s
pe
e
c
h
s
ignal
x
(
t
)
is
de
gr
a
de
d
by
a
ddit
ive
s
ignal
no
is
e
n(
t)
by:
n(
t)
=
aw
gn
(
x
,
s
nr
)
(
5)
S
o,
the
obs
e
r
ve
d
nois
y
s
pe
e
c
h
y
(
t)
c
a
n
be
e
xpr
e
s
s
e
d
as
:
y
(
t)
=
x
(
t)
+
n
(
t)
(
6)
An
e
xa
mpl
e
of
the
c
or
r
upt
ion
of
the
c
lea
n
s
ignal
w
it
h
a
n
a
ddit
ive
whi
te
Ga
us
s
ian
nois
e
(
AW
GN
)
is
gi
ve
n
in
F
igur
e
5
.
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
A
r
abic
digi
ts
s
pe
e
c
h
r
e
c
ognit
ion
and
s
pe
ak
e
r
identif
ication
in
…
(
A
bde
lkbir
Ouis
aadane
)
2199
F
igur
e
5
.
T
he
magnitude
s
pe
c
tr
um
f
or
m
of
the
or
ig
inal
c
lea
n
a
nd
the
nois
y
s
ignal
of
the
wor
d
“
/wa
hid/
”
r
e
c
or
de
d
in
S
NR
=
5
dB
3.
E
XP
E
RI
M
E
NT
AL
RE
S
U
L
T
S
3.
1.
Cor
p
u
s
p
r
e
p
ar
at
ion
I
n
thi
s
pa
pe
r
,
we
ha
ve
c
ons
ider
e
d
two
c
a
tegor
i
e
s
of
a
da
taba
s
e
,
c
ontaining
the
no
is
y
da
ta
a
nd
c
lea
n
da
ta
;
−
C
lea
n
da
taba
s
e
W
e
ha
ve
r
e
gis
ter
e
d
a
da
taba
s
e
AR
B
DI
GI
T
S
of
1
00
Ar
a
b
M
or
oc
c
a
n
s
pe
a
ke
r
s
including
male
s
a
nd
f
e
male
s
ha
ve
be
e
n
buil
t
f
r
om
‘
s
if
f
e
r
’
(
z
e
r
o
)
to
‘
T
is
a
a
’
(
nine)
whic
h
a
r
e
us
e
d
f
or
t
r
a
ini
ng
pur
pos
e.
T
he
s
pe
a
ke
r
s
pe
a
ks
the
wor
d
s
e
ve
r
a
l
ti
mes
is
olate
d.
T
he
s
e
voice
s
ha
ve
be
e
n
r
e
c
or
de
d
a
t
s
a
mpl
ing
f
r
e
que
nc
ies
8000
Hz
.
W
e
us
e
the
nois
e
r
e
moval
tool
a
va
il
a
ble
in
"
a
uda
c
it
y"
s
of
twa
r
e
to
de
lete
ba
c
kgr
ound
nois
e
f
r
om
the
or
igi
na
l
r
e
c
or
ding
a
nd
then
we
ge
t
t
he
c
lea
n
da
ta.
−
Nois
y
da
taba
s
e
F
or
the
da
ta
that
we
us
e
it
in
the
tes
ti
ng
pha
s
e
,
the
nois
e
is
a
dde
d
by
the
M
AT
L
AB
f
unc
ti
on
a
dding
white
Ga
us
s
ian
nois
e
(
AW
GN
)
to
the
c
lea
n
da
taba
s
e
AR
B
DI
GI
T
S
a
t
va
r
ious
leve
ls
of
s
ignal
to
no
is
e
r
a
ti
o
(
S
NR
)
va
r
ying
f
r
om
5
dB
to
20
dB
.
T
he
s
e
tes
ti
ng
s
a
mpl
e
s
c
ons
is
t
of
10
male
voice
s
a
nd
5
f
e
male
voice
s
both
a
ge
d
be
twe
e
n
15
a
nd
40
ye
a
r
s
.
T
he
y
r
e
c
or
de
d
the
f
ir
s
t
ten
Ar
a
bic
digi
ts
r
e
pe
a
ted
thr
e
e
ti
mes
.
All
the
pr
e
vious
a
na
lys
is
wa
s
made
only
in
one
nois
y
c
ondit
ion
(
S
NR
s
leve
l
f
r
om
5
dB
int
o
20
dB
a
nd
AW
GN
nois
e
)
.
