I
nte
rna
t
io
na
l J
o
urna
l o
f
E
lect
rica
l a
nd
Co
m
pu
t
er
E
ng
ineering
(
I
J
E
CE
)
Vo
l.
1
6
,
No
.
4
,
A
u
g
u
s
t
20
2
6
,
p
p
.
2
1
0
6
~
2
1
1
9
I
SS
N:
2088
-
8
7
0
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijece.
v
1
6
i
4
.
pp
2
1
0
6
-
2
1
1
9
2106
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ec
e.
ia
esco
r
e.
co
m
Spa
tial a
nd
cha
n
nel atte
ntion me
c
ha
nism
f
o
r spee
c
h disflue
ncy
detec
tion usin
g
de
ep learning
techni
que
K
us
um
a
H
.
R
.
1
, G
.
Ses
hik
a
la
2
1
D
e
p
a
r
t
me
n
t
o
f
El
e
c
t
r
o
n
i
c
s a
n
d
C
o
m
mu
n
i
c
a
t
i
o
n
,
G
l
o
b
a
l
A
c
a
d
e
m
y
o
f
Te
c
h
n
o
l
o
g
y
,
B
e
n
g
a
l
u
r
u
,
I
n
d
i
a
2
S
c
h
o
o
l
o
f
E
l
e
c
t
r
o
n
i
c
s
a
n
d
C
o
mm
u
n
i
c
a
t
i
o
n
E
n
g
i
n
e
e
r
i
n
g
,
R
EV
A
U
n
i
v
e
r
si
t
y
,
B
e
n
g
a
l
u
r
u
,
I
n
d
i
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Sep
2
,
2
0
2
5
R
ev
is
ed
Ap
r
1
7
,
2
0
2
6
Acc
ep
ted
Ap
r
2
6
,
2
0
2
6
S
tu
tt
e
ri
n
g
is
a
sp
e
e
c
h
c
o
m
m
u
n
ica
ti
o
n
d
iso
rd
e
r,
it
is
c
h
a
ra
c
teriz
e
d
b
y
re
p
e
ti
ti
o
n
s,
p
ro
l
o
n
g
a
ti
o
n
,
a
n
d
u
n
u
su
a
l
p
a
u
se
s
t
h
a
t
c
a
u
se
i
n
terfe
re
n
c
e
with
th
e
n
a
tu
ra
l
fl
o
w
o
f
sp
e
e
c
h
.
I
n
re
c
e
n
t
ti
m
e
s,
a
u
to
m
a
ti
c
sp
e
e
c
h
re
c
o
g
n
it
io
n
a
n
d
sp
e
e
c
h
p
ro
c
e
ss
in
g
sy
ste
m
s
h
a
v
e
g
a
in
e
d
e
n
o
rm
o
u
s
a
tt
e
n
ti
o
n
b
e
c
a
u
se
th
e
y
a
re
u
se
d
i
n
m
o
st
o
f
th
e
h
u
m
a
n
m
a
c
h
in
e
in
tera
c
ti
o
n
a
p
p
li
c
a
ti
o
n
s.
H
o
we
v
e
r,
th
e
p
e
rfo
rm
a
n
c
e
o
f
th
e
se
s
y
ste
m
s
is
a
ffe
c
ted
b
y
stu
tt
e
r
sp
e
e
c
h
,
stu
tt
e
r
d
e
tec
ti
o
n
is
th
e
m
a
jo
r
c
h
a
ll
e
n
g
e
d
u
e
t
o
s
p
e
e
c
h
d
isflu
e
n
c
ies
.
To
a
d
d
re
ss
th
is
m
a
jo
r
c
h
a
ll
e
n
g
e
,
th
is
p
a
p
e
r
in
tr
o
d
u
c
e
d
a
n
o
v
e
l
d
e
e
p
lea
rn
in
g
(DL)
b
a
se
d
p
a
ra
d
ig
m
,
wh
ich
in
teg
ra
tes
a
h
y
b
rid
fe
a
tu
re
e
x
trac
ti
o
n
a
lg
o
rit
h
m
,
wit
h
th
e
S
p
a
ti
a
l
a
n
d
Ch
a
n
n
e
l
a
tt
e
n
ti
o
n
m
e
c
h
a
n
ism
to
r
e
fin
e
th
e
fe
a
tu
re
s
a
n
d
fo
r
re
li
a
b
le
d
e
tec
ti
o
n
o
f
s
p
e
e
c
h
d
isflu
e
n
c
y
.
Th
is
stu
d
y
is
c
o
n
d
u
c
ted
o
n
m
u
lt
i
p
le
st
u
tt
e
r
d
a
ta
se
t
wh
ich
in
c
l
u
d
e
s
UCLASS
(Re
lea
se
1
,
Re
lea
se
2
),
F
lu
e
n
c
y
Ba
n
k
a
n
d
S
EP
-
2
8
k
.
Th
e
m
a
jo
r
d
ra
wb
a
c
k
o
f
a
ll
t
h
e
se
d
a
ta
se
ts
is
d
a
ta
imb
a
lan
c
e
.
To
r
e
d
u
c
e
th
is
imb
a
lan
c
e
,
th
e
a
u
th
o
r
u
se
d
d
a
ta
a
u
g
m
e
n
tati
o
n
tec
h
n
i
q
u
e
s
,
wh
ich
in
c
l
u
d
e
s
,
n
o
ise
,
m
u
sic
,
re
v
e
r
b
e
ra
ti
o
n
a
n
d
p
it
c
h
sh
ift
in
g
m
e
t
h
o
d
s.
Ho
we
v
e
r,
i
n
c
re
a
sin
g
th
e
stu
tt
e
r
d
e
tec
ti
o
n
a
c
c
u
ra
c
y
re
m
a
in
s
a
c
h
a
ll
e
n
g
in
g
issu
e
.
T
o
a
d
d
re
ss
th
i
s
issu
e
,
th
e
a
u
t
h
o
r
p
ro
p
o
se
d
a
h
y
b
r
id
fe
a
tu
re
e
x
trac
ti
o
n
m
o
d
e
l
,
wh
ic
h
e
x
trac
ts
tem
p
o
ra
l,
c
o
n
tex
t
u
a
l,
sp
e
c
tral,
a
n
d
p
i
tch
i
n
fo
rm
a
ti
o
n
fr
o
m
th
e
sp
e
e
c
h
sig
n
a
l.
Th
e
o
b
tain
e
d
fe
a
tu
re
s
a
re
th
e
n
p
ro
c
e
ss
e
d
th
ro
u
g
h
t
h
e
a
tt
e
n
ti
o
n
m
e
c
h
a
n
ism
wh
e
re
c
h
a
n
n
e
l
a
n
d
s
p
a
ti
a
l
a
tt
e
n
ti
o
n
m
o
d
e
ls
h
e
l
p
t
o
re
fin
e
th
e
fe
a
tu
re
s.
F
in
a
ll
y
,
a
m
u
lt
icla
ss
c
o
n
v
o
lu
ti
o
n
a
l
n
e
u
ra
l
n
e
two
rk
(CNN
)
c
las
sifier
is
u
se
d
to
d
e
tec
t
t
h
e
stu
t
ter
e
v
e
n
t
in
th
e
sp
e
e
c
h
sig
n
a
ls.
Th
e
re
su
lt
s
sh
o
w
th
a
t
o
u
r
m
o
d
e
l
wit
h
sp
a
ti
a
l
a
n
d
c
h
a
n
n
e
l
a
tt
e
n
ti
o
n
m
e
c
h
a
n
ism
p
e
rf
o
rm
s
b
e
tt
e
r
th
a
n
e
x
isti
n
g
d
e
e
p
lea
rn
in
g
a
p
p
ro
a
c
h
e
s
a
n
d
a
c
c
u
ra
tely
d
e
tec
ts
stu
tt
e
ri
n
g
.
K
ey
w
o
r
d
s
:
C
las
s
if
icatio
n
Featu
r
e
ex
tr
ac
tio
n
Flu
en
cy
B
an
k
Mu
lticlas
s
C
NN
SEP
-
28k
Stu
tter
T
h
is i
s
a
n
o
p
e
n
a
c
c
e
ss
a
rticle
u
n
d
e
r th
e
CC B
Y
-
SA
li
c
e
n
se
.
C
o
r
r
e
s
p
o
nd
ing
A
uth
o
r
:
Ku
s
u
m
a
H
.
R
.
Dep
ar
tm
en
t o
f
E
lectr
o
n
ics an
d
C
o
m
m
u
n
icatio
n
,
Glo
b
al
Aca
d
em
y
o
f
T
ec
h
n
o
lo
g
y
B
en
g
alu
r
u
,
I
n
d
ia
E
m
ail:
k
u
s
u
m
ar
am
an
n
a.
h
@
g
m
ail.
co
m
1.
I
NT
RO
D
UCT
I
O
N
Sp
e
ec
h
,
as
a
p
r
i
m
a
r
y
m
o
d
e
o
f
co
m
m
u
n
i
ca
t
io
n
am
o
n
g
h
u
m
a
n
s
,
p
l
a
y
s
v
it
al
r
o
le
i
n
h
u
m
a
n
i
n
te
r
ac
ti
o
n
.
No
w
a
d
a
y
s
a
u
t
o
m
ati
c
s
p
ee
c
h
r
e
co
g
n
it
io
n
(
ASR
)
s
y
s
te
m
s
a
r
e
w
id
el
y
u
s
e
d
.
T
h
e
y
a
r
e
u
s
e
d
i
n
m
o
s
t
o
f
t
h
e
s
p
ee
ch
r
el
ate
d
ap
p
l
ic
ati
o
n
s
,
f
r
o
m
v
i
r
t
u
al
la
n
g
u
a
g
e
ass
is
t
an
ts
t
o
r
ea
l
ti
m
e
v
o
ic
e
t
o
te
x
t
tr
a
n
s
c
r
i
p
ti
o
n
s
er
v
i
ce
s
,
es
p
e
cia
ll
y
wh
e
r
e
s
u
b
t
itl
es
a
r
e
n
e
ed
e
d
t
o
h
el
p
u
n
d
e
r
s
ta
n
d
s
p
o
k
e
n
co
n
t
e
n
t
.
T
h
es
e
s
p
ee
c
h
t
o
te
x
t
c
o
n
v
e
r
te
r
s
p
r
ec
is
el
y
d
o
es
th
e
c
o
n
v
e
r
s
i
o
n
b
y
l
ea
r
n
i
n
g
n
at
u
r
al
s
p
ee
ch
p
at
te
r
n
s
b
as
ed
o
n
p
h
o
n
e
tic
a
n
d
li
n
g
u
is
tic
s
t
r
u
ct
u
r
es
[
1
]
.
Des
p
i
te
t
h
e
r
a
p
i
d
g
r
o
w
th
o
f
ASR
s
y
s
t
em
s
,
th
ey
a
r
e
s
t
ill
f
a
ci
n
g
n
o
t
a
b
le
c
h
al
le
n
g
es
,
p
a
r
ti
cu
la
r
l
y
i
n
n
o
is
y
c
o
n
d
iti
o
n
s
,
w
h
e
n
d
e
ali
n
g
w
it
h
d
i
f
f
er
e
n
t
s
p
e
a
k
e
r
s
,
a
n
d
m
o
s
t
im
p
o
r
ta
n
tl
y
,
w
h
e
n
p
r
o
ce
s
s
in
g
d
is
f
l
u
en
t
s
p
e
ec
h
.
Dis
f
l
u
en
cies
s
u
c
h
as
r
e
p
e
titi
o
n
s
,
p
r
o
l
o
n
g
at
io
n
s
,
in
t
er
j
ec
t
io
n
s
,
a
n
d
a
b
n
o
r
m
a
l
p
a
u
s
es
i
n
t
er
r
u
p
t
th
e
n
o
r
m
al
f
l
o
w
o
f
s
p
ee
ch
,
w
h
ic
h
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
S
p
a
tia
l a
n
d
ch
a
n
n
el
a
tten
tio
n
mec
h
a
n
is
m
fo
r
s
p
ee
ch
d
is
flu
en
cy
d
etec
tio
n
…
(
K
u
s
u
ma
H
.
R
.
)
2107
n
e
g
a
ti
v
el
y
a
f
f
ec
ts
ASR
p
e
r
f
o
r
m
an
ce
[
2
]
.
B
ec
a
u
s
e
tr
ad
iti
o
n
al
ASR
m
o
d
els
a
r
e
p
r
i
m
a
r
i
l
y
tr
ai
n
e
d
o
n
f
l
u
e
n
t
s
p
e
ec
h
p
att
e
r
n
s
,
t
h
e
y
s
t
r
u
g
g
l
e
to
u
n
d
er
s
ta
n
d
d
is
f
l
u
en
t
s
p
ee
ch
e
f
f
e
cti
v
e
ly
.
As
a
r
esu
l
t,
th
ese
ir
r
e
g
u
la
r
it
ies
o
f
te
n
lea
d
to
t
r
a
n
s
c
r
i
p
ti
o
n
er
r
o
r
s
,
a
f
f
ec
ti
n
g
th
e
r
e
lia
b
il
it
y
o
f
t
h
e
o
u
t
p
u
t
i
n
r
ea
l
-
w
o
r
l
d
ap
p
l
ica
ti
o
n
s
[
3
]
.
T
h
e
i
n
ef
f
i
cie
n
t
b
e
h
a
v
i
o
r
o
f
t
h
e
A
SR
s
y
s
te
m
s
ig
n
it
e
d
r
ese
ar
c
h
i
n
t
er
est
i
n
s
p
e
ec
h
p
r
o
ce
s
s
i
n
g
t
ec
h
n
i
q
u
es
t
h
at
ca
n
d
ete
ct
an
d
h
a
n
d
le
d
is
f
l
u
e
n
cies
ef
f
e
cti
v
el
y
.
B
y
i
n
t
e
g
r
ati
n
g
s
u
c
h
m
o
d
els
i
n
t
o
ASR
s
y
s
te
m
s
,
it
b
e
co
m
es
p
o
s
s
ib
le
t
o
i
m
p
r
o
v
e
b
o
t
h
r
ec
o
g
n
iti
o
n
ac
c
u
r
a
cy
a
n
d
t
h
e
o
v
er
all
u
s
er
e
x
p
e
r
ie
n
ce
in
s
p
e
ec
h
-
d
r
i
v
en
a
p
p
lic
ati
o
n
s
.
Fo
r
m
a
n
y
s
p
ee
c
h
-
b
as
ed
a
p
p
li
ca
t
io
n
s
,
it
m
i
g
h
t
b
e
h
el
p
f
u
l
t
o
k
n
o
w
w
h
et
h
er
d
is
f
l
u
e
n
c
ies
a
r
e
p
r
ese
n
t
a
n
d
h
o
w
lo
n
g
t
h
ey
l
ast
[
4
]
.
