I
AE
S In
t
er
na
t
io
na
l J
o
urna
l o
f
Art
if
icia
l In
t
ellig
ence
(
I
J
-
AI
)
Vo
l.
15
,
No
.
5
,
Octo
b
er
20
26
,
p
p
.
4
8
9
3
~
490
6
I
SS
N:
2
2
5
2
-
8
9
3
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijai.v
15
.i
5
.
p
p
4
8
9
3
-
490
6
4893
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e.
co
m
Em
pirica
l eva
lua
t
io
n of genera
tive
fusio
n decoding
f
o
r
Indo
nesia
n
g
ener
a
l and medica
l sp
eech r
e
co
g
nition
Asril J
a
rin,
L
y
la
Rus
la
na
Ai
ni,
Ag
un
g
Sa
nto
s
a
,
G
un
a
rso
,
M
o
ha
m
m
a
d T
eduh
Ulini
a
n
s
y
a
h,
E
lv
ira
Nurf
a
dh
ila
h,
Si
s
k
a
P
e
bia
na
R
e
se
a
r
c
h
C
e
n
t
e
r
f
o
r
D
a
t
a
a
n
d
I
n
f
o
r
ma
t
i
o
n
S
c
i
e
n
c
e
s
,
R
e
se
a
r
c
h
O
r
g
a
n
i
z
a
t
i
o
n
f
o
r
El
e
c
t
r
o
n
i
c
s a
n
d
I
n
f
o
r
ma
t
i
c
s
,
N
a
t
i
o
n
a
l
R
e
se
a
r
c
h
a
n
d
I
n
n
o
v
a
t
i
o
n
A
g
e
n
c
y
(
B
R
I
N
)
,
B
a
n
d
u
n
g
,
I
n
d
o
n
e
s
i
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
No
v
5
,
2
0
2
5
R
ev
is
ed
Au
g
3
,
2
0
2
6
Acc
ep
ted
Au
g
12
,
2
0
2
6
Th
is
st
u
d
y
e
m
p
iri
c
a
ll
y
e
v
a
l
u
a
tes
g
e
n
e
ra
ti
v
e
f
u
sio
n
d
e
c
o
d
i
n
g
(
G
F
D)
with
o
p
e
n
-
we
ig
h
t
larg
e
la
n
g
u
a
g
e
m
o
d
e
ls
f
o
r
I
n
d
o
n
e
sia
n
a
u
t
o
m
a
ti
c
sp
e
e
c
h
re
c
o
g
n
it
i
o
n
(ASR)
u
n
d
e
r
two
c
o
n
tr
o
ll
e
d
o
ffli
n
e
c
o
n
d
i
ti
o
n
s:
g
e
n
e
ra
l
re
a
d
sp
e
e
c
h
(F
LE
URS
-
ID)
a
n
d
m
e
d
ic
a
l
d
icta
ti
o
n
(Ha
rk
it
)
.
T
h
e
train
i
n
g
-
fre
e
b
y
te
-
lev
e
l
fu
sio
n
i
n
teg
ra
tes
th
re
e
fr
o
z
e
n
lan
g
u
a
g
e
-
m
o
d
e
l
b
a
c
k
e
n
d
s
—
S
EA
-
LION,
S
a
h
a
b
a
t
-
AI,
a
n
d
M
istral
—
wit
h
fr
o
z
e
n
Wh
isp
e
r
-
larg
e
-
v3
d
u
ri
n
g
d
e
c
o
d
in
g
.
A
d
e
v
e
lo
p
m
e
n
t
p
a
rti
ti
o
n
wa
s
u
se
d
to
se
lec
t
th
e
sta
ti
c
fu
sio
n
we
ig
h
t
a
n
d
p
ro
m
p
ti
n
g
c
o
n
d
it
io
n
,
a
n
d
a
h
e
ld
-
o
u
t
tes
t
p
a
rt
it
io
n
wa
s
u
se
d
fo
r
fi
n
a
l
e
v
a
lu
a
ti
o
n
.
On
F
LE
URS
-
ID,
th
e
b
e
st
G
F
D
c
o
n
fig
u
ra
ti
o
n
re
d
u
c
e
d
c
o
rp
u
s
-
lev
e
l
wo
r
d
e
rr
o
r
ra
te
(
WE
R)
fr
o
m
5
.
9
1
%
to
4
.
1
7
%
,
c
o
rre
sp
o
n
d
in
g
t
o
a
29.
44
%
re
lati
v
e
re
d
u
c
t
io
n
.
O
n
Ha
rk
it
,
it
re
d
u
c
e
d
WER
fro
m
1
5
.
8
9
%
t
o
1
2
.
3
0
%
,
a
2
2
.
5
9
%
re
lati
v
e
re
d
u
c
ti
o
n
.
P
a
ired
u
tt
e
ra
n
c
e
-
lev
e
l
tes
ts
y
ield
e
d
p
<
0
.
0
1
fo
r
b
o
th
d
a
tas
e
ts.
S
e
v
e
r
a
l
n
o
n
-
p
r
o
m
p
ted
c
o
n
fi
g
u
ra
ti
o
n
s
a
lrea
d
y
imp
ro
v
e
d
o
v
e
r
t
h
e
u
n
f
u
se
d
W
h
i
sp
e
r
-
larg
e
-
v
3
b
a
se
li
n
e
,
su
g
g
e
stin
g
th
a
t
t
h
e
e
x
tern
a
l
lan
g
u
a
g
e
-
m
o
d
e
l
p
rio
r
c
o
n
tri
b
u
te
d
m
o
re
c
o
n
siste
n
tl
y
t
h
a
n
e
x
p
li
c
i
t
p
ro
m
p
ti
n
g
.
P
ro
m
p
ti
n
g
p
ro
d
u
c
e
d
l
imited
c
h
a
n
g
e
s
o
n
F
L
EURS
-
ID
a
n
d
larg
e
r
b
u
t
b
a
c
k
e
n
d
-
d
e
p
e
n
d
e
n
t
c
h
a
n
g
e
s
o
n
Ha
rk
i
t
.
T
h
e
re
su
lt
s
su
p
p
o
rt
G
F
D
a
s
a
train
in
g
-
fre
e
d
e
c
o
d
i
n
g
stra
teg
y
u
n
d
e
r
th
e
e
v
a
lu
a
ted
c
o
n
d
it
io
n
s,
wh
il
e
b
ro
a
d
e
r
v
a
li
d
a
ti
o
n
is
re
q
u
ired
f
o
r
n
o
is
y
,
sp
o
n
ta
n
e
o
u
s,
stre
a
m
in
g
,
a
n
d
c
li
n
ica
l
d
e
p
lo
y
m
e
n
t
se
tt
i
n
g
s.
K
ey
w
o
r
d
s
:
Au
to
m
atic
s
p
ee
ch
r
ec
o
g
n
itio
n
Do
m
ain
ad
ap
tatio
n
Gen
er
ativ
e
f
u
s
io
n
d
ec
o
d
i
n
g
I
n
d
o
n
esian
s
p
ee
ch
L
ar
g
e
lan
g
u
ag
e
m
o
d
els
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
:
Asril
J
ar
in
R
esear
ch
C
en
ter
f
o
r
Data
an
d
I
n
f
o
r
m
atio
n
Scien
ce
s
,
R
esear
ch
Or
g
an
izatio
n
f
o
r
E
lectr
o
n
ics
an
d
I
n
f
o
r
m
atics
Natio
n
al
R
esear
ch
an
d
I
n
n
o
v
a
tio
n
Ag
en
cy
B
an
d
u
n
g
,
I
n
d
o
n
esia
E
m
ail:
asri0
0
3
@
b
r
in
.
g
o
.
id
1.
I
NT
RO
D
UCT
I
O
N
Au
to
m
atic
s
p
ee
ch
r
ec
o
g
n
itio
n
(
ASR
)
h
as
ad
v
an
ce
d
s
u
b
s
tan
tially
o
v
er
th
e
p
ast
d
ec
ad
e
,
tr
a
n
s
itio
n
in
g
f
r
o
m
h
y
b
r
id
h
id
d
e
n
Ma
r
k
o
v
m
o
d
el
(
HM
M)
-
b
ased
s
y
s
tem
s
to
f
u
lly
n
e
u
r
al,
en
d
-
to
-
e
n
d
a
p
p
r
o
ac
h
es
[
1
]
,
[
2
]
.
Pio
n
ee
r
in
g
a
r
ch
itectu
r
es
s
u
ch
as
d
ee
p
r
ec
u
r
r
en
t
n
eu
r
al
n
etw
o
r
k
s
[
3
]
an
d
th
e
atten
tio
n
-
b
ased
en
co
d
er
–
d
ec
o
d
e
r
m
o
d
el
lis
ten
,
atten
d
,
an
d
s
p
ell
[
4
]
p
av
ed
th
e
way
f
o
r
d
ir
ec
t
ac
o
u
s
tic
-
to
-
tex
t
m
ap
p
in
g
.