T
a
ble
3
r
e
por
ts
mo
r
e
tec
hnica
l
de
tails
a
bout
AR
AD
I
GI
T
S
da
taba
s
e
us
e
d
in
e
xpe
r
im
e
ntal
e
va
luation.
T
a
ble
3
.
I
nf
o
r
mation
a
nd
c
ondit
ion
o
f
the
AR
B
DI
GI
T
S
c
o
r
pus
us
e
d
P
r
oc
e
s
s
D
e
s
c
r
ip
ti
on
P
a
r
ti
c
ip
a
nt
100 S
pe
a
ke
r
s
(
70 M
a
le
30 f
e
ma
le
s
)
E
nvi
r
onme
nt
R
e
ve
r
be
r
a
nt
a
nd t
w
o c
ha
nn
e
ls
-
s
te
r
e
o mode
.
W
or
ds
10 A
r
a
bi
c
s
poke
n D
ig
it
s
T
r
a
in
in
g S
e
t
85
S
pe
a
ke
r
s
T
e
s
ti
ng S
e
t
15 S
pe
a
ke
r
s
N
umbe
r
of
c
le
a
n w
or
ds
s
e
le
c
te
d
1
0
3
1
0
0
3
0
0
0
=
T
ot
a
l
N
umbe
r
of
noi
s
y
w
or
ds
3
0
0
0
1
4
1
2
0
0
0
=
noi
s
e
t
ype
us
e
d
A
W
G
N
S
N
R
l
e
ve
l
us
e
d
C
le
a
n,5 dB
, 10 dB
, 15 d
B
, 20 dB
T
ot
a
l
S
iz
e
of
D
a
ta
ba
s
e
1 G
B
S
a
mpl
in
g F
r
e
que
nc
y, f
s
8000Hz
S
of
twa
r
e
us
e
d f
or
mi
xi
ng
M
A
T
L
A
B
R
2016
b v
T
r
ia
l
3.
2.
Re
s
u
lt
s
F
ir
s
tl
y,
the
pe
r
f
o
r
manc
e
of
the
s
ys
tem
is
e
va
luate
d
by
r
e
c
ognit
ion
r
a
te.
I
t
is
c
a
lcula
ted
by:
=
.
×
100
(
7)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
1693
-
6930
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
,
Vol.
18
,
No
.
4
,
Augus
t
2020
:
2193
-
2204
2200
T
he
a
ve
r
a
ge
r
e
c
ognit
ion
r
a
tes
obtaine
d
f
r
o
m
the
t
e
n
Ar
a
b
digi
ts
in
c
lea
n
e
nvi
r
onment
a
nd
in
nois
y
e
nvir
onment,
f
or
di
f
f
e
r
e
nt
S
NR
va
lues
,
a
r
e
r
e
pr
e
s
e
nted
in
T
a
ble
4
a
nd
e
f
f
icie
nc
y
c
ha
r
t
is
s
hown
in
F
igur
e
6
r
e
s
pe
c
ti
ve
ly.
F
r
om
the
r
e
s
ult
s
s
hown
in
T
a
ble
4
,
w
e
c
a
n
c
onc
lude
that
the
e
f
f
e
c
t
of
the
nois
e
is
not
im
por
tant
if
the
S
NR
is
s
upe
r
ior
to
20dB
,
in
thi
s
c
a
s
e
we
obt
a
in
a
ppr
oxim
a
tely
the
s
a
me
‘
a
ve
r
a
ge
’
va
lue
of
r
e
c
ognit
ion
obtaine
d
in
c
lea
ne
d
da
ta.
I
n
tes
ti
ng
pha
s
e
f
or
s
pe
a
ke
r
identif
ica
ti
on
,
the
s
poke
n
s
a
mp
les
a
r
e
r
e
c
or
d
e
d
by
15
s
pe
a
ke
r
s
;
(
10
male
s
pe
a
ke
r
s
a
nd
5
f
e
male
s
pe
a
k
e
r
s
)
c
hos
e
n
f
r
om
our
da
taba
s
e
AR
B
DI
GI
T
S
of
1
00
Ar
a
b
M
or
oc
c
a
n
s
pe
a
ke
r
s
(
the
s
pe
e
c
h
wa
ve
is
with
8
K
Hz
s
a
mpl
ing
f
r
e
que
nc
y
us
ing
AU
DI
OR
E
C
OR
D
f
unc
ti
on
of
M
AT
L
AB
2016
e
nvir
onment
in
windows
pl
a
tf
or
m
in
64
bit
)
.