Sp
ee
ch
d
is
o
r
d
er
s
ar
e
g
e
n
er
al
ly
class
if
ied
in
to
f
o
llo
win
g
ca
teg
o
r
ies
s
u
ch
as
d
y
s
ar
th
r
ia
,
ap
r
a
x
ia,
clu
tter
in
g
,
lis
p
in
g
,
a
n
d
s
tu
tter
in
g
[
5
]
.
Ho
wev
er
,
th
e
s
tu
tter
in
g
is
a
co
m
m
o
n
t
y
p
e
o
f
s
p
ee
ch
d
is
f
lu
en
cy
,
an
d
it
is
also
ca
lled
as
s
tam
m
er
in
g
an
d
it
af
f
ec
ts
o
v
er
all
1
%
o
f
p
o
p
u
l
atio
n
g
lo
b
ally
[
6
]
,
th
is
m
ea
n
s
ar
o
u
n
d
7
0
m
illi
o
n
p
eo
p
le
wo
r
ld
wid
e.
Dev
elo
p
m
en
tal
s
tu
tter
in
g
s
tar
ts
in
ch
ild
r
en
ag
e
ar
o
u
n
d
2
to
6
y
ea
r
s
,
in
s
p
ee
ch
d
ev
el
o
p
in
g
an
d
lan
g
u
ag
e
d
e
v
elo
p
m
e
n
t
s
ta
g
e.
Gen
er
all
y
,
5
%
-
1
0
%
o
f
c
h
i
ld
r
en
will
s
tu
tter
in
th
e
lear
n
i
n
g
s
tag
e,
b
u
t
m
o
s
t
will
r
ec
o
v
er
with
o
u
t
an
y
ass
is
tan
ce
ev
en
tu
ally
.
Ap
p
r
o
x
im
ately
2
5
%
o
f
ch
ild
r
en
wh
o
h
av
e
th
is
co
n
d
itio
n
p
er
s
is
ted
in
to
ad
u
lth
o
o
d
,
b
ec
o
m
in
g
a
p
er
m
a
n
en
t
ch
allen
g
e
.
W
h
en
it
co
m
es
to
g
en
d
er
,
s
tu
tter
in
g
is
m
o
r
e
co
m
m
o
n
in
m
ales
th
a
n
f
em
al
es,
with
a
4
:1
r
ati
o
.
T
h
e
m
aj
o
r
ch
ar
ac
te
r
is
tics
o
f
th
e
s
tu
tter
in
g
co
n
tain
s
o
u
n
d
r
ep
etitio
n
s
,
ab
n
o
r
m
al
p
a
u
s
es,
p
r
o
lo
n
g
atio
n
s
,
s
o
u
n
d
r
e
p
etitio
n
,
wo
r
d
r
ep
etitio
n
,
in
ter
jectio
n
s
,
an
d
b
lo
c
k
s
wh
ile
s
p
ea
k
in
g
.
T
h
e
s
tu
tter
in
g
s
y
m
p
to
m
s
ar
e
n
o
t
s
am
e
i
n
in
d
iv
i
d
u
als,
an
d
d
is
o
r
d
e
r
p
atter
n
is
also
d
if
f
e
r
en
t
f
r
o
m
p
er
s
o
n
to
p
er
s
o
n
a
n
d
ev
en
wi
th
th
e
s
am
e
p
er
s
o
n
in
d
if
f
er
en
t
s
u
r
r
o
u
n
d
i
n
g
s
an
d
em
o
tio
n
al
co
n
d
itio
n
s
.
Du
e
to
all
th
ese
u
n
ce
r
tain
c
o
n
d
itio
n
d
is
f
lu
en
cy
d
etec
tio
n
s
is
m
o
r
e
c
h
allen
g
in
g
t
o
r
esear
ch
e
r
s
.
Mu
ch
r
esear
ch
is
g
o
in
g
on
but
ac
h
iev
in
g
s
ig
n
i
f
ican
t
m
iles
to
n
e
in
ac
cu
r
ac
y
is
s
till
ch
allen
g
in
g
.
T
h
e
s
tu
tter
i
n
g
ca
u
s
es
s
ev
er
al
ch
allen
g
es to
au
to
m
atic
s
p
ee
c
h
r
ec
o
g
n
itio
n
s
y
s
tem
,
m
o
s
t lik
ely
ad
d
r
ess
ed
ch
allen
g
es a
r
e
d
is
cu
s
s
ed
:
A
SR
s
y
s
te
m
s
a
r
e
t
r
a
i
n
e
d
o
n
h
u
g
e
f
l
u
e
n
t
s
p
e
e
c
h
d
a
ta
s
et
s
,
a
n
d
w
o
r
k
s
m
o
o
t
h
l
y
w
it
h
f
l
u
e
n
t
s
p
e
e
c
h
,
b
u
t
t
h
e
y
s
t
r
u
g
g
l
e
w
i
t
h
s
p
e
e
c
h
a
b
n
o
r
m
a
l
i
t
i
es
li
k
e
s
t
u
t
t
e
r
i
n
g
.
S
in
c
e
t
h
e
t
r
ai
n
i
n
g
d
a
t
a
h
a
r
d
l
y
c
o
n
t
a
i
n
s
r
a
r
e
s
p
e
ec
h
p
a
t
t
e
r
n
s
s
u
c
h
a
s
s
t
u
tt
e
r
i
n
g
,
d
u
e
t
o
t
h
i
s
,
t
h
e
A
SR
m
o
d
e
l
s
u
f
f
e
r
s
f
r
o
m
u
n
d
e
r
r
a
t
e
d
a
c
c
u
r
a
c
y
.
T
h
e
r
e
f
o
r
e
,
d
e
t
e
c
t
i
o
n
a
n
d
h
a
n
d
l
i
n
g
t
h
e
s
e
d
i
s
f
l
u
e
n
ci
es
is
v
e
r
y
i
m
p
o
r
t
a
n
t
t
o
i
m
p
r
o
v
e
t
h
e
a
c
c
u
r
a
c
y
o
f
A
SR
s
y
s
tem
s
.
T
h
e
d
e
t
ec
t
i
o
n
o
f
t
h
e
s
e
d
is
f
l
u
e
n
c
i
es
c
a
n
b
e
b
e
n
ef
i
c
i
a
l
f
o
r
AS
R
s
y
s
t
e
m
s
i
n
i
m
p
r
o
v
i
n
g
t
h
e
t
r
a
n
s
c
r
i
p
t
i
o
n
q
u
a
li
ty
a
n
d
i
m
p
r
o
v
i
n
g
t
h
e
A
SR
a
c
c
u
r
ac
y
.
I
n
la
n
g
u
a
g
e
le
a
r
n
i
n
g
a
n
d
s
p
e
ec
h
t
h
e
r
a
p
y
,
d
i
s
f
l
u
e
n
c
y
d
et
e
c
ti
o
n
p
r
o
v
i
d
e
s
v
a
l
u
a
b
l
e
f
e
e
d
b
a
c
k
t
o
u
s
e
r
s
a
n
d
p
r
o
v
i
d
e
s
a
d
v
i
c
e
o
n
h
o
w
t
o
i
n
c
r
e
as
e
f
l
u
e
n
c
y
.
T
h
e
s
e
s
y
s
t
e
m
s
c
a
n
b
e
a
b
l
e
t
o
f
i
n
d
in
s
t
a
n
c
es
i
n
w
h
i
c
h
a
s
p
e
a
k
e
r
h
e
s
i
t
a
te
s
o
r
c
o
r
r
e
c
ts
t
h
e
m
s
e
l
v
es
,
t
h
es
e
i
n
f
o
r
m
a
t
i
o
n
’
s
a
r
e
m
o
r
e
h
e
l
p
f
u
l
i
n
t
h
e
t
h
e
r
a
p
y
s
e
s
s
i
o
n
.
R
esear
ch
er
s
h
av
e
co
n
ce
n
tr
ate
d
o
n
d
ev
elo
p
in
g
ef
f
icien
t
tech
n
iq
u
es
to
d
etec
t
an
d
r
ec
o
g
n
i
ze
s
p
ee
ch
d
is
f
lu
en
cies,
h
en
ce
en
h
an
cin
g
th
e
p
er
f
o
r
m
a
n
ce
o
f
ASR
s
y
s
te
m
s
.
I
n
itially
,
ea
r
ly
r
u
le
-
b
ased
s
y
s
tem
s
wer
e
u
s
ed
f
o
r
d
is
f
lu
en
cy
d
etec
tio
n
an
d
r
ec
o
g
n
itio
n
.
T
h
ese
d
etec
tio
n
m
eth
o
d
s
d
e
p
en
d
e
n
t
o
n
m
a
n
u
al
a
n
n
o
tatio
n
,
in
wh
ich
s
p
ee
ch
p
ath
o
lo
g
is
ts
wo
u
ld
an
aly
ze
r
ec
o
r
d
i
n
g
s
an
d
f
in
d
d
is
f
lu
en
cies
ac
co
r
d
in
g
to
th
at,
th
e
y
estab
lis
h
ed
r
u
les
an
d
p
atter
n
s
.
T
h
e
m
o
d
e
r
n
a
u
to
m
atio
n
s
y
s
tem
s
em
p
lo
y
e
d
alg
o
r
ith
m
s
to
i
d
en
tify
d
is
f
lu
e
n
cies
v
ia
ac
o
u
s
tic
ch
ar
ac
ter
is
tics
,
d
is
f
lu
en
t
ch
ar
ac
ter
is
tics
s
u
ch
as
p
au
s
es,
r
ep
ea
ts
an
d
h
esit
atio
n
s
.
T
h
e
p
r
o
g
r
ess
io
n
in
th
e
f
ield
th
r
o
u
g
h
o
u
t
th
e
1
9
9
0
s
an
d
ea
r
l
y
2
0
0
0
s
,
in
c
o
r
p
o
r
ated
s
tatis
tical
m
o
d
els
lik
e
h
id
d
en
Ma
r
k
o
v
m
o
d
els
(
HM
Ms)
ar
e
r
aised
.
T
h
ese
m
o
d
els
f
ac
ilit
ated
in
ca
p
tu
r
in
g
th
e
s
eq
u
en
tial
ch
ar
ac
ter
is
tics
o
f
s
p
ee
ch
b
y
f
in
d
i
n
g
p
r
o
b
a
b
ilit
y
d
is
tr
ib
u
tio
n
o
f
s
p
ee
ch
.
T
h
ese
m
eth
o
d
o
lo
g
ies
s
u
f
f
er
ed
f
r
o
m
lim
itatio
n
s
s
u
ch
as
u
n
p
r
ed
ictab
ilit
y
an
d
co
n
tex
t
d
e
p
en
d
e
n
cy
.
T
h
e
2
0
0
0
s
witn
ess
ed
a
r
ap
id
g
r
o
wth
in
m
ac
h
i
n
e
lear
n
i
n
g
alg
o
r
ith
m
s
lik
e
“su
p
p
o
r
t
v
ec
to
r
m
ac
h
in
es”
(
SVMs),
w
h
ich
ef
f
icien
tly
class
if
ies
s
p
ee
ch
s
eg
m
en
t
in
to
f
lu
en
t
an
d
d
i
s
f
lu
en
t
o
n
th
e
b
asis
o
f
ex
tr
ac
te
d
f
ea
tu
r
es,
p
er
f
o
r
m
an
ce
o
f
th
ese
m
o
d
els
ar
e
g
o
o
d
,
b
u
t
n
ee
d
s
m
o
r
e
h
u
m
an
in
te
r
v
en
tio
n
f
o
r
f
ea
tu
r
e
ex
tr
ac
tio
n
,
an
d
wh
ich
is
h
a
v
in
g
lim
ited
ab
ilit
y
to
ad
o
p
t to
t
h
e
d
iv
er
s
e
s
p
ea
k
in
g
s
ty
les an
d
l
an
g
u
ag
es
Dee
p
lear
n
in
g
tech
n
iq
u
es
ar
e
in
tr
o
d
u
ce
d
in
th
e
y
ea
r
2
0
1
0
s
an
d
th
er
e
is
a
m
ajo
r
b
r
ea
k
t
h
r
o
u
g
h
in
d
etec
tio
n
o
f
d
is
f
lu
e
n
cy
.
T
h
ese
tech
n
iq
u
es
co
u
ld
a
u
to
m
atica
lly
lear
n
th
e
p
atter
n
o
f
th
e
d
i
v
er
s
e
s
p
ee
ch
p
atter
n
s
an
d
ex
tr
ac
t
th
e
r
elev
a
n
t
f
ea
tu
r
es.
Few
ex
am
p
les
f
o
r
d
ee
p
le
ar
n
in
g
n
etwo
r
k
s
ar
e
“r
ec
u
r
r
en
t
n
eu
r
al
n
etwo
r
k
s
”
(
R
NNs)
an
d
“lo
n
g
s
h
o
r
t
-
te
r
m
m
em
o
r
y
”
(
L
STM
)
n
etwo
r
k
s
.
T
h
ese
tech
n
i
q
u
es
r
e
d
u
ce
h
u
m
an
in
ter
v
en
tio
n
in
f
ea
tu
r
e
ex
tr
ac
tio
n
a
n
d
n
etwo
r
k
s
ar
e
ef
f
icien
t
to
ca
p
tu
r
e
t
h
e
co
n
tex
tu
al
an
d
s
p
atial
in
f
o
r
m
atio
n
o
f
s
p
ee
ch
wh
ich
im
p
r
o
v
es
th
e
d
etec
tio
n
o
f
d
is
f
lu
en
cies.
T
h
e
m
o
d
el
p
er
f
o
r
m
a
n
ce
ca
n
b
e
in
cr
ea
s
ed
as
lo
n
g
as
y
o
u
p
r
o
v
id
e
atten
tio
n
to
m
o
s
t
r
ele
v
an
t
p
ar
ts
o
f
th
e
s
p
ee
ch
s
ig
n
al
,
it
lead
s
to
b
etter
id
en
tific
atio
n
o
f
d
is
f
lu
en
cies
in
b
etwe
en
f
lu
en
t sp
ee
c
h
.
Pre
s
en
tly
,
co
n
v
o
l
u
tio
n
n
e
u
r
al
n
etwo
r
k
s
-
b
ased
m
o
d
els
h
av
e
g
ain
ed
m
ass
iv
e
atten
tio
n
i
n
s
p
ee
ch
p
r
o
ce
s
s
in
g
s
y
s
tem
b
ec
au
s
e
o
f
th
eir
co
m
p
eten
ce
to
lear
n
th
e
f
ea
tu
r
e
v
ar
iatio
n
ef
f
icie
n
tly
.