Fu
r
t
h
er
b
r
ea
k
th
r
o
u
g
h
s
in
d
ata
au
g
m
e
n
tatio
n
[
5
]
,
s
elf
-
s
u
p
er
v
is
ed
r
ep
r
esen
tatio
n
lea
r
n
in
g
[
6
]
–
[
8
]
,
a
n
d
ar
ch
itectu
r
al
in
n
o
v
atio
n
s
s
u
ch
as
th
e
co
n
f
o
r
m
e
r
[
9
]
h
a
v
e
s
u
b
s
tan
tially
im
p
r
o
v
ed
r
o
b
u
s
tn
es
s
an
d
s
ca
lab
ilit
y
.
T
h
ese
ad
v
an
ce
s
,
co
m
b
in
ed
with
m
ass
iv
e
tr
ain
in
g
d
atasets
,
h
av
e
en
ab
led
s
tate
-
of
-
th
e
-
a
r
t
s
y
s
tem
s
s
u
ch
as
W
h
i
s
p
er
[
1
0
]
to
ac
h
iev
e
n
ea
r
-
h
u
m
an
tr
an
s
cr
ip
tio
n
ac
cu
r
ac
y
u
n
d
er
c
lean
,
h
ig
h
-
r
eso
u
r
ce
co
n
d
itio
n
s
.
Desp
ite
th
is
p
r
o
g
r
ess
,
I
n
d
o
n
esian
ASR
co
n
tin
u
es
to
f
ac
e
u
n
iq
u
e
c
h
allen
g
es.
T
h
e
I
n
d
o
n
esian
lan
g
u
ag
e
ex
h
ib
its
ag
g
lu
tin
ati
v
e
m
o
r
p
h
o
lo
g
y
an
d
p
r
o
d
u
cti
v
e
d
er
i
v
atio
n
al
af
f
ix
atio
n
,
w
h
ich
s
u
b
s
tan
tially
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
15
,
No
.
5
,
Octo
b
er
20
26
:
4
8
9
3
-
490
6
4894
in
cr
ea
s
e
lex
ical
v
ar
iab
ilit
y
an
d
s
u
r
f
ac
e
-
f
o
r
m
d
iv
er
s
ity
in
s
p
o
k
en
lan
g
u
ag
e
[
1
1
]
,
[
1
2
]
.
I
n
ad
d
itio
n
,
f
r
eq
u
en
t
I
n
d
o
n
esian
–
E
n
g
lis
h
c
o
d
e
-
s
w
itch
in
g
—
a
co
m
m
o
n
p
h
en
o
m
en
o
n
in
b
o
th
i
n
f
o
r
m
al
an
d
p
r
o
f
ess
io
n
al
co
m
m
u
n
icatio
n
—
in
tr
o
d
u
c
es
s
tr
u
ctu
r
al
an
d
p
h
o
n
o
lo
g
ical
v
ar
iab
ilit
y
,
f
u
r
t
h
er
co
m
p
licatin
g
ac
o
u
s
tic
an
d
lan
g
u
ag
e
m
o
d
elin
g
f
o
r
ASR
s
y
s
tem
s
[
1
3
]
–
[
1
5
]
.
T
h
ese
lin
g
u
is
tic
ch
ar
ac
ter
is
tics
,
co
m
b
in
ed
with
th
e
lim
ited
av
ailab
ilit
y
o
f
an
n
o
tated
d
o
m
ain
-
s
p
ec
if
ic
co
r
p
o
r
a
,
d
e
g
r
ad
e
r
ec
o
g
n
itio
n
p
e
r
f
o
r
m
an
ce
in
s
p
ec
ialized
co
n
te
x
ts
s
u
ch
as m
ed
ical
tr
an
s
cr
ip
tio
n
.
Prio
r
s
tu
d
ies
o
n
m
e
d
ical
ASR
co
n
s
is
ten
tly
r
ep
o
r
t
s
u
b
s
tan
tial in
cr
ea
s
es
in
er
r
o
r
w
h
en
co
n
f
r
o
n
ted
with
ter
m
in
o
lo
g
y
-
d
e
n
s
e
u
tter
an
ce
s
an
d
d
o
m
ain
-
s
p
ec
if
ic
ex
p
r
e
s
s
io
n
s
,
ev
en
with
lar
g
e
-
s
ca
le
n
eu
r
al
m
o
d
els.
Alth
o
u
g
h
m
u
ltil
in
g
u
al
m
o
d
el
s
s
u
ch
as
cr
o
s
s
-
lin
g
u
al
s
p
ee
ch
r
ep
r
esen
tatio
n
(
XL
S
-
R
)
an
d
W
h
is
p
er
-
lar
g
e
-
v
3
d
em
o
n
s
tr
ate
s
tr
o
n
g
c
r
o
s
s
-
lin
g
u
al
g
en
e
r
aliza
tio
n
,
t
h
eir
ac
cu
r
ac
y
co
n
s
is
ten
tly
d
ec
lin
es
u
n
d
e
r
d
o
m
ain
-
m
is
m
atch
co
n
d
itio
n
s
,
p
ar
ticu
lar
ly
wh
e
n
en
co
u
n
ter
i
n
g
r
ar
e
o
r
m
o
r
p
h
o
l
o
g
ically
c
o
m
p
lex
wo
r
d
f
o
r
m
s
an
d
co
d
e
-
s
witch
ed
u
tter
an
ce
s
th
at
f
all
o
u
ts
id
e
th
eir
p
r
e
-
tr
ai
n
in
g
d
is
tr
ib
u
tio
n
s
[
1
6
]
–
[
1
8
]
.
R
ec
en
t
p
r
o
g
r
e
s
s
in
m
u
ltil
in
g
u
al
d
atasets
—
in
clu
d
in
g
f
ew
-
s
h
o
t
lear
n
in
g
ev
alu
atio
n
o
f
u
n
iv
er
s
al
r
ep
r
esen
tatio
n
s
o
f
s
p
ee
ch
(
FLE
UR
S
)
[
1
9
]
,
m
u
ltil
in
g
u
al
L
ib
r
iSp
ee
ch
(
ML
S
)
[
2
0
]
,
an
d
m
ass
iv
ely
m
u
ltil
in
g
u
al
s
p
ee
ch
(
MMS
)
[
2
1
]
—
h
av
e
s
ig
n
if
ican
tly
ad
v
an
ce
d
c
r
o
s
s
-
lin
g
u
al
tr
an
s
f
er
f
o
r
ASR
.
Ho
wev
er
,
th
ese
co
r
p
o
r
a
s
till
lack
s
u
f
f
icien
t
co
v
er
ag
e
o
f
d
o
m
ain
-
s
p
ec
if
ic
lin
g
u
is
tic
p
h
e
n
o
m
en
a
,
p
ar
ticu
lar
ly
in
l
o
w
-
r
eso
u
r
ce
d
o
m
ain
s
lik
e
h
ea
lth
ca
r
e.
I
n
s
u
ch
s
ettin
g
s
,
en
d
-
to
-
en
d
m
o
d
els
o
f
ten
m
is
r
ec
o
g
n
ize
s
p
ec
ialized
ter
m
s
,
ab
b
r
ev
i
atio
n
s
,
an
d
co
n
te
x
t
-
d
ep
e
n
d
en
t
ex
p
r
ess
io
n
s
th
at
d
ev
iate
f
r
o
m
th
eir
p
r
e
-
tr
ain
in
g
d
is
tr
ib
u
tio
n
.
E
x
ter
n
al
lan
g
u
ag
e
-
m
o
d
el
in
teg
r
atio
n
in
en
d
-
to
-
en
d
ASR
s
p
an
s
s
ev
er
al
d
is
tin
ct
p
ar
ad
ig
m
s
.
C
o
n
v
en
tio
n
al
s
h
allo
w
f
u
s
io
n
p
er
f
o
r
m
s
lo
g
-
lin
ea
r
in
ter
p
o
latio
n
o
f
in
d
ep
e
n
d
en
tly
tr
ain
ed
ASR
an
d
lan
g
u
ag
e
-
m
o
d
el
s
co
r
es
d
u
r
in
g
b
ea
m
-
s
ea
r
ch
in
f
er
en
ce
an
d
d
o
es
n
o
t
r
eq
u
ir
e
r
et
r
ain
in
g
th
e
ASR
m
o
d
el
[
2
2
]
.
T
r
ain
in
g
-
tim
e
f
u
s
io
n
m
eth
o
d
s
,
s
u
ch
as
co
ld
f
u
s
io
n
,
in
s
t
ea
d
tr
ain
th
e
s
eq
u
en
ce
-
to
-
s
e
q
u
en
ce
r
ec
o
g
n
ize
r
to
g
eth
er
with
a
f
ix
ed
p
r
etr
ai
n
e
d
lan
g
u
ag
e
m
o
d
el
[
2
3
]
.