T
he
s
a
mpl
e
c
oll
e
c
ti
on
pr
oc
e
s
s
is
a
c
c
ompl
is
he
d
by
us
ing
the
mi
c
r
ophone
to
r
e
c
or
d
the
s
pe
e
c
h
of
m
a
le/f
e
male
.
T
he
f
ir
s
t
tes
ti
ng
pha
s
e
,
in
c
lea
n
c
ondit
ion,
a
f
ter
t
his
,
we
ha
ve
tes
ted
with
a
dding
AW
GN
nois
e
a
t
dif
f
e
r
e
nt
S
NR
s
leve
ls
va
lue
s
of
5,
10,
15,
20
dB
.
N
ote
that
the
s
pe
e
c
h
s
e
gment
wa
s
de
gr
a
de
d
whe
n
S
NR
<
5dB
.
T
he
pe
r
c
e
ntage
r
e
c
ognit
ion
o
f
a
s
pe
a
ke
r
is
given
i
n
the
T
a
ble
5
a
nd
the
e
f
f
icie
nc
y
c
ha
r
t
is
s
hown
in
F
igur
e
7
r
e
s
pe
c
ti
ve
ly.
T
he
a
ve
r
a
ge
o
f
15
s
pe
a
ke
r
’
s
r
e
c
ognit
ion
r
a
tes
obtaine
d
with
tr
a
ini
ng
by
GM
M
in
c
lea
n
e
nvir
onment
a
nd
in
nois
y
e
nvir
onment
,
f
or
di
f
f
e
r
e
n
t
S
NR
va
lues
a
r
e
r
e
pr
e
s
e
nted
in
T
a
ble
6
a
nd
e
f
f
icie
nc
y
c
ha
r
t
is
s
hown
in
F
igur
e
8
r
e
s
pe
c
ti
ve
ly.
T
a
ble
7
s
hows
the
ove
r
a
ll
r
e
c
ognit
ion
r
a
tes
(
%
)
f
or
the
s
pe
a
ke
r
i
de
nt
if
ica
ti
on
s
ys
tem
us
ing
the
c
ombi
na
ti
on
of
VQ
a
nd
GM
M
a
l
gor
it
hms
a
nd
ba
r
c
ha
r
t
plot
is
s
hown
in
F
igur
e
9
.
T
a
ble
4
.
Ave
r
a
ge
r
e
c
ognit
ion
r
a
te
(
%
)
us
ing
mf
c
c
+
vq
N
oi
s
e
L
e
ve
l
D
ig
it
s
C
le
a
n
5dB
10dB
15dB
20dB
0
82.37
45.17
67.27
79.66
80.33
1
80.11
42.34
65.17
75.11
79.28
2
81.54
44.06
62.66
78.09
80.05
3
79.08
40.47
60.87
75.33
78.00
4
85.52
41.52
68.12
77.18
84.11
5
80.17
38.78
58.34
74.23
80.17
6
84.42
42.10
66.52
76.12
84.42
7
83.67
39.27
62.71
75.25
82.05
8
77.02
38.96
60.33
71.78
77.02
9
86.63
47.03
61.51
79.00
85.28
A
ve
r
a
ge
82,05
41,97
63,35
76,17
81,07
T
a
ble
5
.
S
pe
a
ke
r
I
de
nt
if
ica
ti
on
r
a
te
(
%
)
f
or
tes
ti
ng
s
pe
e
c
h
in
c
lea
n
a
nd
in
AW
GN
nois
e
f
o
r
di
f
f
e
r
e
nt
S
NR
M
e
th
ods
#s
pe
a
ke
r
s
c
le
a
n
5dB
10dB
15dB
20dB
M
F
C
C
+
V
Q
10 ma
le
s
89.43
62.11
77.27
79.66
85.12
5 f
e
ma
le
s
90.66
63.72
75.17
77.26
86.06
F
igur
e
6.
Ar
a
b
digi
ts
r
e
c
ognit
ion
s
uc
c
e
s
s
r
a
tes
(
%
)
in
c
lea
n
a
nd
in
pr
e
s
e
nc
e
of
a
wgn
nois
e
F
igur
e
7.
S
pe
a
ke
r
I
de
nti
f
ica
ti
on
r
e
s
ult
(
%
)
in
c
lea
n
a
nd
f
or
dif
f
e
r
e
nt
S
NR
va
lues
T
a
ble
6
.