I
n
th
is
p
ap
er
r
esear
ch
er
s
u
s
ed
a
n
o
v
el
d
ee
p
lear
n
in
g
(
DL
)
a
p
p
r
o
ac
h
an
d
b
r
o
u
g
h
t
o
u
t
n
ew
ar
ch
itectu
r
e
f
o
r
s
tu
tter
d
etec
tio
n
.
T
h
e
m
ain
s
tep
s
f
o
llo
wed
in
th
i
s
n
ew
ar
ch
itectu
r
e
ar
e
as f
o
llo
ws.
a.
B
y
d
o
wn
s
am
p
lin
g
th
e
d
o
m
in
an
t
class
f
o
r
ea
ch
task
—
wh
ich
is
s
till
th
e
f
lu
en
t
class
—
we
ar
e
ab
le
to
f
ix
th
e
class
im
b
alan
ce
is
s
u
e.
W
e
h
a
v
e
lim
itatio
n
s
with
th
e
s
p
ee
c
h
d
is
f
lu
en
c
y
d
atasets
th
er
ef
o
r
e
ex
p
e
r
im
en
tin
g
with
“d
ee
p
n
eu
r
al
n
etwo
r
k
(
DNN)
”
m
eth
o
d
s
b
ec
o
m
e
ch
allen
g
in
g
is
s
u
e.
T
o
ad
d
r
ess
th
is
is
s
u
e,
we
h
av
e
in
co
r
p
o
r
ated
a
r
o
b
u
s
t
d
ata
au
g
m
en
tatio
n
m
eth
o
d
wh
ich
c
o
n
s
id
er
s
n
o
is
e,
r
ev
e
r
b
er
atio
n
,
m
u
s
ic
an
d
p
itch
s
h
if
t to
p
r
o
d
u
ce
th
e
a
u
g
m
e
n
ted
d
ata.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
1
6
,
No
.
4
,
Au
g
u
s
t
20
2
6
:
2
1
0
6
-
2
119
2108
b.
I
n
n
ex
t stag
e,
we
f
o
cu
s
o
n
in
tr
o
d
u
cin
g
a
n
ew
“f
ea
tu
r
e
ex
tr
ac
t
io
n
”
m
eth
o
d
wh
er
e
we
e
x
tr
ac
t
s
p
ec
tr
al,
p
itch
,
tem
p
o
r
al
an
d
co
n
te
x
tu
al
f
ea
tu
r
e.
c.
I
n
o
r
d
e
r
to
r
ef
in
e
th
ese
f
ea
tu
r
es,
we
in
co
r
p
o
r
ate
atten
tio
n
m
ec
h
an
is
m
s
wh
er
e
f
ea
tu
r
es
ar
e
r
ef
in
ed
th
r
o
u
g
h
ch
an
n
el
an
d
s
p
atial
atten
tio
n
m
ec
h
an
is
m
s
.
T
h
e
f
in
al
o
b
tain
ed
f
ea
tu
r
es a
r
e
f
u
s
ed
an
d
p
r
o
c
ess
ed
th
r
o
u
g
h
th
e
m
u
lticlas
s
C
NN
clas
s
if
ier
to
g
et
th
e
f
in
al
o
u
tc
o
m
e.
R
em
ain
in
g
p
ar
t
o
f
th
is
ar
ticl
e
is
s
tr
u
ctu
r
ed
as:
s
ec
tio
n
2
p
r
esen
ts
a
b
r
ie
f
d
is
cu
s
s
io
n
o
n
ex
is
tin
g
m
eth
o
d
s
o
f
s
p
ee
ch
d
is
f
lu
en
cy
d
etec
tio
n
,
s
ec
tio
n
3
illu
s
tr
ates th
e
p
r
o
p
o
s
ed
DL
b
ased
s
o
lu
tio
n
f
o
r
th
is
r
esear
ch
,
s
ec
tio
n
4
p
r
esen
ts
th
e
o
u
tco
m
e
o
f
p
r
o
p
o
s
e
m
o
d
el
a
n
d
c
o
m
p
ar
ativ
e
an
al
y
s
is
with
s
tan
d
ar
d
m
eth
o
d
s
,
last
ly
,
s
ec
tio
n
5
co
m
p
letes th
e
ar
ticle
b
y
p
r
esen
tin
g
co
n
clu
d
in
g
r
em
ar
k
s
ab
o
u
t t
h
is
r
esear
ch
.
2.
L
I
T
E
R
AT
U
RE
R
E
VI
E
W
T
h
is
s
ec
tio
n
p
r
esen
ts
an
o
v
er
v
iew
ab
o
u
t
ex
is
tin
g
m
eth
o
d
s
f
o
r
s
p
ee
c
h
an
al
y
s
is
an
d
s
tu
tte
r
d
etec
tio
n
b
y
em
p
l
o
y
in
g
m
ac
h
in
e
lear
n
i
n
g
an
d
d
ee
p
lear
n
in
g
f
r
a
m
ewo
r
k
.
T
h
e
tr
ad
itio
n
al
m
eth
o
d
s
m
an
u
ally
an
al
y
ze
th
e
in
s
tan
ce
s
o
f
d
if
f
er
en
t ty
p
es o
f
s
tu
tter
in
g
an
d
ex
p
r
ess
as r
atio
o
f
s
tu
tter
ev
en
t to
to
tal
wo
r
d
s
in
s
p
ee
ch
s
eg
m
en
t.
Ho
wev
er
,
th
is
m
an
u
al
p
r
o
ce
s
s
in
cr
ea
s
es
tim
e
an
d
co
m
p
le
x
ities
,
m
o
r
eo
v
er
,
it
r
eq
u
ir
es
h
u
m
an
in
v
o
lv
em
en
t
th
u
s
it
is
m
o
r
e
ch
allen
g
in
g
is
s
u
e
to
ad
d
r
ess
.
T
h
e
m
a
n
u
al
d
et
ec
tio
n
o
f
s
tu
tter
in
g
p
o
s
es
s
ev
er
al
ch
allen
g
es
s
u
ch
as
d
is
tin
g
u
is
h
in
g
s
tu
tter
in
g
a
m
o
n
g
o
th
e
r
s
p
ee
ch
d
is
f
lu
e
n
cies.
Mo
r
eo
v
er
,
co
n
s
is
ten
t
d
ete
ctio
n
o
f
s
tu
tter
in
g
ac
cu
r
ac
y
is
af
f
ec
ted
to
th
e
s
ev
er
ity
an
d
s
tu
tter
in
g
f
r
eq
u
en
cy
wh
ich
also
v
ar
ies
f
o
r
ea
c
h
s
p
e
ak
er
.
Fu
r
th
e
r
m
o
r
e,
th
e
ag
e
o
f
s
p
ea
k
er
,
g
e
n
d
er
an
d
lan
g
u
ag
e
also
co
m
p
licates
th
e
s
tu
tter
id
en
tific
atio
n
p
r
o
ce
s
s
.
T
h
er
ef
o
r
e
,
r
esear
ch
er
s
h
av
e
u
s
ed
m
ac
h
i
n
e
lear
n
in
g
f
r
am
ewo
r
k
s
f
o
r
v
ar
io
u
s
s
p
ee
ch
p
r
o
ce
s
s
in
g
task
s
s
u
ch
as
s
p
ee
ch
f
ilter
in
g
,
s
p
ee
ch
r
ec
o
g
n
itio
n
,
s
p
ee
ch
d
en
o
is
in
g
.
T
h
e
ef
f
icie
n
t
p
atter
n
lea
r
n
in
g
n
atu
r
e
o
f
m
ac
h
in
e
an
d
d
ee
p
lear
n
in
g
ap
p
r
o
ac
h
es h
as a
ttra
cted
r
esear
ch
er
s
to
im
p
lem
en
t i
t f
o
r
s
tu
tter
d
etec
tio
n
.
R
ec
en
t
s
tu
d
ies
o
n
au
to
m
atic
s
tu
tter
d
etec
tio
n
h
av
e
b
ee
n
r
elied
o
n
d
ee
p
lear
n
in
g
f
r
am
e
wo
r
k
s
th
at
u
s
es
ac
o
u
s
tic
f
ea
tu
r
es
an
d
tem
p
o
r
al
m
o
d
ellin
g
.
Ko
u
r
k
o
u
n
ak
is
et
a
l.
[
7
]
p
r
o
p
o
s
ed
,
“Fl
u
en
tNet
,
”
a
h
y
b
r
i
d
f
r
am
ewo
r
k
co
n
s
is
ts
o
f
“Sq
u
ee
ze
-
an
d
-
ex
citatio
n
r
esid
u
al
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etw
o
r
k
”
alo
n
g
with
a
b
id
ir
ec
tio
n
al
lo
n
g
-
s
h
o
r
t
ter
m
m
em
o
r
y
(
B
i
-
L
STM
)
lay
er
s
.
T
h
is
p
er
f
o
r
m
a
n
ce
o
f
th
is
m
o
d
el
m
ain
ly
f
o
cu
s
s
es
o
n
ac
o
u
s
tic
q
u
ality
o
f
th
e
d
ata
in
p
u
t
an
d
ef
f
icien
tl
y
ca
p
tu
r
es
b
o
th
tem
p
o
r
al
an
d
s
p
ec
tr
al
f
ea
tu
r
es
o
f
th
e
s
tu
tter
s
p
ee
ch
.
He
also
in
tr
o
d
u
ce
d
a
d
ee
p
lear
n
in
g
-
b
ased
s
o
lu
tio
n
to
ad
d
r
ess
th
e
d
ata
im
b
alan
ce
is
s
u
e
ex
is
t
in
th
e
s
tu
tter
d
atasets
.
Similar
ly
,
Sh
eik
h
et
a
l.
[
8
]
,
[
9
]
i
n
tr
o
d
u
ce
d
a
d
ee
p
lear
n
in
g
f
r
am
ewo
r
k
f
o
r
s
tu
tter
d
etec
tio
n
b
y
h
ig
h
lig
h
tin
g
im
p
ac
t
o
f
m
u
lti
-
task
lear
n
in
g
an
d
ad
v
er
s
ar
ia
l
lear
n
in
g
to
lear
n
th
e
r
o
b
u
s
t
s
tu
tter
attr
ib
u
tes.
Ho
wev
er
,
b
o
th
th
e
s
tu
d
ies
ar
e
lim
ited
t
o
co
n
tr
o
lled
en
v
ir
o
n
m
e
n
ts
,
n
o
t
co
m
p
atib
le
with
u
n
co
n
tr
o
lled
en
v
ir
o
n
m
en
ts
lik
e
n
o
is
y
e
n
v
ir
o
n
m
en
ts
.
Ma
n
y
wo
r
k
s
h
av
e
b
ee
n
f
o
c
u
s
ed
o
n
f
ea
tu
r
e
r
ep
r
esen
tatio
n
b
y
in
teg
r
atin
g
atten
tio
n
m
ec
h
an
is
m
an
d
m
u
lti
-
f
ea
tu
r
e
f
u
s
io
n
.
Al
-
B
an
n
a
et
a
l.
[
1
0
]
in
tr
o
d
u
ce
d
s
p
a
tial
an
d
tem
p
o
r
al
atten
tio
n
-
b
ased
f
r
am
ewo
r
k
in
wh
ich
,
th
e
f
ea
tu
r
e
ex
tr
ac
tio
n
m
ain
ly
f
o
cu
s
es
o
n
ac
o
u
s
tic
f
ea
tu
r
es
b
y
co
n
s
id
er
i
n
g
d
if
f
er
en
t
p
itch
,
tim
e
a
n
d
f
r
eq
u
e
n
cy
d
o
m
ain
f
ea
tu
r
es.
Fu
r
th
er
,
s
p
atial
an
d
tem
p
o
r
al
atten
tio
n
m
ec
h
an
is
m
s
ar
e
also
in
co
r
p
o
r
ated
to
o
b
tain
r
o
b
u
s
t
f
ea
tu
r
e
r
ep
r
esen
tatio
n
.
T
o
in
cr
ea
s
e
th
e
lear
n
i
n
g
p
er
f
o
r
m
an
ce
b
id
ir
ec
tio
n
al
L
STM
(
B
i
-
L
STM
)
class
if
ier
is
u
s
ed
.
Simh
a
et
a
l.
[
1
1
]
p
r
o
p
o
s
ed
f
ea
tu
r
e
s
p
ac
e
-
b
ased
class
if
icatio
n
m
o
d
el
t
o
d
etec
t
th
e
s
tu
tter
ev
en
ts
in
s
p
ee
ch
in
w
h
ich
.
T
h
ey
u
s
ed
“z
e
r
o
-
tim
e
win
d
o
w
in
g
ce
p
s
tr
al
co
ef
f
icien
ts
(
Z
T
W
C
C
)
”
f
o
r
f
ea
tu
r
e
ex
tr
ac
tio
n
,
th
e
e
x
tr
ac
ted
f
ea
t
u
r
es
ar
e
p
r
o
ce
s
s
ed
th
r
o
u
g
h
m
u
ltip
le
class
if
icatio
n
alg
o
r
ith
m
,
s
u
ch
as
L
STM
,
SVM
an
d
B
i
-
L
STM
to
m
ea
s
u
r
e
th
e
o
v
e
r
all
p
e
r
f
o
r
m
an
ce
t
h
is
m
o
d
el
u
s
ed
h
an
d
cr
a
f
ted
a
co
u
s
tic
f
ea
tu
r
es
f
o
r
th
is
s
tu
d
y
,
an
d
it is
n
o
t
f
lex
ib
l
e.
I
n
co
n
tr
ast
to
co
n
v
en
tio
n
al
ac
o
u
s
tic
b
ased
m
eth
o
d
s
,
Das
et
a
l.
[
1
2
]
h
i
g
h
lig
h
ted
th
e
k
ey
li
m
itatio
n
s
o
f
th
e
ex
is
tin
g
tech
n
i
q
u
es,
s
u
ch
a
s
th
eir
d
ep
en
d
en
cy
o
n
th
e
r
ec
o
r
d
in
g
c
o
n
d
itio
n
s
.
I
n
o
r
d
e
r
to
d
ea
l
with
th
is
is
s
u
e,
au
th
o
r
s
in
co
r
p
o
r
ated
v
is
u
al
cu
es
b
y
ca
p
tu
r
in
g
th
e
f
ac
ial
m
u
s
cle
m
o
v
em
en
t
u
s
in
g
a
C
NN
b
ased
ap
p
r
o
ac
h
to
p
r
ed
ict
th
e
s
tu
tter
in
g
f
ea
tu
r
e
s
an
d
d
em
o
n
s
tr
atin
g
th
e
ch
e
ek
an
d
lip
d
y
n
am
ic
m
o
v
e
m
en
ts
to
p
r
ed
ict
t
h
e
u
p
co
m
in
g
u
tter
an
ce
an
d
s
tu
tte
r
ev
en
t.