C
o
n
te
x
tu
al
ap
p
r
o
ac
h
es
s
u
ch
as
co
n
tex
tu
al
lis
ten
,
atten
d
an
d
s
p
ell
(
C
L
AS
)
jo
in
tly
o
p
tim
ize
th
e
r
ec
o
g
n
izer
with
em
b
ed
d
in
g
s
o
f
task
-
o
r
u
s
er
-
s
p
ec
if
ic
co
n
tex
t
p
h
r
ases
[
2
4
]
,
wh
e
r
ea
s
s
ec
o
n
d
-
p
ass
an
d
s
y
n
ch
r
o
n
o
u
s
m
et
h
o
d
s
s
u
ch
as
R
esco
r
eBER
T
[
2
5
]
an
d
s
p
ee
d
y
ASR
-
lar
g
e
lan
g
u
ag
e
m
o
d
el
(
LLM
)
s
y
n
c
h
r
o
n
o
u
s
a
g
g
r
e
g
atio
n
(
SALSA
)
[
2
6
]
in
t
r
o
d
u
ce
li
n
g
u
is
tic
in
f
o
r
m
atio
n
at
d
if
f
er
en
t
s
tag
es
o
f
th
e
r
e
co
g
n
itio
n
p
r
o
ce
s
s
.
T
h
ese
ap
p
r
o
ac
h
es
d
if
f
er
in
th
eir
tr
ai
n
in
g
r
eq
u
ir
em
en
ts
,
co
n
tex
tu
al
r
eso
u
r
ce
s
,
s
co
r
e
c
o
m
p
atib
ilit
y
,
ar
ch
itectu
r
al
ass
u
m
p
tio
n
s
,
an
d
in
f
er
e
n
ce
-
tim
e
c
o
s
t.
Hsu
et
a
l.
[
2
7
]
in
tr
o
d
u
ce
d
g
en
er
ativ
e
f
u
s
io
n
d
ec
o
d
in
g
(
GFD)
as
a
g
en
e
r
ativ
e,
s
h
allo
w
-
f
u
s
io
n
f
r
am
ew
o
r
k
f
o
r
co
m
b
in
in
g
au
to
r
eg
r
ess
iv
e
r
ec
o
g
n
itio
n
m
o
d
els
with
lar
g
e
lan
g
u
ag
e
m
o
d
els
wh
o
s
e
to
k
en
v
o
ca
b
u
lar
ie
s
m
ay
d
if
f
e
r
.
GFD
m
ap
s
m
o
d
el
lik
elih
o
o
d
s
in
to
a
s
h
ar
ed
b
y
te
-
lev
el
r
ep
r
esen
tati
o
n
an
d
s
y
n
ch
r
o
n
izes
ASR
an
d
L
L
M
p
r
o
g
r
ess
io
n
d
u
r
in
g
b
ea
m
-
s
ea
r
ch
d
ec
o
d
i
n
g
.
T
h
is
d
esig
n
p
er
m
its
in
s
tr
u
ctio
n
-
co
n
d
itio
n
ed
L
L
M
p
r
io
r
s
to
b
e
in
teg
r
ate
d
with
o
u
t r
etr
ain
in
g
eith
er
m
o
d
e
l.
T
h
e
p
r
esen
t
s
tu
d
y
ev
alu
ates
t
h
e
ef
f
ec
tiv
en
ess
o
f
t
h
is
tr
ain
in
g
-
f
r
ee
b
y
te
-
le
v
el
f
u
s
io
n
f
o
r
I
n
d
o
n
esian
ASR
in
b
o
th
g
en
er
al
r
ea
d
s
p
e
ec
h
an
d
ter
m
in
o
lo
g
y
-
d
en
s
e
m
ed
ical
d
ictatio
n
.
Acc
o
r
d
in
g
l
y
,
th
is
s
tu
d
y
co
n
d
u
cts
an
em
p
ir
ical
ev
alu
atio
n
o
f
GFD
f
o
r
I
n
d
o
n
esian
ASR
ac
r
o
s
s
two
r
ep
r
esen
tativ
e
d
o
m
ain
s
:
i)
g
en
er
al
s
p
ee
ch
:
r
ep
r
esen
tin
g
co
n
tr
o
lled
,
clea
n
r
ea
d
-
s
p
ee
ch
co
n
d
itio
n
s
;
an
d
ii)
m
ed
ical
s
p
ee
ch
:
ch
ar
ac
ter
i
ze
d
b
y
s
p
ec
ialized
ter
m
in
o
lo
g
y
an
d
s
ty
lis
tic
v
ar
iab
ilit
y
.
W
e
em
p
lo
y
th
r
ee
o
p
e
n
-
weig
h
t
L
L
Ms
—
L
lam
a
-
SEA
-
L
I
ON
-
v
3
.
5
-
8B
-
R
[
2
8
]
,
Gem
m
a2
-
9B
-
cp
t
-
s
ah
ab
a
tai
-
v1
-
b
ase
[
2
9
]
,
a
n
d
Mistra
l
-
7
B
[
3
0
]
—
f
u
s
ed
with
W
h
is
p
er
-
lar
g
e
-
v
3
u
s
in
g
th
e
GFD
f
r
am
ewo
r
k
.
T
h
e
in
v
esti
g
atio
n
s
y
s
tem
atica
lly
an
aly
ze
s
th
e
ef
f
ec
ts
o
f
th
e
f
u
s
io
n
weig
h
t
p
a
r
am
eter
(
r
)
,
in
s
tr
u
ctio
n
p
r
o
m
p
tin
g
,
an
d
d
o
m
ain
s
en
s
itiv
ity
o
n
d
ec
o
d
in
g
im
p
r
o
v
em
e
n
ts
ac
r
o
s
s
g
en
er
al
an
d
m
ed
ical
ASR
.
T
o
g
u
id
e
th
is
in
v
esti
g
atio
n
,
we
ad
d
r
ess
th
e
f
o
llo
win
g
r
esear
ch
q
u
esti
o
n
s
:
−
R
Q1
:
d
o
es
GFD
im
p
r
o
v
e
A
SR
p
er
f
o
r
m
a
n
ce
c
o
m
p
ar
e
d
t
o
a
s
tr
o
n
g
W
h
is
p
er
b
aselin
e
in
I
n
d
o
n
esian
g
en
er
al
an
d
m
ed
ical
d
o
m
ain
s
?
−
R
Q2
:
h
o
w
d
o
es
th
e
f
u
s
io
n
weig
h
t
p
ar
am
eter
(
r
)
i
n
f
lu
en
ce
r
ec
o
g
n
itio
n
ac
cu
r
ac
y
a
cr
o
s
s
d
o
m
ain
co
n
d
itio
n
s
,
an
d
wh
at
v
al
u
es y
i
eld
o
p
tim
al
tr
ad
e
-
o
f
f
s
b
etwe
en
ac
o
u
s
tic
ev
id
en
ce
a
n
d
lin
g
u
is
tic
p
r
io
r
s
?
−
R
Q3
:
h
o
w
d
o
th
e
ch
o
ice
o
f
o
p
en
-
weig
h
t
L
L
M
b
ac
k
en
d
a
n
d
f
ix
ed
in
s
tr
u
ctio
n
p
r
o
m
p
tin
g
af
f
ec
t
GFD
p
er
f
o
r
m
an
ce
ac
r
o
s
s
th
e
two
ev
alu
ated
s
p
ee
ch
d
o
m
ain
s
?
T
h
ese
r
esear
ch
q
u
esti
o
n
s
co
ll
ec
tiv
ely
ex
am
in
e
th
e
ef
f
ec
tiv
en
ess
o
f
GFD
ac
r
o
s
s
d
if
f
er
en
t
d
o
m
ain
s
,
f
u
s
io
n
s
ettin
g
s
,
an
d
L
L
M
b
ac
k
e
n
d
s
.
T
h
is
s
tu
d
y
d
o
es
n
o
t
p
r
o
p
o
s
e
a
n
ew
f
u
s
io
n
alg
o
r
ith
m
.
I
n
s
tead
,
it
ad
o
p
ts
th
e
o
r
ig
in
al
b
y
te
-
l
ev
el
GFD
f
r
am
ewo
r
k
p
r
o
p
o
s
ed
b
y
Hsu
et
a
l
.
an
d
in
v
esti
g
ates
h
o
w
e
x
ter
n
al
L
L
M
p
r
io
r
s
in
ter
ac
t
wit
h
W
h
is
p
er
-
lar
g
e
-
v3
u
n
d
er
I
n
d
o
n
esian
ASR
co
n
d
itio
n
s
.
T
h
e
co
n
tr
ib
u
tio
n
s
o
f
th
is
p
ap
er
ar
e
as f
o
llo
ws:
i)
Do
m
ain
-
f
o
c
u
s
ed
em
p
ir
ical
e
v
alu
atio
n
:
p
r
esen
t
a
co
n
tr
o
lled
em
p
ir
ical
ev
al
u
atio
n
o
f
GFD
f
o
r
I
n
d
o
n
esian
ASR
ac
r
o
s
s
two
r
ep
r
esen
tativ
e
s
p
ee
ch
co
n
d
itio
n
s
:
g
en
e
r
al
r
ea
d
s
p
ee
ch
(
FLE
UR
S
-
I
D)
an
d
co
n
tr
o
lled
,
ter
m
in
o
lo
g
y
-
d
e
n
s
e
m
ed
ical
d
i
ctatio
n
(
Har
k
it).