S
pe
a
ke
r
I
de
nt
if
ica
ti
on
r
a
te
(
%
)
f
or
tes
ti
ng
s
pe
e
c
h
in
c
lea
n
a
nd
in
AW
GN
nois
e
f
or
di
f
f
e
r
e
nt
s
nr
us
ing
GM
M
M
e
th
ods
#s
pe
a
ke
r
s
c
le
a
n
5dB
10dB
15dB
20dB
M
F
C
C
+
G
M
M
10 ma
le
s
96
.
12
70
.
26
88
.
33
92
.
66
95
5 f
e
ma
le
s
94
.
66
70
87
.
67
92
.
33
96
.
33
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
A
r
abic
digi
ts
s
pe
e
c
h
r
e
c
ognit
ion
and
s
pe
ak
e
r
identif
ication
in
…
(
A
bde
lkbir
Ouis
aadane
)
2201
T
a
ble
7.
S
pe
a
ke
r
I
de
nt
if
ica
ti
on
r
a
te
(
%
)
f
or
tes
ti
ng
s
pe
e
c
h
in
c
lea
n
a
nd
with
AW
GN
f
o
r
di
f
f
e
r
e
nt
s
nr
us
ing
GM
M
+
VQ
C
le
a
n
5dB
10dB
15dB
20dB
10 M
a
le
98.33
88.66
90.26
95.12
98
5
f
e
ma
le
97.12
86.33
90.66
94
97
F
igur
e
8
.
S
pe
a
ke
r
I
de
nti
f
ica
ti
on
r
e
s
ult
(
%
)
in
c
lea
n
a
nd
a
ll
S
NR
leve
ls
with
AW
GN
nois
e
us
ing
GM
M
modeling
F
igur
e
9
.
S
pe
a
ke
r
I
de
nti
f
ica
ti
on
r
a
te
in
c
lea
n
a
nd
nois
e
us
ing
GM
M
+
VQ
T
he
r
e
s
ult
s
a
r
e
im
pleme
nted
unde
r
M
AT
L
AB
R
2
016b.
F
or
thi
s
,
we
ha
ve
bui
lt
a
GU
I
i
nter
f
a
c
e
a
s
il
lus
tr
a
ted
in
F
igur
e
10
to
s
im
pli
f
y
the
tes
ti
ng
pr
oc
e
s
s
whe
r
e
the
s
pe
a
k
e
r
c
a
n
be
tes
ted
dir
e
c
tl
y
by
a
ne
w
voice
r
e
c
or
ding
or
f
r
om
tes
t
ba
s
e
.
Dur
ing
r
e
c
or
ding
,
the
us
e
r
a
dds
AW
GN
nois
e
to
he
r
\
his
voice
if
he
\
s
he
de
s
ir
e
d
a
nd
he
\
s
he
s
e
lec
ts
the
S
NR
leve
l.
T
his
GU
I
e
na
bles
the
r
e
c
or
ding
o
r
plot
t
ing
of
a
s
ound
a
s
we
ll
a
s
the
r
e
c
or
ding
of
a
ne
w
tes
t
da
ta
a
nd
identif
ica
ti
on
of
the
s
pe
a
ke
r
I
D.
F
igur
e
1
0
.
GU
I
M
a
in
M
AT
L
AB
s
ys
tem
4.
DI
S
CU
S
S
I
ONS
I
n
thi
s
wor
k,
f
o
r
the
e
xpe
r
im
e
nts
tes
ted
with
the
c
lea
n
da
ta
the
maximum
pe
r
f
or
manc
e
r
e
c
e
ived
f
or
M
F
C
C
+
G
M
M
+
V
Q
is
97.
92%
,
f
or
M
F
C
C
+
GM
M
is
95
.
39%
%
a
nd
f
o
r
M
F
C
C
+
VQ
is
90.
04
%
.
Als
o,
f
or
the
tes
ts
with
nois
y
da
ta,
the
maximum
a
c
c
ur
a
c
y
r
e
c
e
ived
f
or
M
F
C
C
+
GM
M
+
VQ
is
92.
50
%
,
f
o
r
M
F
C
C
+
GM
M
is
86.
57%
a
nd
f
o
r
M
F
C
C
+
VQ
is
65.
64
%
.
I
t
is
c
l
e
a
r
ly
obs
e
r
ve
d
that
be
tt
e
r
pe
r
f
or
manc
e
ha
s
be
e
n
s
e
e
n
whe
n
us
ing
the
thr
e
e
tec
hniques
togethe
r
M
F
C
C
+
GM
M
+
VQ
.