Alth
o
u
g
h
t
h
is
m
u
ltimo
d
al
ap
p
r
o
ac
h
im
p
r
o
v
es
ef
f
icien
cy
,
it
in
tr
o
d
u
ce
s
h
ar
d
war
e
co
m
p
lex
ity
in
o
th
er
h
an
d
,
s
o
it
is
m
ajo
r
ch
allen
g
e
to
im
p
lem
en
t
in
r
ea
l
wo
r
l
d
.
Fu
r
th
er
ex
te
n
s
io
n
o
f
d
ee
p
lear
n
in
g
alg
o
r
ith
m
f
o
r
s
tu
tter
ev
e
n
t
d
etec
tio
n
,
Al
-
B
an
n
a
et
a
l
.
[
1
3
]
em
p
l
o
y
ed
lo
g
Me
l
s
p
ec
tr
o
g
r
am
al
o
n
g
with
2
D
atr
o
u
s
co
n
v
o
lu
tio
n
d
esig
n
wh
ich
h
elp
s
to
lear
n
b
o
th
s
p
ec
tr
al
an
d
tem
p
o
r
al
attr
ib
u
tes
f
o
r
th
e
g
iv
en
s
p
ee
ch
s
ig
n
al.
O
n
o
th
er
h
a
n
d
Sh
eik
h
et
a
l.
[
1
4
]
i
n
tr
o
d
u
ce
d
an
o
th
e
r
d
ee
p
lea
r
n
in
g
-
b
ased
ar
c
h
itectu
r
e
S
tu
tter
N
et
wh
ich
is
u
s
ed
to
tim
e
d
elay
n
e
u
r
al
n
etwo
r
k
to
c
ap
tu
r
e
th
e
c
o
n
tex
tu
al
i
n
f
o
r
m
at
io
n
ab
o
u
t
th
e
g
i
v
en
s
p
ee
ch
s
ig
n
al,
th
is
wo
r
k
m
ain
l
y
f
o
cu
s
es o
n
th
e
ac
o
u
s
tic
s
ig
n
al
an
aly
s
is
.
W
ah
ee
d
et
a
l.
[
1
5
]
p
r
esen
te
d
a
“b
r
ain
c
o
m
p
u
ter
in
ter
f
ac
e”
(
B
C
I
)
b
ased
m
o
d
el
to
a
n
aly
ze
th
e
elec
tr
o
en
ce
p
h
al
o
g
r
am
(
E
E
G)
s
ig
n
al
o
f
t
h
e
p
e
r
s
o
n
wh
o
s
tu
tt
er
(
PW
S)
in
o
r
d
er
to
id
e
n
tify
t
h
e
th
at
wh
at
k
in
d
o
f
em
o
tio
n
s
ar
e
d
o
m
in
ati
n
g
f
o
r
th
e
PW
S
ca
s
es.
An
E
E
G
b
as
ed
m
in
d
-
wav
e
s
en
s
o
r
was
u
s
ed
to
ac
q
u
ir
e
n
e
u
r
al
s
ig
n
als
u
s
in
g
th
e
E
E
G
m
a
ch
in
e
f
o
r
f
u
r
th
er
an
aly
s
is
b
y
h
i
g
h
lig
h
tin
g
th
e
d
if
f
er
e
n
t
p
o
ten
tials
o
f
th
e
p
h
y
s
io
lo
g
ical
s
ig
n
als
to
u
n
d
e
r
s
tan
d
s
tu
tter
r
elate
d
co
g
n
itiv
e
an
d
em
o
tio
n
al
p
atter
n
.
Sh
o
n
ib
ar
e
et
a
l.
[
1
6
]
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
S
p
a
tia
l a
n
d
ch
a
n
n
el
a
tten
tio
n
mec
h
a
n
is
m
fo
r
s
p
ee
ch
d
is
flu
en
cy
d
etec
tio
n
…
(
K
u
s
u
ma
H
.
R
.
)
2109
p
r
o
p
o
s
ed
a
m
eth
o
d
ca
lled
as
“d
etec
t
an
d
p
ass
”
to
im
p
r
o
v
e
th
e
au
to
m
atic
s
p
ee
ch
r
ec
o
g
n
itio
n
s
y
s
tem
f
o
r
PW
S,
wh
ich
is
tr
ain
ed
o
n
lim
ited
d
ata
an
d
f
o
cu
s
o
n
ly
o
n
ac
o
u
s
tic
f
ea
tu
r
es
o
f
th
e
s
tu
tter
s
p
ee
ch
.
Similar
ly
,
Ko
u
r
k
o
u
n
ak
is
et
a
l.
[
1
7
]
d
e
v
elo
p
ed
d
ee
p
r
esid
u
al
d
ee
p
le
ar
n
in
g
f
r
am
ewo
r
k
,
wh
ich
is
a
ls
o
d
ep
en
d
s
o
n
th
e
ac
o
u
s
tic
f
ea
tu
r
es
an
d
wh
ich
em
p
lo
y
e
d
B
id
ir
ec
tio
n
al
lo
n
g
s
h
o
r
t
-
ter
m
m
em
o
r
y
(
B
i
-
L
STM
)
alg
o
r
ith
m
to
class
if
y
s
tu
tter
d
is
f
lu
en
cies.
Alh
ar
b
i
et
a
l.
[
1
8
]
in
tr
o
d
u
ce
d
a
lig
h
tly
s
u
p
er
v
is
ed
ap
p
r
o
a
ch
f
o
r
d
etec
tio
n
o
f
s
tu
tter
s
p
ee
ch
in
ch
ild
r
e
n
’
s
s
p
ee
ch
.
Ov
er
all
,
ex
is
tin
g
s
tu
d
ies
d
em
o
n
s
tr
ated
s
ig
n
if
ican
t
p
r
o
g
r
ess
in
s
tu
tter
s
p
ee
ch
d
etec
tio
n
u
s
in
g
m
ac
h
in
e
lear
n
in
g
a
n
d
d
ee
p
lea
r
n
in
g
tech
n
iq
u
es.
Ho
wev
e
r
,
th
e
m
ajo
r
ity
o
f
th
ese
ap
p
r
o
ac
h
es
ar
e
p
r
e
d
o
m
in
a
n
tly
d
ep
en
d
i
n
g
m
ain
ly
o
n
ac
o
u
s
tic
f
ea
tu
r
es,
an
d
th
ese
f
ea
tu
r
es
ar
e
m
o
r
e
o
f
ten
ev
alu
ated
u
n
d
er
th
e
co
n
tr
o
lled
en
v
ir
o
n
m
en
t,
th
ese
ar
e
th
e
m
ain
lim
itatio
n
s
o
f
th
e
ex
is
tin
g
s
tu
d
ies.
Mo
r
eo
v
er
,
b
y
g
i
v
in
g
litt
le
atten
tio
n
to
lig
h
tweig
h
t
ar
ch
itectu
r
es
is
m
ak
in
g
m
o
d
el
m
o
r
e
ca
p
ab
le
o
f
ca
p
tu
r
in
g
b
o
th
s
p
ec
tr
al
a
n
d
s
p
atial
-
tem
p
o
r
al
s
p
ee
ch
ch
ar
ac
ter
is
tics
.
T
h
ese
ar
e
th
e
m
ain
lim
itatio
n
s
,
wh
ich
em
p
h
asize
th
e
n
ee
d
f
o
r
t
h
e
r
o
b
u
s
t
an
d
g
en
er
alize
d
f
r
am
ewo
r
k
s
wh
ic
h
ar
e
m
o
r
e
r
eliab
le
ac
r
o
s
s
d
if
f
er
en
t
d
atasets
,
en
v
ir
o
n
m
en
tal
co
n
d
itio
n
s
an
d
r
ea
l
-
wo
r
ld
s
p
ea
k
i
n
g
co
n
d
itio
n
s
.
T
h
e
au
th
o
r
m
ai
n
ly
c
o
n
ce
n
t
r
ated
o
n
th
ese
lim
itatio
n
s
an
d
th
e
p
r
esen
t
wo
r
k
aim
s
t
o
ad
d
r
ess
th
ese
is
s
u
es b
y
in
teg
r
a
tin
g
ch
an
n
el
an
d
s
p
atial
atten
ti
o
n
m
ec
h
a
n
is
m
s
.
Stu
tter
in
g
d
etec
tio
n
alg
o
r
ith
m
s
ar
e
b
r
o
ad
l
y
class
if
ied
in
t
o
m
ac
h
i
n
e
lear
n
in
g
,
d
ee
p
le
ar
n
in
g
,
an
d
h
y
b
r
id
f
r
am
ew
o
r
k
s
.
T
h
e
tr
a
d
itio
n
al
m
ac
h
in
e
lear
n
in
g
alg
o
r
ith
m
s
m
ain
ly
d
ep
e
n
d
s
o
n
th
e
m
an
u
al
f
ea
tu
r
e
ex
tr
ac
tio
n
,
an
d
wh
ich
h
as
lim
ited
ad
ap
tab
ilit
y
to
s
p
ee
ch
v
ar
iab
ilit
y
.
Dee
p
lear
n
in
g
an
d
h
y
b
r
id
f
r
am
ewo
r
k
s
s
h
o
w
m
o
r
e
ac
cu
r
ac
y
th
an
th
e
tr
ad
itio
n
al
m
eth
o
d
.
T
h
ese
ar
e
th
e
m
ain
ad
v
a
n
tag
es
th
at
m
ak
e
th
e
atten
tio
n
-
b
ased
d
ee
p
o
r
h
y
b
r
id
f
r
am
ew
o
r
k
s
m
o
r
e
r
o
b
u
s
t
an
d
s
u
itab
le
tech
n
iq
u
e
f
o
r
s
tu
tter
d
etec
ti
o
n
ac
r
o
s
s
d
if
f
er
e
n
t
d
ata
s
ets an
d
d
iv
er
s
e
s
p
ee
ch
c
o
n
d
itio
n
s
.
3.
P
RO
P
O
SE
D
M
O
D
E
L
T
h
is
s
ec
tio
n
p
r
esen
ts
th
e
p
r
o
p
o
s
ed
d
ee
p
lear
n
in
g
-
b
ased
m
o
d
el
f
o
r
s
tu
tter
e
v
en
t
d
ete
ctio
n
.
T
h
e
co
m
p
lete
ar
ch
itectu
r
e
is
d
iv
id
ed
in
to
f
o
llo
win
g
s
tep
s
:
a.
Data
co
llectio
n
an
d
d
ata
lo
ad
in
g
:
T
h
is
s
tag
e
we
co
n
s
id
er
ed
m
u
lti
-
d
o
m
ain
s
p
ee
ch
d
ata
f
o
r
f
u
r
th
er
p
r
o
ce
s
s
in
g
.
I
n
o
u
r
w
o
r
k
we
c
o
n
s
id
er
ed
Flu
en
c
y
B
an
k
,
SEP
-
2
8
k
an
d
UC
lass
(
R
elea
s
e
1
&
2
)
d
ataset.
T
h
e
o
v
er
all
d
ataset
s
ize
is
4
0
h
r
s
.
c
o
m
b
in
in
g
u
n
la
b
eled
an
d
lab
ell
ed
d
ata.
b.
D
a
t
a
a
u
g
m
e
n
ta
t
i
o
n
:
T
h
e
d
a
t
ase
t
w
h
i
c
h
w
e
a
r
e
c
o
n
s
i
d
e
r
i
n
g
h
a
s
li
m
i
t
e
d
s
a
m
p
le
s
w
it
h
l
a
b
el
s
t
h
e
r
e
f
o
r
e
,
we
i
n
c
o
r
p
o
r
a
t
e
d
a
t
a
a
u
g
m
e
n
t
a
t
i
o
n
s
t
e
p
s
t
o
i
n
c
r
e
a
s
e
t
h
e
a
v
a
il
ab
l
e
d
a
t
a
f
o
r
t
r
a
i
n
i
n
g
.
w
h
i
c
h
a
p
p
l
i
e
s
t
e
m
p
o
r
a
l
a
u
g
m
e
n
t
a
t
i
o
n
s
u
c
h
as
r
a
n
d
o
m
p
i
t
c
h
s
h
i
f
t
,
r
e
v
e
r
b
e
r
a
t
i
o
n
a
n
d
wh
i
t
e
n
o
i
s
e
.
c.
F
e
a
t
u
r
e
e
x
t
r
a
c
t
i
o
n
:
T
h
e
a
u
g
m
e
n
t
e
d
d
a
t
a
i
s
p
r
o
c
ess
e
d
t
h
r
o
u
g
h
“
f
e
a
t
u
r
e
e
x
t
r
a
ct
i
o
n
”
p
h
a
s
e
w
h
e
r
e
w
e
a
p
p
l
y
C
N
N
b
as
e
m
o
d
e
l
t
o
e
x
t
r
a
c
t
th
e
r
o
b
u
s
t
f
e
a
t
u
r
e
s
s
u
c
h
a
s
s
p
e
c
t
r
a
l
f
e
at
u
r
e
s
,
p
i
tc
h
,
t
e
m
p
o
r
a
l
f
e
a
t
u
r
e
s
a
n
d
c
o
n
t
e
x
t
u
a
l
f
e
a
t
u
r
es
.
t
h
e
d
e
ta
i
l
ed
f
e
a
t
u
r
e
e
x
t
r
a
c
t
i
o
n
al
g
o
r
i
t
h
m
s
d
i
s
c
u
s
s
e
d
i
n
n
e
x
t
s
e
c
ti
o
n
.
d.
Featu
r
e
r
ef
in
em
en
t:
Fo
r
f
ea
tu
r
e
r
ef
in
em
en
t
s
tag
e,
ea
r
lier
ex
tr
ac
ted
f
ea
tu
r
es
ar
e
g
iv
en
a
s
in
p
u
t,
in
th
is
p
ar
ticu
lar
s
tag
e
we
h
av
e
in
c
o
r
p
o
r
ated
atten
tio
n
m
ec
h
a
n
is
m
wh
ich
h
elp
s
to
f
o
c
u
s
o
n
im
p
o
r
tan
t
attr
ib
u
tes
with
r
ed
u
ce
d
c
o
m
p
u
tatio
n
al
c
o
m
p
lex
ity
.
T
h
e
f
ea
tu
r
e
ex
tr
ac
t
io
n
p
h
ase
also
in
clu
d
es
s
h
o
r
t
t
er
m
en
er
g
y
an
d
co
r
r
elatio
n
f
ac
t
o
r
attr
ib
u
tes.
e.
C
o
n
ca
ten
atio
n
an
d
m
u
lti
-
class
class
if
icatio
n
:
On
ce
th
e
f
ea
tu
r
es
ar
e
ex
tr
ac
ted
,
we
co
n
ca
te
n
a
te
all
attr
ib
u
tes
an
d
ap
p
ly
C
NN
b
ased
m
u
lti
class
if
ier
m
o
d
el.