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
E
mp
ir
ica
l e
va
lu
a
tio
n
o
f g
en
er
a
tive
fu
s
io
n
d
ec
o
d
in
g
fo
r
I
n
d
o
n
esia
n
g
en
era
l a
n
d
med
ic
a
l
…
(
A
s
r
il J
a
r
in
)
4895
ii)
Fu
s
io
n
-
weig
h
t
ch
ar
ac
ter
izatio
n
:
s
h
o
w
th
at
th
e
f
u
s
io
n
weig
h
t
p
ar
am
eter
g
o
v
er
n
s
th
e
b
alan
ce
b
etwe
en
ac
o
u
s
tic
g
r
o
u
n
d
i
n
g
a
n
d
lin
g
u
i
s
tic
p
r
io
r
s
tr
en
g
th
,
a
n
d
th
at
it
s
o
p
tim
al
v
alu
e
is
d
o
m
ain
-
d
e
p
en
d
en
t
r
ath
er
th
an
u
n
i
v
er
s
al.
iii)
Statis
t
ical
co
n
f
ir
m
atio
n
o
f
g
ai
n
s
:
u
s
in
g
p
air
e
d
o
n
e
-
tailed
t
-
test
s
an
d
ef
f
ec
t
-
s
ize
an
al
y
s
is
,
we
co
n
f
ir
m
th
a
t
GFD
y
ield
s
s
tatis
tica
lly
s
ig
n
if
ican
t
an
d
p
r
ac
tically
m
ea
n
i
n
g
f
u
l
wo
r
d
er
r
o
r
r
ate
(
W
E
R
)
r
ed
u
ctio
n
s
r
elativ
e
to
a
s
tr
o
n
g
W
h
is
p
er
-
lar
g
e
-
v
3
b
aselin
e.
iv
)
LLM
-
p
r
io
r
an
d
p
r
o
m
p
tin
g
ab
l
atio
n
:
d
is
tin
g
u
is
h
th
e
c
o
n
tr
ib
u
tio
n
o
f
th
e
ex
ter
n
al
L
L
M
p
r
io
r
f
r
o
m
ex
p
licit
task
co
n
d
itio
n
in
g
b
y
co
m
p
a
r
in
g
wh
is
p
er
-
o
n
ly
b
aselin
e,
n
o
n
-
p
r
o
m
p
ted
GFD,
an
d
in
s
tr
u
c
tio
n
-
p
r
o
m
p
te
d
GFD
ac
r
o
s
s
th
r
ee
o
p
en
-
weig
h
t
L
L
M
b
ac
k
en
d
s
.
T
h
e
r
esu
lts
s
h
o
w
th
at
p
r
o
m
p
tin
g
p
r
o
d
u
ce
s
a
s
ec
o
n
d
ar
y
an
d
co
n
f
ig
u
r
atio
n
-
d
e
p
en
d
e
n
t e
f
f
ec
t r
elativ
e
to
th
e
co
n
tr
ib
u
tio
n
o
f
th
e
ex
ter
n
al
L
L
M
p
r
io
r
.
T
h
r
o
u
g
h
t
h
i
s
s
t
u
d
y
,
e
m
p
i
r
i
ca
l
e
v
i
d
en
c
e
i
s
p
r
o
v
i
d
e
d
t
h
at
G
F
D
c
an
i
m
p
r
o
v
e
I
n
d
o
n
es
i
a
n
A
S
R
u
n
d
e
r
t
h
e
e
v
a
l
u
a
t
ed
o
f
f
l
i
n
e
c
o
n
d
i
t
io
n
s
.
T
h
i
s
i
m
p
r
o
v
em
e
n
t
i
s
a
c
h
i
ev
ed
b
y
i
n
t
eg
r
a
t
in
g
e
x
t
er
n
a
l
L
L
M
p
r
i
o
r
s
d
u
r
in
g
d
ec
o
d
i
n
g
w
i
th
o
u
t
r
e
tr
a
in
i
n
g
e
i
th
e
r
m
o
d
e
l
.
T
h
e
r
e
s
u
l
t
s
f
u
r
t
h
er
ch
ar
a
c
t
e
r
iz
e
h
o
w
t
h
e
e
f
f
e
c
t
iv
en
e
s
s
o
f
th
i
s
in
t
e
g
r
a
t
i
o
n
d
e
p
en
d
s
o
n
s
p
e
e
ch
c
o
n
d
i
t
io
n
,
s
t
a
t
i
c
f
u
s
i
o
n
w
e
i
g
h
t,
p
r
o
m
p
t
in
g
c
o
n
f
i
g
u
r
a
t
io
n
,
an
d
L
L
M
b
ac
k
e
n
d
,
p
r
o
v
id
i
n
g
a
f
o
u
n
d
a
t
i
o
n
f
o
r
f
u
tu
r
e
d
o
m
a
in
-
s
en
s
i
t
i
v
e
d
e
c
o
d
i
n
g
s
t
r
a
t
e
g
ie
s
a
n
d
d
ep
l
o
y
m
en
t
-
o
r
i
en
t
e
d
e
v
a
lu
a
t
i
o
n
.
2.
M
E
T
H
O
DS
2
.
1
.
O
v
er
a
ll
ex
perim
ent
a
l desig
n
T
h
is
s
tu
d
y
ev
alu
ates
th
e
ef
f
ec
tiv
en
ess
o
f
GFD
[
2
7
]
f
o
r
I
n
d
o
n
esian
ASR
in
two
r
ep
r
esen
tativ
e
d
o
m
ain
s
:
g
en
er
al
an
d
m
ed
ic
al.
T
h
e
ex
p
er
im
e
n
tal
p
ip
elin
e
co
n
s
is
ts
o
f
f
o
u
r
k
ey
c
o
m
p
o
n
en
ts
:
i)
d
ataset
s
elec
tio
n
an
d
p
r
ep
a
r
atio
n
f
o
r
d
o
m
ain
-
g
en
er
al
an
d
d
o
m
ain
-
s
p
ec
if
ic
task
s
;
ii)
b
aselin
e
A
SR
d
ec
o
d
in
g
s
etu
p
u
s
in
g
W
h
is
p
er
-
lar
g
e
-
v3;
iii)
i
m
p
lem
en
tatio
n
o
f
GFD
with
s
y
s
tem
atic
v
ar
iatio
n
o
f
th
e
f
u
s
i
o
n
weig
h
t
p
a
r
am
eter
(
r
)
;
an
d
iv
)
c
o
m
p
r
e
h
en
s
iv
e
ev
alu
atio
n
an
d
s
tatis
tical
v
alid
atio
n
o
f
p
e
r
f
o
r
m
an
ce
i
m
p
r
o
v
e
m
en
ts
.
All
ex
p
er
im
en
ts
wer
e
co
n
d
u
cted
u
s
in
g
o
f
f
lin
e
u
tter
an
ce
-
le
v
el
d
ec
o
d
i
n
g
.
E
ac
h
u
tter
an
ce
was
ev
alu
ated
as
a
co
m
p
lete
in
p
u
t
in
s
tan
ce
,
an
d
th
e
s
tu
d
y
d
i
d
n
o
t
im
p
lem
en
t
o
r
ass
es
s
ch
u
n
k
-
wis
e
au
d
io
p
r
o
c
ess
in
g
,
in
cr
em
en
tal
tr
an
s
cr
ip
t d
eliv
er
y
,
en
d
p
o
in
t
d
etec
tio
n
,
o
r
laten
c
y
-
co
n
s
tr
ain
ed
s
tr
ea
m
in
g
in
f
e
r
en
ce
.
Fig
u
r
e
1
p
r
esen
ts
th
e
co
m
p
lete
o
f
f
lin
e
e
x
p
er
im
e
n
tal
wo
r
k
f
lo
w.
Fo
r
ea
ch
d
ataset,
co
m
p
lete
u
tter
an
ce
s
f
r
o
m
FLE
UR
S
-
I
D
an
d
Har
k
i
t
ar
e
d
ec
o
d
e
d
u
s
in
g
th
e
s
am
e
f
r
o
ze
n
W
h
is
p
er
-
lar
g
e
-
v
3
m
o
d
el,
eith
er
as
th
e
u
n
f
u
s
ed
b
aselin
e
o
r
th
r
o
u
g
h
GFD
wi
th
o
n
e
o
f
th
r
ee
f
r
o
ze
n
L
L
M
b
ac
k
en
d
s
.
T
h
e
GFD
ex
p
er
im
en
ts
ev
alu
ate
p
r
o
m
p
te
d
an
d
n
o
n
-
p
r
o
m
p
te
d
co
n
d
itio
n
s
ac
r
o
s
s
th
e
p
r
ed
ef
in
e
d
s
tatic
f
u
s
io
n
weig
h
ts
r
.
T
h
e
r
esu
ltin
g
tr
an
s
cr
ip
ts
ar
e
ev
alu
ated
u
s
in
g
co
r
p
u
s
-
lev
el
W
E
R
an
d
p
air
ed
u
tter
an
ce
-
l
ev
el
s
tatis
tical
an
aly
s
i
s
.
Fig
u
r
e
1
.
Of
f
lin
e
ex
p
e
r
im
en
tal
wo
r
k
f
lo
w
f
o
r
e
v
alu
atin
g
GFD
f
o
r
I
n
d
o
n
esian
ASR
2
.
2
.
Da
t
a
s
et
s
a
nd
s
ce
na
rio
s
T
wo
d
atasets
wer
e
u
s
ed
in
th
is
s
tu
d
y
:
FLE
UR
S
-
I
D
[
1
9
]
an
d
Har
k
it
m
ed
ical
[
3
1
]
.