F
r
om
the
r
e
s
ult
s
,
it
is
c
lea
r
ly
f
ound
that
the
Ar
a
bic
digi
ts
s
pe
e
c
h
r
e
c
ognit
ion
s
ys
tem
a
nd
the
s
pe
a
ke
r
i
de
nti
f
ica
ti
on
s
ys
tem
pe
r
f
o
r
med
we
ll
in
both
c
lea
n
a
nd
nois
y
e
nvir
onment
us
ing
M
F
C
C
a
s
f
e
a
tur
e
e
xtr
a
c
ti
on
a
nd
ve
c
tor
qua
nti
z
a
ti
on.
F
o
r
our
pr
opos
e
d
model
of
he
c
ombi
na
ti
on
a
mong
thr
e
e
methods
M
F
C
C
+
GM
M
+
VQ
,
we
obs
e
r
ve
d
that
the
a
c
c
ur
a
c
y
of
the
r
e
s
ult
s
obtaine
d
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
1693
-
6930
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
,
Vol.
18
,
No
.
4
,
Augus
t
2020
:
2193
-
2204
2202
is
high
with
e
it
he
r
c
lea
n
or
nois
y
da
ta
c
ompar
e
d
to
the
r
e
s
ult
s
obtaine
d
by
M
F
C
C
+
GM
M
or
M
F
C
C
+
VQ
methods
.
W
e
c
a
n
a
ls
o
c
onc
lude
that
the
e
f
f
e
c
t
of
th
e
nois
e
is
not
im
por
tant
if
the
S
NR
is
s
upe
r
ior
to
2
0
dB
,
in
thi
s
c
a
s
e
we
obtain
a
ppr
oxim
a
tely
s
a
me
‘
a
ve
r
a
ge
’
va
lue
of
r
e
c
ognit
ion
obtaine
d
in
c
lea
ne
d
da
ta.
N
o
te
t
ha
t
the
s
pe
a
ke
r
i
de
n
t
if
ic
a
t
io
n
s
ys
te
m
is
mo
r
e
pe
r
f
or
ma
nt
u
s
i
ng
t
he
GM
M
m
e
t
ho
d
c
o
mp
a
r
e
d
t
o
VQ
m
e
th
od
,
b
u
t
wh
e
n
we
h
a
ve
us
e
d
t
he
c
o
mb
in
a
t
i
on
o
f
b
ot
h
me
th
ods
(
GM
M
+
VQ
)
we
o
bt
a
i
n
be
t
te
r
s
pe
a
ke
r
i
de
n
ti
f
ica
ti
on
r
a
t
e
t
ha
n
i
nd
iv
id
ua
l
mo
de
ls
.
A
ls
o
,
w
e
r
e
ma
r
k
a
n
in
c
r
e
a
s
e
o
f
the
i
de
n
t
if
ic
a
t
io
n
r
a
te
by
1
4%
is
o
b
ta
ine
d
whe
n
S
N
R
=
1
5
dB
a
n
d
GM
M
a
s
m
ode
l
li
ng
m
e
th
od
,
bu
t
th
i
s
r
a
t
e
is
r
e
duc
e
d
t
o
7%
w
he
n
S
NR
=
5
d
B
.
W
e
c
a
n
e
xp
la
in
th
e
di
f
f
e
r
e
nc
e
be
tw
e
e
n
th
os
e
tw
o
pe
r
c
e
n
ta
ge
s
in
the
tw
o
e
x
pe
r
im
e
n
ts
b
y
th
e
ba
c
k
g
r
o
un
d
noi
s
e
a
f
f
e
c
t
o
n
th
e
f
e
a
tu
r
e
ve
c
t
o
r
s
a
nd
on
t
he
a
c
o
us
t
ic
da
ta
,
e
s
pe
c
i
a
ll
y
i
f
t
he
n
o
is
e
in
te
ns
i
ty
is
ve
r
y
s
t
r
on
g
t
he
n
i
t
c
a
n
h
id
e
d
a
t
a
,
c
ons
e
q
ue
n
t
ly
th
e
v
e
r
i
f
ica
t
io
n
is
f
a
i
le
d
.
T
h
us
,
the
ve
r
i
f
i
c
a
t
i
on
is
pe
r
f
o
r
med
o
nl
y
o
n
s
ma
ll
e
r
a
m
ou
n
t
o
f
v
a
lua
b
le
d
a
t
a
.