Fu
r
th
er
th
e
o
u
tco
m
e
o
f
class
if
icatio
n
m
o
d
u
les
is
u
s
ed
to
p
er
f
o
r
m
th
e
class
if
icatio
n
an
al
y
s
is
.
Fig
u
r
e
1
d
ep
icts
th
e
d
etailed
a
r
ch
itectu
r
e
o
f
p
r
o
p
o
s
ed
s
tu
tter
d
etec
tio
n
m
o
d
el.
T
h
e
e
x
p
er
i
m
en
tal
d
ata
u
s
ed
in
th
is
wo
r
k
is
o
b
tain
ed
f
r
o
m
p
u
b
licly
av
ailab
le
r
eso
u
r
ce
s
an
d
f
u
r
th
er
d
etails
ab
o
u
t
th
e
d
a
tab
ase
d
escr
ib
ed
in
s
ec
tio
n
4
.
2
.
Fig
u
r
e
1
.
Sp
ee
c
h
f
ea
tu
r
e
ex
tr
a
ctio
n
an
d
s
p
atial
an
d
ch
an
n
el
atten
tio
n
-
b
ased
m
u
lticlas
s
clas
s
if
icatio
n
b
lo
ck
d
iag
r
am
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
1
6
,
No
.
4
,
Au
g
u
s
t
20
2
6
:
2
1
0
6
-
2
119
2110
3
.
1
.
Da
t
a
a
ug
m
ent
a
t
io
n
As
d
is
cu
s
s
ed
b
ef
o
r
e
th
at
we
h
av
e
lim
ited
lab
elled
d
ataset
th
er
ef
o
r
e
d
ata
au
g
m
en
tatio
n
p
r
o
ce
s
s
is
in
co
r
p
o
r
ated
in
th
is
wo
r
k
.
Data
au
g
m
en
tatio
n
is
a
wid
el
y
u
tili
ze
d
m
eth
o
d
aim
ed
at
en
r
ich
in
g
b
o
th
th
e
q
u
an
tity
an
d
d
iv
er
s
ity
o
f
a
n
n
o
tated
tr
ain
in
g
d
ata
,
th
er
eb
y
e
n
h
an
cin
g
th
e
r
o
b
u
s
tn
ess
o
f
d
e
ep
n
eu
r
al
n
etwo
r
k
s
(
DNNs)
wh
ile
m
itig
atin
g
o
v
e
r
f
itti
n
g
.
I
n
th
e
r
ea
lm
o
f
n
o
r
m
al
s
p
ee
ch
r
ec
o
g
n
itio
n
,
d
ata
au
g
m
e
n
tatio
n
h
as
p
r
o
v
e
n
to
b
e
p
ar
ticu
lar
l
y
ef
f
ec
tiv
e
in
ad
d
r
ess
in
g
d
ata
s
ca
r
city
an
d
im
p
r
o
v
in
g
t
h
e
p
er
f
o
r
m
an
ce
o
f
v
a
r
io
u
s
DNN
ac
o
u
s
tic
tech
n
iq
u
es
[
1
]
,
[
4
]
,
[
1
6
]
.
Nu
m
e
r
o
u
s
au
g
m
en
tatio
n
tech
n
iq
u
es
h
av
e
b
ee
n
e
x
p
lo
r
ed
,
in
clu
d
in
g
p
itch
ad
ju
s
tm
en
t,
s
p
ec
tr
al
d
i
s
to
r
tio
n
,
tem
p
o
p
er
t
u
r
b
atio
n
,
s
p
ee
d
p
er
t
u
r
b
atio
n
,
cr
o
s
s
-
d
o
m
ain
ad
a
p
tatio
n
,
ad
d
itio
n
o
f
n
o
is
e
to
clea
n
s
p
ee
ch
,
s
p
ec
tr
o
g
r
am
m
an
ip
u
latio
n
in
v
o
lv
in
g
f
r
eq
u
e
n
cy
an
d
t
im
e
m
ask
in
g
,
m
ix
-
s
p
ee
ch
,
am
o
n
g
o
th
e
r
s
[
1
]
,
[
4
]
,
[
7
]
,
[
1
7
]
.
Ho
wev
er
,
t
h
er
e
h
as
b
ee
n
n
o
tab
ly
s
ca
n
t
atten
tio
n
d
ir
ec
ted
to
war
d
s
d
ata
au
g
m
e
n
tatio
n
tailo
r
ed
s
p
ec
if
ically
f
o
r
th
e
s
p
ee
c
h
d
is
o
r
d
er
d
o
m
ain
[
5
]
,
[
6
]
,
[
9
]
.
T
h
er
ef
o
r
e,
we
co
n
s
id
er
4
d
if
f
er
en
t ty
p
es o
f
s
am
p
les to
a
u
g
m
en
t th
e
d
ata
wh
ich
a
r
e:
a.
Mu
s
ic:
T
h
e
in
itial
clea
n
s
tu
tter
in
g
v
o
ice
s
am
p
le
is
co
m
b
in
ed
with
o
n
e
r
an
d
o
m
m
u
s
i
c
s
am
p
le
f
r
o
m
MU
SAN.
(
s
ig
n
al
-
to
-
n
o
is
e
r
atio
(
SNR
)
: 5
-
15
d
B
)
.
b.
No
is
e:
Sam
p
les
o
f
MU
SAN
s
o
u
n
d
s
ar
e
p
lace
d
in
to
th
e
s
tu
tter
in
g
s
p
ee
c
h
at
1
-
s
ec
o
n
d
i
n
ter
v
als.
(
SNR
:
0
-
15
d
B
)
.
c.
B
ab
b
le:
T
h
e
in
itial
clea
n
s
tu
tt
er
ed
s
p
ee
ch
s
am
p
le
is
s
u
p
p
le
m
en
ted
b
y
s
p
ee
ch
s
am
p
les
f
r
o
m
th
r
ee
t
o
s
ev
en
r
an
d
o
m
l
y
ch
o
s
en
s
p
ea
k
er
s
.
(
S
NR
: 1
3
-
20
d
B
)
.
d.
R
ev
er
b
: T
h
e
“c
lean
”
tr
ai
n
in
g
s
et
u
n
d
er
g
o
es c
o
n
v
o
lu
tio
n
with
s
im
u
lated
r
o
o
m
im
p
u
ls
e
r
esp
o
n
s
es.
T
h
is
m
o
d
el
u
s
es
wav
Au
g
m
en
t
an
d
to
r
ch
au
d
io
lib
r
ar
ies
to
i
n
co
r
p
o
r
ate
th
ese
a
u
g
m
en
tatio
n
s
.
L
et
u
s
co
n
s
id
er
th
at
r
ep
r
esen
t
th
e
as
th
e
o
r
ig
in
al
clea
n
s
tu
tter
in
g
s
p
ee
ch
s
am
p
le,
as
th
e
s
in
g
le
m
u
s
ic
s
am
p
le
r
an
d
o
m
l
y
s
elec
ted
f
r
o
m
MU
S
AN,
as
th
e
n
o
is
e
s
am
p
le
r
an
d
o
m
ly
s
elec
ted
f
r
o
m
MU
SAN
n
o
is
es
at
th
e
ℎ
s
ec
o
n
d
in
ter
v
al,
as
th
e
co
m
b
in
ed
s
p
ee
ch
s
am
p
les
f
r
o
m
r
an
d
o
m
ly
s
elec
ted
3
-
7
s
p
ea
k
er
s
,
an
d
as
th
e
clea
n
tr
ain
in
g
s
et
co
n
v
o
lv
e
d
with
s
i
m
u
lated
r
o
o
m
im
p
u
ls
e
r
esp
o
n
s
es.
B
ased
o
n
th
ese
s
ig
n
als
th
e
au
g
m
e
n
tatio
n
ca
n
b
e
ex
p
r
ess
ed
as:
a.
Mu
s
ic
au
g
m
en
tatio
n
=
+
b.
No
is
e
au
g
m
en
tatio
n
=
+
∑
=
1
wh
er
e
is
th
e
to
tal
n
u
m
b
e
r
o
f
1
-
s
ec
o
n
d
in
ter
v
al
c.
B
ab
b
le
au
g
m
en
tatio
n
=
+
d.
R
ev
er
b
au
g
m
e
n
tatio
n
=
3
.
2
.
F
e
a
t
ure
ex
t
r
a
ct
io
n
T
h
e
f
ea
tu
r
e
ex
tr
ac
tio
n
is
a
m
ain
s
tep
in
an
y
s
p
ee
ch
an
aly
s
i
s
.
T
h
er
ef
o
r
e,
in
o
u
r
wo
r
k
we
u
s
ed
m
o
s
t
ef
f
icien
t
C
NN
b
ased
DL
alg
o
r
ith
m
f
o
r
f
ea
tu
r
e
ex
tr
ac
tio
n
p
r
o
ce
s
s
to
d
etec
t
th
e
s
tu
tter
ev
en
ts
[
7
]
,
[
1
4
]
,
[
1
9
]
,
[
2
0
]
.
W
e
u
s
ed
s
tu
tter
ev
en
ts
with
an
n
o
tatio
n
,
an
d
th
is
tem
p
o
r
al
an
n
o
tatio
n
s
p
ec
if
ies
s
p
e
cif
ic
ty
p
e
o
f
s
tu
tter
ev
en
t
in
th
e
s
p
ee
ch
s
ig
n
al.
A
d
d
itio
n
ally
,
th
ese
an
n
o
tatio
n
s
g
iv
e
d
etail
in
f
o
r
m
atio
n
o
f
th
e
s
p
ec
if
ic
s
eg
m
en
t,
alo
n
g
with
s
tar
tin
g
p
o
in
t o
f
t
h
e
ev
en
t
an
d
en
d
i
n
g
p
o
in
t
o
f
th
e
ev
en
t
with
in
s
p
ee
ch
s
ig
n
al
[
1
8
]
,
[
2
1
]
.
T
y
p
ically
,
in
SED
m
o
d
els,
th
e
d
u
r
atio
n
o
f
t
h
e
tem
p
o
r
al
r
eg
io
n
r
a
n
g
es
b
etwe
en
th
r
ee
an
d
f
iv
e
s
ec
o
n
d
s
.
Fix
ed
-
s
ized
r
eg
io
n
s
u
p
er
v
is
ed
lear
n
i
n
g
m
o
d
els
ar
e
well
-
s
u
ited
f
o
r
th
e
SED
task
[
7
]
,
[
1
7
]
.
I
n
wh
ich
th
e
ac
o
u
s
tic
r
eg
io
n
s
an
d
an
n
o
tatio
n
s
s
er
v
e
as tr
ain
i
n
g
d
ata
f
o
r
th
e
m
o
d
el.
T
h
e
o
b
tain
ed
s
p
ee
ch
s
ig
n
al
is
th
en
p
r
o
ce
s
s
ed
th
r
o
u
g
h
t
h
e
f
ea
tu
r
e
ex
tr
ac
tio
n
m
o
d
u
le
wh
er
e
we
ex
tr
ac
t
s
p
ec
tr
al,
p
itch
tem
p
o
r
al
an
d
co
n
tex
t
u
al
f
ea
tu
r
es
o
f
s
p
ee
ch
s
ig
n
al
[
2
2
]
,
[
2
3
]
.
3
.
3
.
Sp
ec
t
ra
l f
e
a
t
ure
ex
t
ra
ct
io
n
T
h
e
ac
o
u
s
tic
en
er
g
y
ca
n
b
e
d
e
tecte
d
with
th
e
h
elp
o
f
s
p
ec
tr
a
l
r
ep
r
esen
tatio
n
.
I
n
t
h
is
wo
r
k
,
we
u
tili
ze
th
e
MFC
C
f
ea
tu
r
e
ex
tr
ac
tio
n
m
o
d
el.
I
t
co
n
v
er
ts
th
e
au
d
i
o
s
ig
n
al
in
to
co
ef
f
icien
ts
to
ca
p
tu
r
e
th
e
s
p
ec
tr
al
ch
ar
ac
ter
is
tics
o
n
a
q
u
asi
-
lo
g
ar
ith
m
ic
f
r
eq
u
e
n
cy
s
ca
le
[
2
2
]
,
[
2
3
]
.
T
h
e
p
r
o
ce
s
s
ed
s
ig
n
al
is
tr
an
s
f
o
r
m
ed
in
to
th
e
f
r
eq
u
e
n
cy
d
o
m
ai
n
u
tili
zin
g
th
e
s
h
o
r
t
-
tim
e
Fo
u
r
ier
tr
an
s
f
o
r
m
(
STFT
)
,
em
p
lo
y
in
g
a
h
o
p
len
g
th
o
f
(
0
.
0
1
0
×
)
an
d
(
0
.
0
2
5
×
)
.
Fo
llo
win
g
th
i
s
,
a
s
et
o
f
6
4
f
ilter
b
an
k
s
is
em
p
lo
y
ed
o
n
t
h
e
f
r
e
q
u
en
c
y
d
o
m
a
in
s
ig
n
al
to
d
er
i
v
e
Me
l
-
f
r
eq
u
e
n
cy
ce
p
s
tr
al
co
ef
f
icien
ts
(
MFC
C
)
v
ec
to
r
r
e
p
r
esen
tin
g
ea
ch
ac
o
u
s
tic
p
o
r
tio
n
[
1
7
]
,
[
2
2
]
,
[
2
4
]
,
[
2
5
]
.
W
h
ile
th
e
s
p
ec
tr
al
f
ea
tu
r
e
r
ep
r
esen
tatio
n
is
cr
u
cial
f
o
r
ca
p
tu
r
in
g
f
r
eq
u
e
n
cy
in
f
o
r
m
atio
n
,
it'
s
eq
u
ally
im
p
er
ativ
e
to
co
n
s
id
er
p
itch
an
d
tim
e
in
f
o
r
m
atio
n
f
o
r
s
tu
tter
d
etec
tio
n
[
2
3
]
.
B
elo
w
g
iv
en
al
g
o
r
ith
m
d
em
o
n
s
tr
ates th
e
MFC
C
co
ef
f
icien
t c
o
m
p
u
tatio
n
.
Alg
o
r
ith
m
1
.
MFC
C
co
ef
f
icien
t e
x
tr
ac
tio
n
p
r
o
ce
s
s
Input: 3 sec down sample speech signals transformed
into
time domain
Output: MFCC coefficients
Step 1:
(
)
←
[
0
:
3
×
8000
)
]
Step 2: apply signal to Nosie estimator
←
0
.
95
and
(
)
←
(
)
−
(
−
1
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
S
p
a
tia
l a
n
d
ch
a
n
n
el
a
tten
tio
n
mec
h
a
n
is
m
fo
r
s
p
ee
ch
d
is
flu
en
cy
d
etec
tio
n
…
(
K
u
s
u
ma
H
.