T
ab
le
1
s
u
m
m
ar
izes
th
e
d
u
r
atio
n
,
n
u
m
b
er
o
f
u
tter
an
ce
s
,
a
n
d
u
tter
an
ce
-
len
g
th
ch
ar
ac
ter
is
tics
o
f
th
e
two
d
atasets
u
s
ed
in
th
e
ev
alu
atio
n
.
i)
FLE
UR
S
-
I
D
is
th
e
I
n
d
o
n
esian
s
u
b
s
et
o
f
th
e
m
u
ltil
in
g
u
al
FLE
UR
S
b
en
ch
m
ar
k
[
1
9
]
an
d
co
n
s
is
ts
o
f
s
h
o
r
t,
co
n
tr
o
lled
r
ea
d
-
s
p
ee
ch
u
tter
an
ce
s
r
ep
r
esen
tin
g
g
en
er
al
-
d
o
m
ain
lan
g
u
ag
e.
I
n
th
is
s
tu
d
y
,
FLE
UR
S
-
I
D
i
s
u
s
ed
as
a
c
o
n
tr
o
lled
,
clea
n
-
s
p
ee
ch
r
ef
e
r
en
ce
co
n
d
itio
n
to
ev
alu
ate
wh
eth
er
GFD
im
p
r
o
v
es
r
e
co
g
n
itio
n
wh
e
n
th
e
ac
o
u
s
tic
s
ig
n
al
an
d
s
p
ea
k
i
n
g
s
ty
le
ar
e
r
elativ
ely
co
n
s
tr
ain
ed
.
I
t
s
h
o
u
ld
th
er
ef
o
r
e
n
o
t
b
e
in
ter
p
r
eted
a
s
r
ep
r
esen
tin
g
s
p
o
n
tan
eo
u
s
c
o
n
v
er
s
atio
n
,
en
v
ir
o
n
m
e
n
tal
n
o
is
e,
f
ar
-
f
ield
s
p
ee
ch
,
o
r
o
th
er
u
n
co
n
s
tr
ain
ed
r
ea
l
-
wo
r
ld
c
o
n
d
itio
n
s
.
ii)
Har
k
it
m
ed
ical
is
a
d
o
m
ain
-
s
p
ec
if
ic
I
n
d
o
n
esian
clin
ical
s
p
ee
ch
co
r
p
u
s
d
e
v
elo
p
ed
u
n
d
e
r
th
e
Natio
n
al
R
esear
ch
an
d
I
n
n
o
v
atio
n
Ag
e
n
cy
(
B
R
I
N)
.
T
h
e
co
r
p
u
s
co
n
ta
in
s
tr
an
s
cr
ib
ed
m
ed
ical
d
ictatio
n
s
in
v
o
lv
in
g
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
15
,
No
.
5
,
Octo
b
er
20
26
:
4
8
9
3
-
490
6
4896
clin
ical
ter
m
in
o
lo
g
y
,
ab
b
r
ev
i
atio
n
s
,
an
d
m
ix
e
d
I
n
d
o
n
esian
–
E
n
g
lis
h
ex
p
r
ess
io
n
s
c
o
m
m
o
n
ly
f
o
u
n
d
in
m
ed
ical
d
o
cu
m
e
n
tatio
n
[
3
1
]
.
I
n
th
is
s
tu
d
y
,
Har
k
it
is
u
s
e
d
as
a
co
n
tr
o
lled
,
ter
m
in
o
l
o
g
y
-
d
e
n
s
e,
an
d
d
o
m
ain
-
m
is
m
atch
e
d
co
n
d
itio
n
to
ev
alu
ate
d
ec
o
d
in
g
-
tim
e
lin
g
u
is
tic
g
u
id
an
ce
.
T
h
e
co
r
p
u
s
d
o
es
n
o
t
r
ep
r
esen
t
th
e
co
m
p
lete
d
iv
e
r
s
ity
o
f
m
ed
ical
s
p
ec
ialties
,
s
p
o
n
tan
eo
u
s
clin
ician
–
p
atie
n
t
in
ter
ac
tio
n
,
en
v
ir
o
n
m
en
tal
n
o
is
e,
o
r
in
s
titu
tio
n
-
s
p
ec
if
ic
clin
ical
wo
r
k
f
lo
ws.
B
o
th
d
atasets
wer
e
s
tan
d
ar
d
ized
to
Un
ico
d
e
NFC
f
o
r
m
[
3
2
]
to
en
s
u
r
e
ca
n
o
n
ical
tex
t
eq
u
iv
alen
ce
wh
ile
p
r
eser
v
in
g
p
u
n
ctu
atio
n
an
d
c
asin
g
.
C
o
d
e
-
s
witch
ed
E
n
g
lis
h
to
k
en
s
(
e
.
g
.
,
“
C
T
s
ca
n
”
an
d
“
ch
ec
k
-
up
”
)
wer
e
r
etain
ed
to
m
ain
tai
n
lex
ical
au
th
en
ticity
.
No
p
r
o
p
r
ietar
y
d
ata
wer
e
u
s
ed
.
T
ab
le
1
.
Su
m
m
a
r
y
o
f
th
e
I
n
d
o
n
esian
s
p
ee
ch
d
atasets
u
s
ed
in
th
e
co
n
tr
o
lled
o
f
f
lin
e
ev
alu
ati
o
n
D
a
t
a
s
e
t
Ev
a
l
u
a
t
i
o
n
c
o
n
d
i
t
i
on
D
u
r
a
t
i
o
n
(
h
)
U
t
t
e
r
a
n
c
e
s
M
e
a
n
u
t
t
e
r
a
n
c
e
d
u
r
a
t
i
o
n
(
s)
M
i
n
i
m
u
m
(
s)
M
a
x
i
m
u
m
(
s)
F
LEU
R
S
-
ID
G
e
n
e
r
a
l
2
.
2
7
6
6
3
1
2
.
3
2
4
.
4
4
4
1
.
3
0
H
a
r
k
i
t
M
e
d
i
c
a
l
d
i
c
t
a
t
i
o
n
1
.
4
0
5
6
3
8
.
9
4
3
.
8
0
2
1
.
5
0
2
.
3
.
B
a
s
eline
a
uto
ma
t
ic
s
peec
h r
ec
o
g
nitio
n
m
o
del
T
h
e
b
aselin
e
s
y
s
tem
is
b
ased
o
n
W
h
is
p
er
-
lar
g
e
-
v3
[
1
0
]
,
a
m
u
ltil
in
g
u
al
tr
an
s
f
o
r
m
e
r
en
co
d
er
–
d
ec
o
d
er
m
o
d
el
tr
ain
ed
o
n
1
m
illi
o
n
h
o
u
r
s
o
f
wea
k
ly
lab
eled
au
d
io
a
n
d
4
m
illi
o
n
h
o
u
r
s
o
f
p
s
eu
d
o
-
lab
eled
au
d
io
,
an
d
u
s
ed
th
r
o
u
g
h
o
u
t
th
is
s
tu
d
y
as
th
e
f
ix
ed
ac
o
u
s
tic
m
o
d
el
.
W
h
is
p
er
r
ep
r
esen
ts
a
s
tr
o
n
g
s
tate
-
of
-
th
e
-
a
r
t
f
o
u
n
d
atio
n
f
o
r
m
u
ltil
in
g
u
al
A
SR
.
T
o
is
o
late
th
e
co
n
tr
ib
u
tio
n
o
f
GFD,
wh
is
p
er
'
s
p
ar
am
ete
r
s
wer
e
f
r
o
ze
n
an
d
all
m
o
d
if
icatio
n
s
wer
e
r
estricte
d
to
th
e
d
ec
o
d
in
g
s
tag
e.
B
o
th
th
e
b
aselin
e
an
d
GFD
s
y
s
tem
s
wer
e
ev
alu
ate
d
u
s
in
g
o
f
f
lin
e
u
tter
an
ce
-
lev
el
d
ec
o
d
in
g
.
T
h
e
b
aselin
e
u
s
ed
b
e
am
s
ea
r
ch
with
a
b
ea
m
wid
th
o
f
5
an
d
n
o
n
u
cleu
s
s
am
p
lin
g
,
wh
e
r
ea
s
th
e
ex
ter
n
a
l
L
L
M
in
GFD
was
d
ec
o
d
ed
with
a
tem
p
e
r
atu
r
e
o
f
1
.
7
a
n
d
n
o
to
p
-
p
s
am
p
lin
g
.
Acc
o
r
d
in
g
ly
,
th
e
r
ep
o
r
ted
p
er
f
o
r
m
an
ce
d
if
f
er
en
ce
s
r
ef
lect
d
ec
o
d
in
g
-
le
v
el
f
u
s
io
n
r
ath
er
th
an
m
o
d
el
r
etr
ain
in
g
o
r
d
if
f
e
r
en
ce
s
in
s
tr
ea
m
in
g
co
n
f
ig
u
r
atio
n
.
2
.
4
.
G
ener
a
t
iv
e
f
us
io
n dec
o
din
g
GFD
is
a
b
y
te
-
lev
el
g
en
er
ativ
e
s
h
allo
w
-
f
u
s
io
n
m
eth
o
d
t
h
at
in
co
r
p
o
r
ates
ex
ter
n
al
L
L
M
p
r
io
r
s
in
to
b
ea
m
-
s
ea
r
ch
d
ec
o
d
in
g
.