As
f
or
t
he
A
r
a
b
ic
s
po
ke
n
d
ig
it
s
s
ys
te
m
us
in
g
V
Q
m
e
t
ho
d
,
the
pa
r
a
me
te
r
s
ob
ta
in
e
d
a
f
te
r
t
r
a
i
n
in
g
th
e
s
ys
te
m
f
o
r
d
ig
i
t
6
a
n
d
7
a
r
e
t
oo
c
los
e
a
s
s
ho
wn
i
n
F
ig
u
r
e
7
.
T
h
e
r
e
f
o
r
e
,
i
f
th
e
di
gi
t
is
no
t
s
po
ke
n
c
le
a
r
l
y
du
r
in
g
r
e
c
o
g
ni
ti
on
,
t
he
s
ys
te
m
f
a
l
te
r
s
.
T
h
e
di
gi
t
8
g
iv
e
s
th
e
lo
we
s
t
a
c
c
ur
a
c
y
,
t
he
r
e
a
s
on
b
e
i
ng
t
he
s
p
e
e
c
h
s
a
m
pl
e
f
or
8
ha
s
t
he
h
ig
he
s
t
a
mo
un
t
of
“
u
nv
oi
c
e
d
”
s
p
e
e
c
h
s
i
gn
a
l
.
T
he
r
e
f
or
e
,
i
t
is
t
r
e
a
ted
a
s
u
nv
oi
c
e
d
s
pe
e
c
h
da
ta
.
5.
CONC
L
USI
ON
I
n
thi
s
pa
pe
r
,
we
ha
ve
p
r
e
s
e
nted
a
n
a
utom
a
ti
c
s
ys
t
e
m
a
ble
to
r
e
c
ognize
the
s
pe
a
ke
r
a
s
we
ll
a
s
s
pe
e
c
h
us
ing
M
F
C
C
,
VQ
a
nd
GM
M
tec
hnique
f
o
r
Ar
a
b
d
igi
ts
wor
ds
.
T
he
r
e
s
ult
s
hows
that
a
ve
r
a
ge
a
c
c
ur
a
c
y
f
or
the
s
ys
tem
is
90.
04%
in
c
lea
n
e
nvir
on
ment
a
nd
75
.
86
%
in
nois
y
e
nvi
r
onment
f
o
r
s
pe
a
ke
r
identif
ica
ti
on
us
ing
VQ
a
nd
f
or
A
r
a
bic
digi
ts
r
e
c
ognit
ion
s
ys
tem
is
82.
05
%
in
c
lea
n
e
nvir
on
ment
a
nd
a
ve
r
a
ge
of
65.
64
%
in
nois
y
e
nvir
onments
r
e
s
pe
c
ti
v
e
ly.
T
he
a
ve
r
a
ge
f
or
s
pe
a
ke
r
identif
ica
ti
on
us
ing
GM
M
is
95.
39%
in
c
lea
n
e
nvir
onment
a
nd
86.
57%
in
nois
y
e
nvir
onment.
F
or
the
a
ve
r
a
ge
of
the
c
ombi
na
ti
on
(
GM
M
+
VQ
)
is
97.
72%
in
c
lea
n
e
nvir
onment
a
nd
92,
50
%
a
dding
AW
GN
nois
e
,
s
o
thi
s
c
ombi
na
ti
on
gives
be
tt
e
r
identif
ica
ti
on
r
a
te
than
indi
vidual
models
.
W
e
c
a
n
im
pr
ove
the
obtaine
d
r
e
s
ult
s
if
we
us
e
other
methods
s
uc
h
a
s
AN
N,
HM
M
or
DN
N
f
or
c
las
s
if
ica
ti
on.
I
n
the
f
utur
e
wor
ks
,
we
will
c
om
pa
r
e
thi
s
method
to
other
methods
in
or
de
r
to
f
ind
a
method
that
c
a
n
im
pr
ove
the
r
obus
tnes
s
in
other
types
of
nois
e
s
,
tes
ti
ng
us
ing
other
tec
hniques
a
nd
by
in
c
r
e
a
s
ing
voc
a
bular
y
s
ize
.
AC
KNOWL
E
DGE
M
E
NT
S
T
he
a
uthor
s
thank
the
r
e
view
e
r
s
f
or
thei
r
s
ugge
s
ti
ons
a
nd
c
or
r
e
c
ti
ons
to
the
or
igi
na
l
manus
c
r
ipt
.