R
.
)
2111
Step 3:
fr
a
me
th
e
o
bt
ai
n
e
d
s
ig
n
al
(
)
in
to
fr
a
me
s
o
f
si
z
e
2
5
ms
wi
th
s
t
ri
de
of
1
0
m
s
Step 4: compute the Hamming window
(
)
(
)
←
0
.
54
−
0
.
46
c
o
s
(
2
−
1
)
Step 5: estimate the power spectrum for the converted frame
←
|
(
)
|
2
Step 6:
transf
orm t
he pow
e
r spectr
um int
o Mel
-
s
c
a
l
e
w
i
t
h
t
h
e
h
e
l
p
o
f
6
4
t
r
i
a
ng
u
l
a
r
f
i
lt
e
r
s
:
←
2595
lo
g
10
(
1
+
100
)
Step 7: Apply DCT on the filter banks to generate the MFCC coefficients
←
(
)
Step 8: return
m
f
c
c
3
.
4
.
P
i
t
ch,
t
em
po
r
a
l a
nd
co
nte
x
t
ua
l f
ea
t
ures
T
h
is
s
ec
tio
n
d
escr
ib
es
th
e
p
r
o
ce
s
s
o
f
p
itch
an
d
tem
p
o
r
al
f
ea
tu
r
e
ex
tr
ac
tio
n
f
o
r
s
tu
tter
ev
en
t
d
etec
tio
n
.
T
h
ese
f
ea
t
u
r
es
h
el
p
to
id
en
tif
y
th
e
u
n
v
o
iced
p
e
r
io
d
s
an
d
s
u
d
d
en
v
ar
iatio
n
in
s
tu
tter
in
g
s
p
ee
ch
.
E
s
tim
atin
g
th
e
u
n
v
o
iced
p
o
r
ti
o
n
s
with
in
s
tu
tter
in
g
s
p
ee
c
h
c
o
u
ld
p
o
ten
tially
im
p
r
o
v
e
t
h
e
r
ec
o
g
n
itio
n
o
f
b
l
o
ck
s
an
d
p
r
o
lo
n
g
atio
n
ev
e
n
ts
.
T
h
is
is
b
ec
au
s
e
v
o
iced
s
p
ee
c
h
ten
d
s
to
ex
h
ib
it
s
m
o
o
th
er
ch
ar
ac
te
r
is
tics
co
m
p
ar
ed
to
u
n
v
o
ice
d
s
p
ee
ch
[
1
7
]
,
[
2
2
]
,
[
2
3
]
.
W
e
h
av
e
co
n
s
id
er
ed
ze
r
o
c
r
o
s
s
in
g
r
ate
(
Z
C
R
)
[
2
6
]
an
d
s
p
ec
tr
al
f
lu
x
o
n
s
et
to
o
b
tain
th
e
p
itch
an
d
tem
p
o
r
al
f
ea
tu
r
es.
T
h
e
Z
C
R
p
r
o
ce
s
s
is
d
escr
ib
ed
in
alg
o
r
ith
m
2
.
Alg
o
r
ith
m
2
.
Z
er
o
c
r
o
s
s
in
g
f
e
atu
r
e
ex
tr
ac
tio
n
Output: estimated value of zero crossing rate
St
ep
1:
in
it
ia
li
ze
th
e
ex
pe
ri
me
nt
al
pa
ra
me
te
rs
su
ch
as
sa
mp
le
ra
te
←
8000
,
wi
nd
ow
si
ze
←
(
×
0
.
025
)
, overlap size
←
(
×
0
.
1
)
, total frames
←
ℎ
−
Step 2: Initialize the
←
[
.
]
Step 3: iterate all frame to get Z
For
=
0
to total frame
-
1
do
←
×
(
−
)
←
m
i
n
(
+
,
ℎ
)
Frame
←
[
:
]
Compute
(
)
End for
Return
I
n
n
ex
t
s
tag
e,
we
co
m
p
u
te
s
p
e
ctr
al
f
lu
x
o
n
s
et
b
y
c
o
m
p
ar
i
n
g
th
e
p
o
wer
s
p
ec
tr
u
m
o
f
two
co
n
s
ec
u
tiv
e
f
r
am
es
an
d
co
m
p
ar
e
with
th
e
s
u
m
o
f
all
p
o
s
itiv
e
d
ev
iatio
n
wh
ich
g
en
er
ates
th
e
v
ec
to
r
o
f
o
n
s
et
s
tr
en
g
th
is
p
r
esen
ted
in
(
1
)
:
(
)
=
∑
(
|
(
,
)
|
−
|
(
−
1
,
)
|
)
=
2
=
1
(
1
)
w
h
er
e
is
th
e
tim
e
in
d
ex
,
r
ep
r
esen
ts
f
r
am
e
n
u
m
b
er
an
d
k
is
th
e
d
is
cr
ete
f
r
eq
u
en
cy
in
d
ex
.
Fin
ally
,
we
ap
p
ly
co
n
tex
tu
al
f
ea
tu
r
e
e
x
tr
ac
tio
n
m
o
d
el
w
h
er
e
we
h
av
e
u
s
ed
W
av
2
Vec
2
.
0
m
o
d
el
[
4
]
,
[
1
6
]
.
W
av
2
Vec
2
.
0
is
co
m
p
r
is
ed
o
f
two
s
elf
-
s
u
p
er
v
is
ed
lear
n
in
g
o
b
jectiv
es:
a
co
n
tr
asti
v
e
task
an
d
a
q
u
an
ti
za
tio
n
task
.
I
n
th
e
co
n
tr
asti
v
e
task
,
th
e
m
o
d
el
u
n
d
er
g
o
es
tr
ain
in
g
t
o
en
h
an
ce
t
h
e
s
im
ilar
ity
am
o
n
g
co
r
r
elate
d
in
s
tan
ce
s
(
r
ef
er
r
ed
to
as
p
o
s
itiv
e
s
am
p
les)
wh
il
e
r
ed
u
cin
g
th
e
s
im
ilar
ity
am
o
n
g
u
n
r
elate
d
in
s
tan
ce
s
(
r
ef
e
r
r
ed
to
as
n
eg
ativ
e
s
am
p
les)
ac
r
o
s
s
d
if
f
er
en
t m
ask
ed
tim
e
s
tep
s
[
4
]
.
T
h
is
task
is
co
m
m
o
n
ly
e
x
ec
u
ted
u
tili
zin
g
a
co
n
tr
asti
v
e
lo
s
s
f
u
n
ctio
n
,
w
h
ich
ev
alu
ates
th
e
s
im
ilar
ity
b
etwe
en
p
air
s
o
f
i
n
s
tan
ce
s
an
d
m
o
d
if
ies
th
e
m
o
d
el
p
ar
am
eter
s
to
m
in
im
ize
th
e
lo
s
s
.
B
y
en
g
ag
i
n
g
in
co
n
tr
asti
v
e
task
s
,
th
e
m
o
d
el
is
eq
u
ip
p
e
d
to
h
a
n
d
le
v
ar
io
u
s
d
o
wn
s
tr
ea
m
task
s
,
in
clu
d
in
g
th
e
d
etec
tio
n
o
f
s
tu
tter
in
g
ev
e
n
ts
.
3.
5
.
F
e
a
t
ure
re
f
inem
ent
T
h
e
o
b
tain
ed
f
ea
tu
r
e
m
a
p
is
p
r
o
ce
s
s
ed
th
r
o
u
g
h
th
e
atten
ti
o
n
m
ec
h
a
n
is
m
wh
ich
m
ain
ly
f
o
cu
s
o
n
im
p
r
o
v
in
g
th
e
f
ea
tu
r
e
m
ap
an
d
f
o
cu
s
o
n
im
p
o
r
ta
n
t
attr
ib
u
tes
to
d
etec
t
th
e
s
tu
tter
ev
en
t
[
7
]
,
[
2
0
]
,
[
2
7
]
.
T
h
e
atten
tio
n
m
ec
h
an
is
m
h
as
two
m
ain
co
m
p
o
n
en
ts
wh
ich
ar
e
k
n
o
wn
as
ch
an
n
el
atten
tio
n
a
n
d
s
p
atial
atten
tio
n
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
1
6
,
No
.
4
,
Au
g
u
s
t
20
2
6
:
2
1
0
6
-
2
119
2112
C
h
an
n
el
atten
tio
n
en
h
an
ce
s
cr
itical
f
ea
tu
r
es
with
in
a
f
ea
tu
r
e
m
ap
b
y
ass
ig
n
in
g
weig
h
ts
to
ea
ch
ch
an
n
el
b
ased
o
n
th
eir
s
ig
n
if
ican
ce
.
Mo
r
e
o
v
er
,
s
p
atial
atten
tio
n
ac
ce
n
tu
ates
s
ig
n
if
ican
t
r
eg
io
n
s
wh
ile
d
am
p
en
in
g
ir
r
elev
an
t
o
n
es
th
r
o
u
g
h
t
h
e
u
tili
za
tio
n
o
f
a
co
n
v
o
lu
tio
n
al
la
y
er
an
d
a
s
k
ip
p
in
g
co
n
n
ec
tio
n
.
I
n
th
e
cu
r
r
en
t
s
tu
d
y
,
c
h
an
n
el
an
d
s
p
atial
atten
tio
n
alg
o
r
ith
m
s
ar
e
co
m
b
in
ed
t
o
en
h
an
ce
th
e
s
p
ee
ch
r
ec
o
g
n
itio
n
s
y
s
tem
s
in
s
tu
tter
s
p
ee
ch
.
B
asically
,
s
p
ee
ch
s
ig
n
als
ar
e
m
o
r
e
co
m
p
lex
a
n
d
d
y
n
am
ic
.
T
o
ca
tch
t
h
is
d
y
n
am
ic
f
ea
t
u
r
e
ch
a
n
g
e
ch
an
n
el
atten
tio
n
is
u
s
ed
b
y
em
p
h
asizin
g
ap
p
r
o
p
r
iate
ch
a
n
n
els,
wh
i
le
s
p
atial
atten
tio
n
d
ea
ls
with
th
e
tem
p
o
r
al
an
d
s
p
ec
tr
al
ch
an
g
es
b
y
f
o
cu
s
in
g
o
n
s
p
ec
if
ic
r
eg
io
n
s
.
Me
r
g
in
g
b
o
th
m
ec
h
a
n
is
m
s
g
iv
es
u
s
a
m
o
s
t
r
o
b
u
s
t
ap
p
r
o
ac
h
to
s
p
ee
ch
d
is
f
lu
en
cy
d
etec
tio
n
.
C
h
an
n
el
atten
tio
n
h
as
th
e
ab
i
lity
to
s
ep
ar
ate
r
elev
a
n
t
f
ea
t
u
r
es
f
r
o
m
o
v
e
r
lap
p
e
d
s
ig
n
als,
wh
ile
s
p
atial
a
tten
tio
n
m
ain
ly
f
o
c
u
s
es
o
n
p
ar
ticu
la
r
r
eg
io
n
s
wh
ich
in
d
icate
d
is
f
l
u
en
cies.
T
h
e
r
esu
lts
f
r
o
m
b
o
t
h
th
e
n
etwo
r
k
s
ar
e
co
m
b
in
ed
an
d
in
p
u
t
to
th
e
s
o
lid
lay
er
with
th
e
h
elp
o
f
s
ig
m
o
id
ac
tiv
atio
n
f
u
n
ctio
n
.
W
h
ich
g
iv
es u
s
p
r
o
b
ab
ilit
y
f
o
r
ea
c
h
d
is
f
lu
en
t c
lass
,
s
o
we
ca
n
ex
p
r
ess
in
th
e
(
2
)
.
(
)
=
(
∗
)
(
,
)
=
∑
∑
(
,
)
(
−
,
−
)
(
2
)
T
h
e
ch
an
n
el
atten
tio
n
m
ec
h
a
n
is
m
p
r
o
d
u
ce
s
f
ea
tu
r
e
m
ap
s
b
y
ap
p
ly
in
g
th
e
m
ax
im
u
m
a
n
d
av
er
ag
e
p
o
o
lin
g
o
n
th
e
in
p
u
t
f
ea
tu
r
e
m
ap
o
b
tain
ed
f
r
o
m
p
r
ev
io
u
s
s
tep
s
.
Fu
r
th
er
,
th
ese
f
ea
tu
r
e
m
ap
s
ar
e
p
ass
ed
th
r
o
u
g
h
th
e
m
u
lti
-
lay
er
p
er
ce
p
tr
o
n
(
ML
P)
n
etwo
r
k
an
d
s
u
m
o
f
o
u
tp
u
t
is
p
r
o
ce
s
s
ed
th
r
o
u
g
h
th
e
ac
tiv
atio
n
f
u
n
ctio
n
to
g
e
n
er
ate
th
e
atten
t
io
n
v
alu
es.
Atten
tio
n
c
h
an
n
el
weig
h
t
an
d
in
p
u
t
f
ea
tu
r
es
ar
e
m
u
ltip
lied
to
f
in
d
th
e
f
in
al
ch
an
n
el
atten
tio
n
.
T
h
is
is
ex
p
r
ess
ed
in
th
e
(
3
)
:
(
)
=
(
(
(
)
)
)
+
(
(
)
)
=
(
1
(
0
(
)
)
+
1
(
0
(
)
)
(
3
)
Similar
ly
,
th
e
s
p
atial
a
tten
tio
n
m
ec
h
an
is
m
ex
tr
ac
ts
th
e
s
p
a
tial
attr
ib
u
tes
b
y
ap
p
ly
in
g
av
er
ag
in
g
an
d
m
ax
im
u
m
p
o
o
lin
g
o
n
t
h
e
o
b
t
ain
ed
f
ea
tu
r
e
m
a
p
in
s
p
atial
d
im
en
s
io
n
.
T
h
is
s
tep
g
e
n
er
ate
s
two
f
ea
tu
r
e
m
ap
s
wh
ich
ar
e
co
n
ca
ten
ated
alo
n
g
with
ch
an
n
el
d
im
en
s
io
n
.
L
at
er
,
th
is
co
n
ca
ten
ate
d
f
ea
tu
r
e
i
s
p
r
o
ce
s
s
ed
th
r
o
u
g
h
th
e
co
n
v
o
lu
tio
n
an
d
s
ig
m
o
id
ac
tiv
atio
n
f
u
n
ctio
n
.
T
h
e
o
b
tai
n
ed
f
ea
tu
r
e
is
m
u
ltip
lied
with
th
e
o
r
i
g
in
al
f
ea
tu
r
e
m
ap
to
g
e
n
er
ate
th
e
r
ef
in
ed
f
e
atu
r
es.