L
ik
e
co
n
v
e
n
tio
n
al
s
h
allo
w
f
u
s
io
n
,
GFD
co
m
b
in
es
in
d
ep
en
d
en
tly
o
b
tain
ed
ASR
an
d
lan
g
u
a
g
e
-
m
o
d
el
p
r
o
b
ab
ilit
ies
d
u
r
in
g
in
f
er
en
ce
.
I
ts
d
is
tin
g
u
is
h
in
g
f
ea
tu
r
e
is
th
e
co
n
v
er
s
io
n
o
f
ca
n
d
id
ate
-
s
eq
u
e
n
ce
p
r
o
b
ab
ilit
ies
in
to
a
s
h
ar
ed
b
y
te
-
lev
el
r
ep
r
esen
tatio
n
,
allo
win
g
th
e
A
SR
m
o
d
el
an
d
th
e
L
L
M
to
p
r
o
g
r
ess
s
y
n
ch
r
o
n
o
u
s
ly
ev
en
wh
en
th
ey
u
s
e
d
if
f
e
r
en
t
to
k
en
v
o
ca
b
u
lar
ies
[
2
7
]
.
Un
lik
e
co
ld
f
u
s
io
n
,
wh
ich
in
co
r
p
o
r
ates
a
f
i
x
ed
p
r
etr
ain
ed
lan
g
u
ag
e
m
o
d
el
d
u
r
i
n
g
s
eq
u
e
n
ce
-
to
-
s
eq
u
en
ce
ASR
tr
ain
in
g
[
2
3
]
,
a
n
d
C
L
AS,
wh
ich
jo
in
tly
lear
n
s
c
o
n
tex
tu
al
p
h
r
ase
em
b
e
d
d
in
g
s
t
o
g
eth
er
with
th
e
ASR
m
o
d
el
[
2
4
]
,
GFD
p
er
f
o
r
m
s
in
teg
r
atio
n
e
n
tire
ly
d
u
r
i
n
g
i
n
f
er
en
ce
.
C
o
n
s
eq
u
en
tly
,
GFD
en
ab
les
ex
ter
n
al
lan
g
u
ag
e
m
o
d
el
g
u
id
a
n
ce
with
o
u
t
r
eq
u
ir
in
g
r
etr
ain
in
g
eith
er
t
h
e
ASR
m
o
d
el
o
r
th
e
L
L
M.
I
n
th
e
p
r
esen
t
s
tu
d
y
,
W
h
is
p
er
-
lar
g
e
-
v
3
an
d
all
ex
ter
n
al
L
L
M
s
wer
e
k
ep
t
f
r
o
ze
n
,
an
d
in
te
g
r
atio
n
was
p
er
f
o
r
m
ed
ex
clu
s
iv
ely
th
r
o
u
g
h
GFD
d
u
r
in
g
d
ec
o
d
in
g
.
No
co
n
v
en
tio
n
al
s
h
allo
w
-
f
u
s
io
n
,
c
o
ld
-
f
u
s
io
n
,
o
r
C
L
AS
s
y
s
tem
was
im
p
lem
en
ted
in
t
h
e
p
r
esen
t
s
tu
d
y
.
I
n
s
tead
,
t
h
e
ex
p
er
im
en
tal
b
aselin
e
co
n
s
is
ts
o
f
th
e
u
n
ch
an
g
ed
W
h
is
p
er
-
lar
g
e
-
v
3
d
ec
o
d
er
,
al
lo
win
g
th
e
ef
f
ec
t
o
f
GFD
to
b
e
is
o
lated
u
n
d
er
o
th
er
wis
e
id
en
tical
d
ec
o
d
in
g
co
n
d
itio
n
s
.
T
h
is
p
r
o
b
a
b
ilis
tic
f
u
s
io
n
is
co
n
tr
o
lled
b
y
a
s
ca
la
r
f
u
s
io
n
weig
h
t
r
,
wh
ic
h
d
eter
m
in
es
th
e
r
elativ
e
in
f
lu
en
ce
o
f
t
h
e
ASR
an
d
L
L
M
d
u
r
in
g
b
ea
m
-
s
ea
r
ch
d
ec
o
d
i
n
g
.
Acc
o
r
d
in
g
l
y
,
at
ea
c
h
d
ec
o
d
in
g
s
tep
,
th
e
b
est
p
ar
tial
b
ea
m
h
y
p
o
t
h
esis
1
:
∗
is
o
b
tain
ed
b
y
m
ax
im
izin
g
th
e
f
o
llo
win
g
weig
h
ted
co
m
b
i
n
atio
n
o
f
ASR
an
d
LLM
lo
g
-
p
r
o
b
a
b
ilit
ies
as sh
o
wn
in
(
1
)
.
1
:
∗
=
{
1
:
}
[
(
1
−
)
l
og
(
1
:
;
)
+
(
)
l
og
(
1
:
−
;
)
]
(
1
)
W
h
er
e
1
:
d
en
o
tes
th
e
p
ar
tial
b
ea
m
s
eq
u
en
ce
u
p
to
tim
e
s
tep
.
AS
R
(
1
:
;
)
is
th
e
p
r
o
b
ab
ilit
y
ass
ig
n
ed
b
y
th
e
tr
an
s
cr
ip
tio
n
m
o
d
el
(
ASR
tr
an
s
cr
ip
tio
n
m
o
d
el)
p
ar
am
eter
ized
b
y
.
L
L
M
(
1
:
−
;
)
is
th
e
au
to
r
eg
r
ess
iv
e
lan
g
u
ag
e
-
m
o
d
e
l
p
r
o
b
ab
ilit
y
o
f
th
e
ex
ter
n
al
L
L
M
—
p
ar
am
eter
ized
b
y
—
f
o
r
co
n
tin
u
in
g
th
e
h
y
p
o
th
esis
s
eq
u
en
ce
,
co
n
d
itio
n
ed
o
n
ly
o
n
th
e
p
r
ev
io
u
s
ly
g
e
n
er
ated
−
tex
t
to
k
en
s
.
∈
[
0
,
1
]
is
th
e
f
u
s
io
n
weig
h
t
co
n
tr
o
llin
g
th
e
b
ala
n
ce
b
etwe
en
th
e
ac
o
u
s
tic
m
o
d
el
an
d
th
e
L
L
M
co
n
tr
ib
u
tio
n
.
k
is
a
d
y
n
am
ic
s
h
if
tin
g
v
alu
e
d
eter
m
in
e
d
s
ep
ar
ately
f
o
r
ea
ch
b
ea
m
h
y
p
o
th
esis
.
Fo
llo
win
g
th
e
o
r
ig
in
al
GFD
im
p
l
em
en
tatio
n
,
k
is
s
et
to
th
e
b
y
te
len
g
th
o
f
th
e
last
to
k
en
p
r
o
p
o
s
ed
b
y
th
e
tex
t
-
r
ec
o
g
n
itio
n
m
o
d
el.
T
h
is
s
elec
tio
n
r
estricts th
e
n
u
m
b
er
o
f
s
eq
u
en
ce
s
r
esco
r
ed
b
y
th
e
L
L
M
to
ap
p
r
o
x
im
ately
th
e
b
e
am
s
ize
wh
ile
r
etain
in
g
th
e
m
ax
im
u
m
av
ailab
le
p
r
e
-
ex
p
an
s
io
n
co
n
tex
t.
I
n
th
e
p
r
esen
t
s
tu
d
y
,
r
is
im
p
l
em
en
ted
as
a
s
tatic
s
ca
lar
in
t
er
p
o
latio
n
co
ef
f
icien
t.
Fo
r
ea
ch
d
ataset,
L
L
M
b
ac
k
en
d
,
an
d
p
r
o
m
p
tin
g
co
n
d
itio
n
,
a
s
in
g
le
r
v
al
u
e
is
ap
p
lied
to
all
u
tter
a
n
ce
s
an
d
au
to
r
eg
r
ess
iv
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
E
mp
ir
ica
l e
va
lu
a
tio
n
o
f g
en
er
a
tive
fu
s
io
n
d
ec
o
d
in
g
fo
r
I
n
d
o
n
esia
n
g
en
era
l a
n
d
med
ic
a
l
…
(
A
s
r
il J
a
r
in
)
4897
d
ec
o
d
in
g
s
tep
s
.
C
an
d
id
ate
v
alu
es
ar
e
ev
alu
ated
u
s
in
g
th
e
d
e
v
elo
p
m
en
t
s
p
lit,
an
d
th
e
s
elec
ted
co
n
f
ig
u
r
atio
n
is
th
en
ap
p
lied
to
th
e
h
eld
-
o
u
t te
s
t set.
No
u
tter
an
ce
-
lev
el,
to
k
e
n
-
lev
el,
co
n
f
i
d
en
ce
-
c
o
n
d
itio
n
e
d
,
o
r
au
to
m
atica
lly
d
o
m
ain
-
c
o
n
d
itio
n
e
d
a
d
ju
s
tm
en
t
o
f
r
is
p
e
r
f
o
r
m
ed
.