T
he
s
tate
ments
made
he
r
e
in
a
r
e
s
olely
the
r
e
s
pons
ibi
li
ty
of
the
a
uthor
s
.
RE
F
E
RE
NC
E
S
[1
]
R.
G
.
May
u
r,
D
.
K
i
n
n
al
,
a
n
d
S.
B.
N
i
n
a
d
,
"
Cl
as
s
i
fi
cat
i
o
n
T
ech
n
i
q
u
e
s
fo
r
S
p
eech
Rec
o
g
n
i
t
i
o
n
:
A
Re
v
i
e
w
,
"
In
t
e
r
n
a
t
i
o
n
a
l
Jo
u
r
n
a
l
o
f
E
m
er
g
i
n
g
Tec
h
n
o
l
o
g
y
a
n
d
A
d
v
a
n
ce
d
E
n
g
i
n
eer
i
n
g
,
p
p
.
5
8
-
6
3
,
2
0
1
5
.
[2
]
D
.
N
i
k
i
t
a
a
n
d
Ba
d
ad
a
p
u
re
P.
R
.
,
"
H
i
n
d
i
S
p
eech
Reco
g
n
i
t
i
o
n
Sy
s
t
em
u
s
i
n
g
MFCC
a
n
d
H
T
K
T
o
o
l
k
i
t
,
"
IJE
S
R
T
Jo
u
r
n
a
l
,
v
o
l
.
5
,
n
o
.
1
2
,
p
p
.
9
0
-
5
,
2
0
1
6
.
[3
]
N
.
Si
n
g
h
,
A
.
A
g
raw
a
l
,
an
d
R.
A
.
K
h
an
,
"
A
Cri
t
i
cal
Rev
i
ew
o
n
A
u
t
o
ma
t
i
c
Sp
ea
k
er
Reco
g
n
i
t
i
o
n
,"
S
ci
e
n
ce
Jo
u
r
n
a
l
o
f
Ci
r
cu
i
t
s
,
S
ys
t
e
m
s
a
n
d
S
i
g
n
a
l
P
r
o
ce
s
s
i
n
g
,
v
o
l
.
4
,
n
o
.
2
,
p
p
.
1
4
-
19
,
2
0
1
5
.
[4
]
G
.
Bi
ag
et
t
i
,
P.
Cri
p
p
a,
L
.
Fal
a
s
ch
e
t
t
i
,
S.
O
rci
o
n
i
,
an
d
C.
T
u
rc
h
et
t
i
,
"
Sp
ea
k
er
Id
e
n
t
i
fi
ca
t
i
o
n
i
n
N
o
i
s
y
C
o
n
d
i
t
i
o
n
s
U
s
i
n
g
Sh
o
rt
Seq
u
en
ce
s
o
f
Sp
eec
h
Frame
s
,"
In
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
e
r
en
ce
o
n
In
t
el
l
i
g
en
t
D
ec
i
s
i
o
n
Tec
h
n
o
l
o
g
i
es
,
p
p
.
4
3
‑
5
2
,
2
0
1
7
.
[5
]
M.
D
u
a,
R.
K
.
A
g
g
ar
w
al
,
a
n
d
M.
B
i
s
w
as
,
"
Perfo
rm
a
n
ce
ev
a
l
u
a
t
i
o
n
o
f
H
i
n
d
i
s
p
eec
h
rec
o
g
n
i
t
i
o
n
s
y
s
t
em
u
s
i
n
g
o
p
t
i
mi
zed
fi
l
t
er
b
an
k
s
,"
E
n
g
.
S
c
i
.
Tech
n
o
l
.
In
t
.
J.
,
v
o
l
.
2
1
,
n
o
.
3
,
p
p
.
3
8
9
-
3
9
8
,
J
u
n
.
2
0
1
8
,
[6
]
G
.
Sáro
s
i
,
M.
Mo
zs
ár
y
,
P.
Mi
h
a
j
l
i
k
,
a
n
d
T
.
Feg
y
ó
,
"
Co
m
p
ari
s
o
n
o
f
fea
t
u
re
e
x
t
rac
t
i
o
n
me
t
h
o
d
s
fo
r
s
p
eec
h
reco
g
n
i
t
i
o
n
i
n
n
o
i
s
e
-
free
a
n
d
i
n
t
raffi
c
n
o
i
s
e
en
v
i
r
o
n
me
n
t
,"
2
0
1
1
6
t
h
Co
n
f
e
r
en
ce
o
n
S
p
eec
h
Tech
n
o
l
o
g
y
a
n
d
H
u
m
a
n
-
Co
m
p
u
t
e
r
D
i
a
l
o
g
u
e
(S
p
e
D
)
,
p
p
.