T
h
is
is
ex
p
r
ess
ed
in
(
4
)
.
(
)
=
(
7
∗
7
(
[
(
)
;
(
)
]
)
)
=
(
7
∗
7
(
;
)
(
4
)
Her
e,
d
en
o
tes th
e
s
ig
m
o
id
ac
t
iv
atio
n
f
u
n
ctio
n
.
Fig
u
r
e
2
d
ep
i
cts th
e
ar
ch
itectu
r
e
o
f
atten
tio
n
m
ec
h
an
is
m
.
Fig
u
r
e
2
.
T
h
e
co
n
v
o
lu
tio
n
al
b
l
o
ck
atten
tio
n
m
o
d
u
le
(
C
B
AM
)
illu
s
tr
a
tin
g
s
eq
u
en
tial a
n
d
ch
an
n
el
atten
tio
n
3.
6
.
M
ulti
-
cla
s
s
c
la
s
s
if
ica
t
io
n m
o
del
T
h
i
s
p
a
r
t
i
l
l
u
s
t
r
a
te
s
t
h
e
s
t
a
g
e
s
in
C
N
N
b
a
s
e
d
m
u
lt
i
-
c
la
s
s
c
la
s
s
i
f
i
c
a
ti
o
n
m
o
d
e
l
.
S
p
e
e
c
h
s
i
g
n
a
l
i
s
a
p
p
li
e
d
a
s
a
n
i
n
p
u
t
l
a
y
e
r
i
n
t
h
e
f
i
r
s
t
s
t
a
g
e
.
S
e
c
o
n
d
s
t
a
g
e
u
s
e
s
c
o
n
v
o
lu
t
i
o
n
a
l
l
a
y
e
r
t
o
e
x
t
r
a
c
t
t
h
e
f
e
a
tu
r
e
s
f
r
o
m
t
h
e
i
n
p
u
t
d
a
t
a
.
I
n
t
h
is
o
p
e
r
a
t
i
o
n
,
we
a
ls
o
i
n
c
l
u
d
e
,
R
e
L
U
la
y
e
r
t
o
i
n
t
r
o
d
u
c
e
n
o
n
-
l
i
n
e
a
r
i
t
y
a
f
te
r
e
a
c
h
c
o
n
v
o
l
u
t
i
o
n
o
p
e
r
a
t
i
o
n
[
2
6
]
.
T
h
e
e
x
t
r
a
c
t
e
d
f
e
a
t
u
r
es
a
r
e
p
a
s
s
e
d
t
h
r
o
u
g
h
t
h
e
M
a
x
P
o
o
li
n
g
l
a
y
e
r
i
ts
m
ai
n
c
o
n
c
e
r
n
i
s
d
i
m
e
n
s
i
o
n
s
t
o
r
e
d
u
ce
t
h
e
f
e
a
t
u
r
e
m
a
p
s
a
n
d
r
e
t
a
i
n
t
h
e
m
o
s
t
i
m
p
o
r
t
a
n
t
f
e
a
t
u
r
es
.
T
h
e
s
a
m
e
p
r
o
c
e
s
s
is
r
e
p
e
a
t
e
d
t
w
o
t
im
e
s
b
e
f
o
r
e
a
p
p
l
y
i
n
g
t
o
t
h
e
“
f
u
l
l
y
c
o
n
n
e
c
t
e
d
l
a
y
e
r
”
o
p
e
r
a
t
i
o
n
s
.
T
h
e
“
f
u
l
l
y
c
o
n
n
ec
t
e
d
l
a
y
e
r
”
is
r
e
s
p
o
n
s
i
b
l
e
f
o
r
l
e
a
r
n
i
n
g
i
m
p
o
r
t
a
n
t
f
e
a
t
u
r
e
s
f
r
o
m
t
h
e
e
x
t
r
a
ct
e
d
f
ea
t
u
r
e
s
a
n
d
d
o
t
h
e
c
l
as
s
i
f
i
c
at
i
o
n
b
a
s
e
d
o
n
R
e
L
U
ac
t
i
v
a
ti
o
n
f
u
n
c
t
i
o
n
.
I
n
t
h
e
n
e
x
t
s
t
a
g
e
,
m
u
l
t
i
d
i
m
e
n
s
i
o
n
f
e
a
t
u
r
es
a
r
e
c
o
n
v
e
r
t
e
d
i
n
t
o
o
n
e
f
e
a
t
u
r
e
v
e
c
t
o
r
a
n
d
f
i
n
a
l
l
y
,
i
n
t
h
e
l
a
s
t
s
t
a
g
e
,
t
h
e
S
o
f
t
M
a
x
l
a
y
e
r
is
u
s
e
d
t
o
d
o
t
h
e
c
l
as
s
i
f
ic
a
t
i
o
n
.
Fi
g
u
r
e
3
il
l
u
s
t
r
a
t
e
t
h
e
a
r
c
h
i
t
e
c
t
u
r
e
o
f
“C
N
N
b
as
e
d
m
u
l
t
i
-
c
l
ass
c
l
ass
i
f
i
e
r
”
m
o
d
e
l
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
S
p
a
tia
l a
n
d
ch
a
n
n
el
a
tten
tio
n
mec
h
a
n
is
m
fo
r
s
p
ee
ch
d
is
flu
en
cy
d
etec
tio
n
…
(
K
u
s
u
ma
H
.
R
.
)
2113
F
i
g
u
r
e
3
.
I
l
l
u
s
t
r
at
i
o
n
o
f
o
v
e
r
a
l
l
C
N
N
-
b
as
e
d
m
u
l
ti
c
l
ass
c
l
as
s
i
f
ic
a
t
i
o
n
a
r
c
h
i
t
ec
t
u
r
e
,
e
n
h
a
n
c
e
d
b
y
a
t
t
e
n
t
i
o
n
m
e
c
h
a
n
i
s
m
4.
RE
SU
L
T
AND
DI
SCUS
SI
O
N
T
h
is
s
ec
tio
n
p
r
esen
ts
a
d
etaile
d
d
escr
ip
tio
n
o
f
th
e
o
u
tco
m
e
o
f
p
r
o
p
o
s
ed
d
ee
p
lear
n
in
g
-
b
a
s
ed
s
tu
tter
d
etec
tio
n
s
y
s
tem
.
f
ir
s
t
s
u
b
s
ec
tio
n
d
escr
ib
es
th
e
ex
p
er
im
en
tal
s
etu
p
,
in
n
ex
t
s
u
b
s
ec
tio
n
we
d
escr
ib
e
th
e
d
ataset
d
etails
u
s
ed
in
th
is
wo
r
k
,
n
e
x
t
s
u
b
s
ec
tio
n
d
escr
ib
es
th
e
p
er
f
o
r
m
a
n
ce
m
ea
s
u
r
em
e
n
t
p
ar
am
eter
s
an
d
f
i
n
ally
co
m
p
ar
ativ
e
a
n
aly
s
is
is
p
r
esen
ted
wh
ich
s
h
o
ws th
e
r
o
b
u
s
tn
es
s
o
f
p
r
o
p
o
s
ed
m
o
d
el.
4
.
1
.
E
x
perim
ent
a
l set
up
T
h
e
p
r
o
p
o
s
ed
m
o
d
el
is
im
p
le
m
en
ted
b
y
u
s
in
g
Py
th
o
n
3
.
8
,
wh
ich
is
in
s
talled
o
n
W
in
d
o
w
s
p
latf
o
r
m
an
d
u
s
es
Nv
id
ia
GPU
with
P
y
T
o
r
ch
f
r
am
ewo
r
k
.
Fo
r
tr
ain
i
n
g
th
e
s
tu
tter
d
etec
tio
n
m
o
d
el
Ad
am
o
p
tim
izer
is
u
s
ed
with
min
i
-
b
a
tch
g
r
a
d
ien
t
d
escen
t
.
T
o
im
p
r
o
v
e
th
e
r
o
b
u
s
tn
ess
o
f
tr
ain
in
g
,
we
u
s
ed
d
ata
au
g
m
e
n
tatio
n
m
eth
o
d
i
n
wh
ich
r
ev
e
r
b
er
atio
n
,
r
a
n
d
o
m
p
itch
s
h
if
t
an
d
ad
d
i
tiv
e
n
o
is
e
f
u
n
ctio
n
s
a
r
e
co
n
s
id
er
ed
.
T
h
e
co
m
p
lete
au
g
m
en
tatio
n
p
r
o
ce
s
s
is
d
iv
id
ed
in
to
two
s
tep
s
,
f
ir
s
t
au
g
m
en
tatio
n
co
n
s
id
er
s
r
a
n
d
o
m
p
itch
s
h
if
t
+
r
ev
er
b
er
atio
n
o
f
au
g
m
en
tatio
n
,
an
d
s
ec
o
n
d
au
g
m
en
tatio
n
co
n
s
id
er
s
r
an
d
o
m
p
itch
s
h
if
t
+
ad
d
itiv
e
wh
it
e
Gau
s
s
ian
n
o
is
e
.
I
n
n
ex
t
s
tag
e,
th
e
f
ea
tu
r
e
e
x
tr
ac
tio
n
m
o
d
el
an
d
th
e
d
ee
p
lear
n
in
g
m
o
d
el
l
ea
r
n
th
e
p
atter
n
s
o
f
th
e
ap
p
r
o
p
r
iate
f
ea
tu
r
es.
4
.
2
.
Da
t
a
s
et
det
a
ils
T
h
e
s
u
p
er
v
is
ed
m
ac
h
i
n
e
lear
n
in
g
an
d
d
ee
p
lea
r
n
in
g
n
ee
d
s
lab
elled
d
ataset
h
o
wev
er
d
ataset
r
elate
d
to
th
ese
d
is
f
lu
en
cies
ar
e
v
er
y
lim
ited
,
m
o
r
e
o
v
er
,
th
e
lab
elle
d
d
ataset
its
elf
is
a
m
ajo
r
c
h
allen
g
e
f
o
r
s
p
ee
ch
d
is
f
lu
en
cy
d
etec
tio
n
.
I
n
th
is
wo
r
k
,
we
h
av
e
co
n
s
id
er
e
d
f
o
u
r
d
if
f
er
e
n
t
d
atasets
to
tr
ain
t
h
e
m
o
d
el
wh
ich
ar
e
UC
las
s
R
elea
s
e
1
,
UC
las
s
R
el
ea
s
e
2
,
Flu
en
cy
B
an
k
a
n
d
SEP
-
28k.
a.
UC
L
AS
S R
elea
s
e
1
an
d
2
:
T
h
e
Un
iv
er
s
ity
C
o
lleg
e
L
o
n
d
o
n
h
as
r
elea
s
ed
s
tu
tter
e
d
s
p
e
ec
h
r
ec
o
r
d
in
g
wh
ich
c
o
n
tain
s
s
tu
tter
in
g
s
p
ee
ch
r
ec
o
r
d
in
g
s
an
d
two
d
if
f
er
en
t
r
ec
o
r
d
i
n
g
s
ar
e
r
elea
s
ed
.
T
h
ese
s
p
ee
ch
r
ec
o
r
d
in
g
s
h
av
e
d
if
f
er
e
n
t
s
ev
er
ity
lev
els
[
2
1
]
.
T
h
e
f
ir
s
t
r
ec
o
r
d
in
g
is
r
elea
s
ed
in
2
0
0
4
an
d
o
th
e
r
is
r
elea
s
ed
in
2
0
0
8
.
B
o
th
r
el
ea
s
es
h
av
e
co
llected
d
ata
f
r
o
m
d
if
f
er
e
n
t
s
u
b
jects,
a
r
an
g
e
o
f
ag
es
f
r
o
m
5
to
4
7
y
ea
r
s
o
ld
wh
er
e
to
tal
o
f
1
3
9
s
p
ea
k
e
r
s
h
av
e
co
n
tr
ib
u
ted
in
d
ata
ac
q
u
is
itio
n
p
r
o
ce
s
s
.
T
h
ese
r
elea
s
es
h
av
e
d
ata
in
th
e
f
o
r
m
o
f
MP3
an
d
W
av
au
d
io
f
o
r
m
ats.
Ho
wev
er
,
th
e
d
ata
is
im
b
alan
c
ed
in
m
ale
v
s
f
e
m
ale
d
ata
s
am
p
les.
b.
Flu
en
cy
B
an
k
an
d
SEP
-
28k:
T
h
ese
d
atasets
ar
e
lab
elled
d
ataset
wh
er
e
u
n
s
u
p
er
v
is
ed
p
r
e
-
tr
ain
in
g
m
eth
o
d
s
h
av
e
b
ee
n
e
m
p
lo
y
ed
.
Flu
en
cy
B
an
k
is
a
d
ataset
o
b
tai
n
ed
f
r
o
m
late
-
s
p
ea
k
in
g
ch
ild
r
e
n
.
R
ec
en
tly
,
L
ea
et
a
l.
[
2
4
]
lab
elled
an
d
r
elea
s
ed
th
e
au
d
io
s
eg
m
e
n
tatio
n
f
r
o
m
F
lu
en
cy
B
an
k
d
atasets
.
I
n
o
r
d
er
to
m
ain
tain
co
n
s
is
ten
cy
with
p
r
ev
io
u
s
wo
r
k
s
,
we
co
n
s
id
er
th
e
5
s
ec
o
r
less
in
ter
v
al
au
d
io
is
id
ea
l
f
o
r
d
is
f
lu
en
cy
d
etec
tio
n
.
Fig
u
r
e
4
d
ep
icts
th
e
s
am
p
le
s
p
ee
ch
s
ig
n
als an
d
th
eir
co
r
r
esp
o
n
d
in
g
h
is
to
g
r
am
s
.
4
.
3
.
P
er
f
o
r
m
a
nce
m
e
a
s
urem
ent
pa
ra
m
et
er
s
T
h
e
tr
ac
e
o
f
Fig
u
r
e
4
(
a)
s
o
u
n
d
r
ep
etitio
n
is
ch
ar
ac
ter
ized
b
y
a
m
u
ltip
le
r
ep
ea
ted
s
p
ec
tr
al
b
u
r
s
t
o
cc
u
r
r
in
g
at
clo
s
e
tim
e
in
ter
v
als.
T
h
ese
b
u
r
s
ts
in
d
icate
th
at,
r
ep
ea
ted
attem
p
ts
t
o
p
r
o
d
u
ce
an
in
itial
p
h
o
n
em
e
,
th
ese
b
u
r
s
ts
ar
e
o
cc
u
r
r
i
n
g
ty
p
ically
with
s
m
all
p
au
s
es
b
etwe
en
s
eg
m
en
ts
.