Acc
o
r
d
in
g
ly
,
r
s
h
o
u
l
d
b
e
in
ter
p
r
eted
as
an
em
p
ir
ically
s
elec
ted
co
ef
f
icien
t
f
o
r
a
p
ar
ti
cu
lar
ex
p
er
im
en
tal
c
o
n
f
ig
u
r
ati
o
n
r
at
h
er
th
a
n
as
a
d
ir
ec
t
p
er
c
en
tag
e
m
ea
s
u
r
e
o
f
th
e
lin
g
u
is
tic
co
n
tr
ib
u
tio
n
.
T
h
e
d
y
n
am
ic
s
h
if
tin
g
v
alu
e
k
an
d
th
e
co
r
r
esp
o
n
d
in
g
d
elay
e
d
co
r
r
ec
tiv
e
-
f
ee
d
b
ac
k
m
ec
h
an
is
m
f
o
llo
w
th
e
o
r
ig
in
al
GFD
alg
o
r
ith
m
a
n
d
im
p
lem
en
tatio
n
with
o
u
t
m
o
d
if
icatio
n
o
r
in
d
e
p
en
d
e
n
t
tu
n
in
g
.
B
ec
au
s
e
th
e
len
g
th
o
f
t
h
e
m
o
s
t
r
ec
en
t
r
ec
o
g
n
itio
n
to
k
e
n
m
ay
d
if
f
er
ac
r
o
s
s
b
ea
m
h
y
p
o
t
h
eses
,
k
is
d
y
n
a
m
ically
d
eter
m
in
ed
r
ath
er
th
an
f
ix
e
d
g
lo
b
ally
.
T
h
is
b
ea
m
-
lev
el
f
o
r
m
u
latio
n
(
1
)
is
f
u
lly
eq
u
iv
alen
t
to
in
cr
em
en
tal
to
k
en
-
lev
el
in
ter
p
o
latio
n
o
f
ASR
an
d
L
L
M
lo
g
-
p
r
o
b
ab
ilit
ies
ex
ec
u
ted
d
u
r
in
g
a
u
to
r
eg
r
ess
iv
e
d
ec
o
d
in
g
;
th
e
b
y
te
-
lev
e
l
r
ep
r
esen
tatio
n
en
s
u
r
es
b
o
t
h
m
o
d
els
o
p
er
ate
o
v
er
th
e
s
am
e
t
o
k
en
s
p
ac
e,
en
ab
lin
g
tr
ain
in
g
-
f
r
ee
f
u
s
io
n
with
o
u
t
em
b
ed
d
in
g
p
r
o
jectio
n
o
r
v
o
c
ab
u
lar
y
r
em
ap
p
i
n
g
.
T
h
is
f
o
r
m
u
latio
n
allo
ws
d
o
m
ain
p
r
io
r
s
to
b
e
in
co
r
p
o
r
ated
th
r
o
u
g
h
p
r
o
b
ab
ilit
y
-
s
p
ac
e
g
u
id
an
ce
r
ath
er
th
an
p
ar
am
ete
r
u
p
d
ates,
m
ak
in
g
GFD
ap
p
licab
le
ac
r
o
s
s
th
e
ev
alu
ated
g
e
n
er
al
an
d
s
p
ec
ialized
I
n
d
o
n
esian
ASR
co
n
d
itio
n
s
with
o
u
t r
etr
ain
in
g
.
Alth
o
u
g
h
GFD
co
m
b
in
es
ASR
an
d
L
L
M
p
r
o
b
ab
ilit
ies
in
cr
em
en
tally
at
ea
c
h
au
t
o
r
eg
r
ess
iv
e
d
ec
o
d
in
g
s
tep
,
th
is
s
tep
wis
e
o
p
er
atio
n
s
h
o
u
ld
n
o
t
b
e
in
te
r
p
r
eted
as
a
s
tr
ea
m
in
g
ASR
im
p
lem
en
tatio
n
.
I
n
t
h
e
p
r
esen
t
ex
p
er
im
en
ts
,
co
m
p
le
te
u
tter
an
ce
s
wer
e
d
ec
o
d
ed
o
f
f
lin
e,
an
d
n
o
ch
u
n
k
-
le
v
el
au
d
io
p
r
o
ce
s
s
in
g
,
p
ar
tial
-
h
y
p
o
th
esis
em
is
s
io
n
,
en
d
p
o
in
ti
n
g
,
o
r
laten
cy
m
ea
s
u
r
em
en
t
was
p
er
f
o
r
m
ed
.
T
h
er
ef
o
r
e,
th
e
r
ep
o
r
ted
r
esu
lts
ch
ar
ac
ter
ize
o
f
f
lin
e
d
ec
o
d
in
g
p
er
f
o
r
m
an
ce
r
ath
er
th
a
n
r
ea
l
-
tim
e
ASR
b
eh
av
io
r
.
T
h
e
d
ec
o
d
in
g
f
r
am
ewo
r
k
n
ev
er
th
eless
s
u
p
p
o
r
ts
in
s
tr
u
ct
io
n
-
co
n
d
itio
n
ed
an
d
d
o
m
ai
n
-
s
en
s
itiv
e
lin
g
u
is
tic
g
u
id
an
ce
with
o
u
t
r
etr
ain
in
g
.
T
o
c
h
ar
ac
ter
ize
t
h
e
s
en
s
itiv
ity
o
f
th
is
g
u
id
a
n
ce
,
we
ev
al
u
ated
r
∈
{0
.
1
0
,
0
.
2
5
,
0
.
5
0
,
0
.
7
5
,
0
.
8
0
,
0
.
8
5
,
0
.
9
0
}.
T
h
e
o
p
tim
al
co
ef
f
icien
t
was
n
o
t
u
n
if
o
r
m
a
cr
o
s
s
th
e
ev
alu
ated
co
n
d
itio
n
s
:
FLE
UR
S
-
I
D
ac
h
i
ev
ed
its
b
est
r
esu
lt
at
r
=
0
.
7
5
,
wh
er
ea
s
Har
k
it
ac
h
ie
v
e
d
its
b
est
r
esu
lt
at
r
=
0
.
5
0
.
T
h
ese
v
alu
es
s
h
o
u
ld
b
e
in
ter
p
r
eted
as
co
n
f
ig
u
r
atio
n
-
s
p
ec
if
ic
em
p
ir
ical
o
p
tim
a
r
a
th
er
th
an
u
n
iv
er
s
al
m
ea
s
u
r
es
o
f
th
e
r
eq
u
ir
e
d
L
L
M
co
n
tr
i
b
u
tio
n
.
T
h
e
d
if
f
er
in
g
o
p
tim
a
m
o
tiv
ate
f
u
tu
r
e
ad
ap
tiv
e
f
u
s
io
n
m
ec
h
an
is
m
s
th
at
co
n
d
itio
n
r
o
n
r
ec
o
g
n
itio
n
co
n
f
id
e
n
ce
,
ASR
–
L
L
M
ag
r
ee
m
en
t,
o
r
d
o
m
ain
-
r
elate
d
cu
es.
2
.
5
.
L
a
rg
e
la
ng
ua
g
e
m
o
del
i
nte
g
ra
t
io
n a
nd
a
bla
t
io
n c
o
n
ditio
ns
T
h
r
ee
f
r
o
ze
n
o
p
e
n
-
weig
h
t
L
L
Ms
wer
e
in
teg
r
ated
in
to
t
h
e
GFD
d
ec
o
d
in
g
p
i
p
elin
e:
Mistra
l
-
7
B
,
L
lam
a
-
SEA
-
L
I
ON
-
v
3
.
5
-
8B
-
R
(
SEA
-
L
I
ON)
,
an
d
Gem
m
a2
-
9B
-
cp
t
-
s
ah
ab
atai
-
v1
-
b
ase
(
SAI
)
.
T
h
ese
m
o
d
els
r
e
p
r
e
s
e
n
t
g
e
n
e
r
al
m
u
lt
i
l
i
n
g
u
a
l
,
S
o
u
t
h
e
a
s
t
A
s
i
a
n
r
e
g
i
o
n
a
l
,
a
n
d
I
n
d
o
n
e
s
i
a
n
-
a
d
a
p
t
e
d
l
a
n
g
u
a
g
e
-
m
o
d
e
l
b
a
c
k
g
r
o
u
n
d
s
,
r
esp
ec
tiv
ely
.
T
h
e
co
m
p
ar
is
o
n
is
em
p
ir
ical;
th
e
p
r
esen
t
s
tu
d
y
d
o
es
n
o
t
in
d
e
p
en
d
e
n
tly
is
o
late
th
e
ef
f
ec
ts
o
f
p
r
e
-
tr
ain
in
g
d
ata
co
m
p
o
s
itio
n
,
to
k
en
izer
d
esig
n
,
in
s
tr
u
ctio
n
tu
n
in
g
,
o
r
m
o
d
el
ar
c
h
itectu
r
e.