1
–
8
,
2
0
1
1
.
[7
]
Bh
ad
ra
g
i
r
i
J
a
g
an
M
o
h
a
n
a
n
d
Rame
s
h
Ba
b
u
N
.
,
"
Sp
eec
h
reco
g
n
i
t
i
o
n
u
s
i
n
g
MFCC
an
d
D
T
W
,
"
2
0
1
4
I
n
t
e
r
n
a
t
i
o
n
a
l
Co
n
f
er
e
n
ce
o
n
A
d
v
a
n
ce
s
i
n
E
l
ec
t
r
i
ca
l
E
n
g
i
n
ee
r
i
n
g
(ICA
E
E
)
,
p
p
.
1
-
4
,
2
0
1
4
.
[8
]
T
.
K
.
D
as
an
d
K
h
a
l
i
d
N
ah
ar,
"
A
V
o
i
ce
Id
e
n
t
i
fi
ca
t
i
o
n
Sy
s
t
em
u
s
i
n
g
H
i
d
d
e
n
Mark
o
v
Mo
d
e
l
,"
In
d
i
a
n
Jo
u
r
n
a
l
o
f
S
ci
e
n
ce
a
n
d
Tec
h
n
o
l
o
g
y
,
v
o
l
.
9
,
n
o
.
4
,
p
p
.
1
-
6
,
2
0
1
6
.
[9
]
L
.
R.
Rab
i
n
er,
"
A
t
u
t
o
r
i
al
o
n
h
i
d
d
en
Mark
o
v
mo
d
el
s
an
d
s
el
ect
e
d
ap
p
l
i
ca
t
i
o
n
s
i
n
s
p
eec
h
reco
g
n
i
t
i
o
n
,"
P
r
o
c
eed
i
n
g
s
IE
E
E
,
p
p
.
2
5
7
–
2
8
6
,
1
9
8
9
.
[1
0
]
R.
T
o
g
n
eri
a
n
d
D
.
P
u
l
l
el
l
a,
"
A
n
O
v
erv
i
ew
o
f
Sp
ea
k
er
Id
e
n
t
i
fi
ca
t
i
o
n
:
A
ccu
racy
an
d
Ro
b
u
s
t
n
es
s
Is
s
u
e
s
,"
IE
E
E
Ci
r
c
u
i
t
s
S
ys
t
.
M
a
g
.
,
v
o
l
.
1
1
,
n
o
.
2
,
p
p
.
2
3
–
6
1
,
2
0
1
1
.
[1
1
]
D
.
N
ag
a
j
y
o
t
h
i
a
n
d
P.
Si
d
d
ai
a
h
,
"
V
o
i
ce
Reco
g
n
i
t
i
o
n
Bas
ed
o
n
V
ec
t
o
r
Q
u
an
t
i
zat
i
o
n
U
s
i
n
g
L
BG
,
"
Co
m
p
u
t
e
r
Co
m
m
u
n
i
c
a
t
i
o
n
,
Net
wo
r
ki
n
g
a
n
d
I
n
t
e
r
n
e
t
S
ecu
r
i
t
y:
P
r
o
ceed
i
n
g
s
o
f
IC
3
T
,
p
p
.
5
0
3
-
5
1
1
,
2
0
1
6
.
[1
2
]
A
.
G
u
p
t
a
a
n
d
H
.
G
u
p
t
a,
"
A
p
p
l
i
cat
i
o
n
s
o
f
MFCC
an
d
V
ec
t
o
r
Q
u
a
n
t
i
zat
i
o
n
i
n
s
p
eak
er
reco
g
n
i
t
i
o
n
,
"
in
2
0
1
3
In
t
e
r
n
a
t
i
o
n
a
l
Co
n
f
er
e
n
ce
o
n
In
t
e
l
l
i
g
e
n
t
S
ys
t
em
s
a
n
d
S
i
g
n
a
l
P
r
o
ces
s
i
n
g
(I
S
S
P
)
,
p
p
.
1
7
0
-
1
7
3
,
2
0
1
3
.
Evaluation Warning : The document was created with Spire.PDF for Python.