T
h
e
d
i
s
co
n
tin
u
ity
in
th
e
s
p
ec
tr
o
g
r
am
s
h
o
ws
th
e
en
er
g
y
p
atter
n
s
,
t
h
e
h
ig
h
tem
p
o
r
al
v
ar
iab
ilit
y
an
d
s
h
o
r
t
d
u
r
atio
n
,
m
ak
i
n
g
d
etec
tio
n
ch
allen
g
in
g
.
Fig
u
r
e
4
(
b
)
illu
s
tr
ates
p
r
o
lo
n
g
atio
n
,
ap
p
ea
r
s
as
a
co
n
tin
u
o
u
s
,
l
o
n
g
d
u
r
ati
o
n
s
p
ec
tr
al
b
an
d
with
a
s
tab
le
f
r
eq
u
en
cy
co
n
ten
t.
I
t
r
e
f
lects
th
e
s
u
s
tain
ed
ar
ticu
latio
n
o
f
a
s
in
g
le
p
h
o
n
e
m
e
with
o
u
t
in
ter
r
u
p
tio
n
.
T
h
is
d
is
f
lu
en
cy
is
ea
s
ier
to
d
etec
t
d
u
e
to
its
clea
r
d
u
r
atio
n
-
b
ased
ch
ar
ac
ter
is
tics
an
d
lo
w
s
p
e
ctr
al
v
ar
iatio
n
o
v
er
tim
e.
Fig
u
r
e
4
(
c)
illu
s
tr
ates
wo
r
d
r
ep
etitio
n
,
in
v
o
lv
es
r
ep
ea
ted
an
d
well
-
s
tr
u
ctu
r
ed
s
p
e
ctr
al
s
eg
m
en
ts
,
ea
ch
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
1
6
,
No
.
4
,
Au
g
u
s
t
20
2
6
:
2
1
0
6
-
2
119
2114
s
eg
m
en
t
is
co
r
r
esp
o
n
d
in
g
t
o
a
f
u
ll
wo
r
d
.
T
h
ese
s
eg
m
en
ts
ar
e
s
ep
ar
ated
b
y
s
h
o
r
t
p
au
s
es
an
d
ex
h
ib
it
s
im
ilar
p
atter
n
s
ac
r
o
s
s
r
ep
etitio
n
s
.
D
u
e
to
th
ei
r
lar
g
e
r
tem
p
o
r
al
s
tr
u
ctu
r
e,
t
h
ey
ar
e
r
elativ
ely
ea
s
ier
to
id
e
n
tify
u
s
in
g
s
eg
m
en
tatio
n
an
d
p
atter
n
r
ec
o
g
n
itio
n
tech
n
iq
u
es.
(
a)
(
b
)
(
c)
Fig
u
r
e
4
.
T
im
e
an
d
f
r
eq
u
en
cy
s
p
ec
tr
a
o
f
(
a)
s
o
u
n
d
r
ep
etitio
n
,
(
b
)
p
r
o
lo
n
g
atio
n
,
an
d
(
c)
wo
r
d
r
ep
etitio
n
Fig
u
r
e
5
illu
s
tr
ates
a
class
-
w
is
e
co
n
f
u
s
io
n
m
atr
ix
[
2
0
]
f
o
r
m
u
lti
-
class
class
if
icatio
n
,
em
p
h
asizin
g
h
o
w
a
m
o
d
el
p
e
r
f
o
r
m
s
o
n
a
s
p
ec
if
ic
class
,
b
y
ca
teg
o
r
izin
g
p
r
e
d
ictio
n
s
in
to
tr
u
e
p
o
s
itiv
es,
f
alse
n
eg
ativ
es,
f
alse
p
o
s
itiv
es,
an
d
tr
u
e
n
e
g
ativ
es,
th
er
eb
y
o
f
f
er
in
g
d
etailed
in
s
ig
h
t
in
t
o
th
e
m
o
d
el’
s
b
e
h
a
v
io
r
f
o
r
th
at
class
.
S
p
e
c
ifi
c
a
ll
y
,
F
i
g
u
re
s
5
(a
)
–
(d
)
e
x
p
li
c
it
l
y
d
e
tail
th
e
p
re
d
ictio
n
re
su
lt
s
fo
r
e
a
c
h
c
las
s:
so
u
n
d
re
p
e
ti
ti
o
n
,
wo
rd
re
p
e
ti
ti
o
n
,
p
ro
l
o
n
g
a
ti
o
n
,
a
n
d
i
n
terje
c
ti
o
n
.
T
h
e
p
er
f
o
r
m
a
n
ce
o
f
s
tu
tter
d
ete
ctio
n
m
eth
o
d
s
is
m
ea
s
u
r
ed
in
ter
m
s
o
f
p
r
ec
is
io
n
,
ac
cu
r
ac
y
a
n
d
F1
-
s
co
r
e
wh
ich
ar
e
o
b
tain
e
d
b
ased
o
n
th
e
co
n
f
u
s
io
n
m
atr
ix
.
B
ased
o
n
th
is
co
n
f
u
s
io
n
m
at
r
ix
,
we
o
b
tain
ed
th
e
ac
cu
r
ac
y
,
F1
-
s
co
r
e
an
d
p
r
ec
is
io
n
wh
ich
ar
e
ex
p
r
ess
ed
in
T
ab
le
1.
Fu
r
th
er
,
we
ex
te
n
d
th
is
ex
p
er
im
en
tal
an
aly
s
is
f
o
r
class
im
b
alan
ce
tr
ain
in
g
s
ce
n
ar
io
wh
er
e
we
co
m
p
ar
e
th
e
p
er
f
o
r
m
a
n
ce
f
o
r
weig
h
ted
cr
o
s
s
en
tr
o
p
y
(
W
C
E
)
,
m
u
lti
b
r
an
ch
tr
ain
in
g
m
et
h
o
d
s
.
T
h
e
o
b
tain
e
d
p
e
r
f
o
r
m
a
n
c
e
f
o
r
t
h
i
s
e
x
p
e
r
i
m
e
n
t
i
s
d
e
m
o
n
s
t
r
at
e
d
i
n
t
h
e
f
o
l
l
o
w
i
n
g
t
a
b
l
e
.
S
i
m
i
l
a
r
l
y
,
we
e
x
t
e
n
d
t
h
e
e
x
p
e
r
i
m
e
n
t
f
o
r
d
a
t
a
a
u
g
m
e
n
t
a
t
i
o
n
m
et
h
o
d
s
w
h
e
r
e
w
e
h
a
v
e
c
o
n
s
i
d
e
r
e
d
b
a
b
b
l
e
,
r
e
v
e
r
b
e
r
a
t
i
o
n
,
m
u
s
i
c
,
a
n
d
n
o
is
e
a
u
g
m
e
n
t
a
t
i
o
n
.
T
ab
le
2
g
iv
es
a
co
m
p
ar
is
o
n
o
f
th
e
d
if
f
er
en
t
s
tu
tter
d
ete
ctio
n
m
o
d
el’
s
p
e
r
f
o
r
m
an
ce
s
am
o
n
g
th
e
d
if
f
er
en
t
ty
p
es
o
f
s
tu
tter
in
g
d
i
s
f
lu
en
cy
ca
teg
o
r
ies.
Ou
t
o
f
t
h
ese
m
o
d
els,
C
o
n
v
L
STM
-
b
ased
b
aselin
e
tech
n
iq
u
e
s
h
o
ws
less
ef
f
ec
tiv
en
ess
,
b
e
ca
u
s
e
wh
ich
is
m
ain
ly
d
ep
e
n
d
en
t
o
n
co
n
tr
o
lled
en
v
ir
o
n
m
en
t
an
d
ac
o
u
s
tic
f
ea
tu
r
es,
an
d
it
h
as
lim
ited
f
l
ex
ib
ilit
y
in
ca
p
tu
r
in
g
co
m
p
le
x
s
tu
tter
p
atter
n
s
.
I
n
co
n
t
r
ast,
R
esNet
+
B
iLST
M
an
d
Stu
tter
Net,
ar
ch
itectu
r
es
ar
e
tem
p
o
r
al
m
o
d
els
wh
ich
p
r
o
v
id
e
im
p
r
o
v
e
d
ef
f
icien
cy
a
n
d
m
o
r
e
f
le
x
ib
ilit
y
,
th
ey
in
co
r
p
o
r
ated
d
ee
p
er
f
ea
t
u
r
e
ex
tr
ac
tio
n
tech
n
iq
u
es
,
to
ca
p
tu
r
e
th
e
tem
p
o
r
al
f
ea
tu
r
es.
T
h
ey
s
u
cc
ess
f
u
lly
ac
h
iev
ed
h
ig
h
e
r
F1
-
s
co
r
es
ac
r
o
s
s
m
o
s
t
s
tu
tter
ty
p
es.
T
h
e
im
p
r
o
v
e
d
r
esu
lts
ar
e
s
h
o
wn
in
t
h
e
f
o
llo
win
g
tab
le
,
an
d
it
ju
s
tify
th
at
f
ea
tu
r
e
ex
tr
ac
tio
n
with
atten
tio
n
-
b
ased
al
g
o
r
ith
m
,
e
n
h
an
ce
s
f
ea
tu
r
e
r
ep
r
esen
tatio
n
s
.
Fro
m
T
ab
le
2
we
ca
n
a
n
aly
ze
t
h
e
p
e
r
f
o
r
m
a
n
ce
o
f
t
h
e
s
tu
tter
d
etec
t
io
n
m
o
d
els
d
e
p
en
d
i
n
g
m
ain
ly
u
p
o
n
th
e
h
o
w
well
s
p
ee
ch
f
ea
tu
r
es
ar
e
r
ep
r
esen
te
d
an
d
h
o
w
ef
f
ec
tiv
ely
s
tu
tter
i
n
g
-
r
elate
d
c
u
es
ar
e
em
p
h
asized
.
B
ased
o
n
th
ese
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
S
p
a
tia
l a
n
d
ch
a
n
n
el
a
tten
tio
n
mec
h
a
n
is
m
fo
r
s
p
ee
ch
d
is
flu
en
cy
d
etec
tio
n
…
(
K
u
s
u
ma
H
.
R
.
)
2115
o
b
s
er
v
atio
n
s
,
th
e
p
r
o
p
o
s
ed
ch
an
n
el
a
n
d
s
p
atial
atten
tio
n
m
e
ch
an
is
m
is
in
tr
o
d
u
ce
d
to
h
el
p
th
e
m
o
d
el
f
o
cu
s
o
n
th
e
m
o
s
t r
elev
an
t a
c
o
u
s
tic
p
atter
n
s
,
th
er
eb
y
im
p
r
o
v
in
g
its
ab
ilit
y
to
d
etec
t stu
tter
in
g
ev
e
n
ts
.
(
a)
(
b
)
(
c)
(
d
)
Fig
u
r
e
5.
C
o
n
f
u
s
io
n
m
atr
i
x
f
o
r
(
a)
s
o
u
n
d
r
e
p
etitio
n
,
(
b
)
wo
r
d
r
ep
etitio
n
,
(
c)
p
r
o
lo
n
g
atio
n
,
an
d
(
d
)
in
ter
jectio
n
T
ab
le
1
.
P
a
r
a
m
e
t
e
r
s
a
n
d
i
ts
r
e
le
v
a
n
t
e
x
p
r
e
s
s
i
o
n
P
a
r
a
me
t
e
r
n
a
m
e
Ex
p
r
e
ssi
o
n
P
r
e
c
i
s
i
o
n
+
R
e
c
a
l
l
+
To
t
a
l
a
c
c
u
r
a
c
y
(
TA
)
+
+
+
+
F1
-
sc
o
r
e
2
×
×
+
T
ab
le
2
.
C
o
m
p
a
r
a
t
i
v
e
a
n
a
l
y
s
is
(
a
c
c
u
r
a
c
y
%
)
o
f
b
a
s
el
i
n
es
M
e
t
h
o
d
R
P
B
In
F
TA
F
1
(
%)
C
o
n
v
LST
M
+
F
(
M
F
C
C
)
(
B
L
1
)
2
2
.
8
3
1
0
.
6
1
6
.
3
4
5
6
.
7
4
7
2
.
3
5
5
2
.
6
8
3
4
.
0
0
C
o
n
v
LST
M
+
F
(
P
h
o
n
e
)
1
0
.
1
8
0
1
.
0
6
0
0
.
3
5
4
3
.
8
8
7
4
.
4
8
4
8
.
4
3
2
4
.
0
0
C
o
n
v
LST
M
+
F
(
F
0
+
M
F
C
C
)
1
9
.
2
8
9
.
5
5
8
.
5
1
5
1
.
7
5
6
6
.
6
0
4
8
.
4
7
3
0
.
8
0
R
e
sN
e
t
+
B
i
LST
M
1
8
.
7
6
4
1
.
2
4
5
.
4
7
5
7
.
1
8
8
8
.
1
9
6
2
.
3
6
4
3
.
1
2
S
t
u
t
t
e
r
N
e
t
2
7
.
1
4
3
2
.
5
5
2
.
9
6
5
7
.
7
4
8
7
.
6
0
6
2
.
5
7
4
2
.
8
4
T
ab
le
3
s
h
o
ws
th
e
r
esu
lt
o
f
th
e
im
p
r
o
v
ed
atten
tio
n
-
b
ased
f
ea
tu
r
e
ex
tr
ac
tio
n
m
o
d
el,
wh
ic
h
in
clu
d
es
ad
d
itio
n
al
tech
n
iq
u
es
s
u
ch
as
weig
h
ted
cr
o
s
s
-
en
tr
o
p
y
a
n
d
m
u
lti
-
b
r
an
ch
s
tr
u
ct
u
r
es,
to
en
h
an
ce
s
tu
tter
d
etec
tio
n
.
T
h
e
u
s
e
o
f
weig
h
t
ed
cr
o
s
s
-
en
tr
o
p
y
an
d
m
u
lti
-
b
r
an
ch
a
r
ch
itectu
r
es
tec
h
n
iq
u
e
s
in
th
e
p
r
o
p
o
s
ed
m
o
d
el
im
p
r
o
v
es
th
e
p
er
f
o
r
m
a
n
ce
m
o
d
e
r
ately
co
m
p
ar
ed
to
b
aselin
e
co
u
n
ter
p
a
r
ts
.
T
h
e
p
r
o
p
o
s
ed
m
o
d
el
is
ab
l
e
to
ac
h
iev
e
th
e
h
ig
h
est
F1
-
s
co
r
e
am
o
n
g
all
c
o
m
p
a
r
ed
m
eth
o
d
s
,
th
is
in
d
icate
s
th
at
th
e
p
r
o
p
o
s
ed
m
o
d
el
p
r
o
v
id
es
m
o
r
e
b
alan
ce
d
an
d
ef
f
ec
tiv
e
d
etec
tio
n
o
f
s
tu
tter
in
g
ev
en
ts
.
Evaluation Warning : The document was created with Spire.PDF for Python.