T
o
d
is
tin
g
u
is
h
th
e
ex
ter
n
al
L
L
M
p
r
io
r
f
r
o
m
e
x
p
li
cit
task
co
n
d
itio
n
in
g
,
th
e
e
v
alu
atio
n
u
s
es th
r
ee
d
ec
o
d
in
g
co
n
d
itio
n
s
:
i)
W
h
is
p
er
-
o
n
ly
b
aselin
e
(
r
=
0
)
:
th
e
ex
ter
n
al
L
L
M
s
co
r
e
is
ex
clu
d
ed
f
r
o
m
th
e
in
te
r
p
o
latio
n
,
an
d
d
ec
o
d
in
g
is
p
er
f
o
r
m
e
d
u
s
in
g
f
r
o
ze
n
W
h
is
p
er
-
lar
g
e
-
v
3
o
n
ly
.
ii)
N
o
n
-
p
r
o
m
p
t
e
d
G
F
D
(
r
>
0
)
:
th
e
e
x
t
e
r
n
a
l
L
L
M
r
e
ce
i
v
e
s
t
h
e
p
a
r
t
i
a
l
A
SR
h
y
p
o
t
h
es
is
wi
t
h
o
u
t
a
n
y
t
as
k
-
o
r
d
o
m
a
i
n
-
s
p
e
c
i
f
i
c
i
n
s
t
r
u
ct
i
o
n
.
T
h
i
s
c
o
n
d
i
ti
o
n
e
v
a
l
u
a
t
es
t
h
e
c
o
n
t
r
i
b
u
t
i
o
n
o
f
t
h
e
p
r
e
t
r
a
i
n
e
d
L
L
M
p
r
i
o
r
a
l
o
n
e
.
iii)
I
n
s
tr
u
ctio
n
-
p
r
o
m
p
ted
GFD
(
r
>
0
)
:
th
e
s
am
e
L
L
M
b
ac
k
en
d
an
d
GFD
p
r
o
ce
d
u
r
e
ar
e
u
s
e
d
,
b
u
t
a
f
ix
ed
d
ataset
-
s
p
ec
if
ic
in
s
tr
u
ctio
n
is
p
r
ep
en
d
ed
to
t
h
e
p
ar
tial h
y
p
o
t
h
esis
.
T
h
e
s
elec
ted
p
r
o
m
p
ts
ar
e:
−
FLE
UR
S
-
I
D:
“
B
eriku
t in
i a
d
a
la
h
tr
a
n
s
krip
s
i k
a
lima
t u
ca
p
a
n
b
a
h
a
s
a
I
n
d
o
n
esia
:
”
E
n
g
lis
h
tr
an
s
latio
n
:
“
B
elo
w
is
th
e
tr
an
s
cr
ip
tio
n
o
f
an
I
n
d
o
n
es
ian
s
p
o
k
en
s
en
te
n
ce
:
”
−
Har
k
it:
“
A
n
d
a
a
d
a
la
h
LL
M
d
a
la
m
ko
n
teks
med
is
.
P
erta
h
a
n
ka
n
is
tila
h
med
is
,
fa
r
ma
s
i,
a
n
a
to
mi,
d
a
n
d
ia
g
n
o
s
is
a
p
a
a
d
a
n
ya
”
.
E
n
g
lis
h
tr
an
s
latio
n
:
“
Yo
u
a
r
e
an
L
L
M
o
p
er
atin
g
in
a
m
ed
ical
co
n
tex
t.
Pre
s
er
v
e
m
ed
ical,
p
h
ar
m
ac
eu
tical,
a
n
ato
m
ical,
a
n
d
d
iag
n
o
s
tic
ter
m
s
ex
ac
tly
as
th
ey
ar
e
”
.
T
h
e
p
r
o
m
p
ted
an
d
n
o
n
-
p
r
o
m
p
ted
co
n
d
itio
n
s
wer
e
ev
alu
ated
u
s
in
g
th
e
s
am
e
ca
n
d
id
ate
v
al
u
es
o
f
th
e
s
tatic
f
u
s
io
n
weig
h
t
r
,
en
a
b
lin
g
d
ir
ec
t
s
am
e
-
r
c
o
m
p
ar
is
o
n
s
o
f
p
r
o
m
p
t
ef
f
ec
ts
.
T
h
e
f
in
al
p
r
o
m
p
t
tem
p
late
an
d
r
v
alu
e
f
o
r
ea
ch
b
ac
k
e
n
d
–
c
o
n
d
itio
n
p
air
wer
e
s
elec
ted
u
s
in
g
th
e
d
e
v
elo
p
m
e
n
t
s
p
lit
an
d
s
u
b
s
eq
u
en
tly
f
ix
e
d
f
o
r
ev
alu
atio
n
o
n
t
h
e
h
eld
-
o
u
t te
s
t set.
No
u
tter
an
ce
-
s
p
ec
if
ic
p
r
o
m
p
t selec
tio
n
o
r
test
-
s
et
p
r
o
m
p
t o
p
tim
izatio
n
was
p
er
f
o
r
m
ed
.
C
o
r
p
u
s
-
lev
el
s
en
s
itiv
ity
to
r
an
d
u
tter
an
ce
-
l
ev
el
d
is
tr
ib
u
tio
n
s
o
f
th
e
d
e
v
elo
p
m
en
t
-
s
elec
ted
co
n
f
ig
u
r
atio
n
s
ar
e
r
e
p
o
r
te
d
in
s
ec
tio
n
s
3
.
2
an
d
3
.
4
,
r
esp
ec
tiv
ely
.
2
.
6
.
E
v
a
lua
t
i
o
n
m
et
rics a
nd
s
t
a
t
is
t
ica
l a
na
ly
s
is
ASR
ac
cu
r
ac
y
was
ass
ess
ed
u
s
in
g
W
E
R
as
th
e
p
r
im
a
r
y
m
etr
ic,
co
m
p
u
ted
as
(
2
)
.
W
h
er
e
is
th
e
n
u
m
b
er
o
f
s
u
b
s
titu
tio
n
s
,
d
eletio
n
s
,
in
s
er
tio
n
s
,
an
d
N
t
h
e
to
ta
l n
u
m
b
er
o
f
r
ef
er
en
ce
wo
r
d
s
.
=
+
+
(
2
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
15
,
No
.
5
,
Octo
b
er
20
26
:
4
8
9
3
-
490
6
4898
Statis
t
ical
s
ig
n
if
ican
ce
was
ass
es
s
ed
u
s
in
g
p
air
ed
o
n
e
-
ta
iled
t
-
test
s
o
n
u
tter
an
ce
-
lev
el
W
E
R
d
if
f
er
en
ce
s
b
etwe
en
th
e
u
n
f
u
s
ed
W
h
is
p
er
-
lar
g
e
-
v
3
b
aselin
e
an
d
th
e
b
est
-
p
er
f
o
r
m
in
g
GFD
co
n
f
i
g
u
r
atio
n
f
o
r
ea
ch
ev
alu
ated
d
ataset.
T
h
e
p
air
ed
d
esig
n
co
m
p
a
r
es
th
e
two
d
e
co
d
in
g
s
y
s
tem
s
o
n
id
e
n
tical
u
tter
an
ce
s
,
th
er
eb
y
co
n
tr
o
llin
g
f
o
r
u
tter
a
n
ce
-
s
p
ec
if
ic
v
ar
iab
ilit
y
.
Statis
tical
test
in
g
was
p
er
f
o
r
m
ed
o
n
ly
af
ter
th
e
f
in
al
co
n
f
ig
u
r
atio
n
h
ad
b
ee
n
s
elec
ted
o
n
t
h
e
d
ev
el
o
p
m
en
t
p
ar
titi
o
n
,
en
s
u
r
i
n
g
th
at
t
h
e
h
eld
-
o
u
t
test
d
ata
wer
e
u
s
ed
ex
clu
s
iv
ely
f
o
r
in
d
ep
e
n
d
en
t
ev
alu
atio
n
.
I
n
ad
d
itio
n
t
o
p
-
v
alu
es,
C
o
h
en
'
s
was
co
m
p
u
t
ed
to
q
u
a
n
tify
th
e
p
r
ac
tical
m
ag
n
itu
d
e
o
f
t
h
e
o
b
s
er
v
ed
im
p
r
o
v
e
m
en
ts
.
2
.
7
.
Repro
du
cibili
t
y
a
nd
im
plem
ent
a
t
io
n
All
ex
p
er
im
en
ts
wer
e
im
p
l
em
en
ted
in
Py
th
o
n
3
.
1
1
u
s
in
g
Py
T
o
r
ch
2
.
2
an
d
Hu
g
g
in
g
Face
T
r
an
s
f
o
r
m
e
r
s
4
.
4
3
a
n
d
wer
e
ex
ec
u
ted
o
n
an
NVI
DI
A
A1
0
0
GPU
with
8
0
GB
o
f
m
em
o
r
y
.
T
h
e
e
v
alu
atio
n
u
s
ed
o
f
f
lin
e
u
tter
an
ce
-
lev
el
d
ec
o
d
in
g
a
n
d
f
o
cu
s
ed
o
n
tr
an
s
cr
ip
tio
n
ac
cu
r
ac
y
.
R
ea
l
-
tim
e
f
ac
to
r
,
f
ir
s
t
-
to
k
en
laten
cy
,
en
d
-
to
-
e
n
d
laten
cy
,
p
ea
k
m
em
o
r
y
u
s
ag
e,
co
m
p
u
tatio
n
al
th
r
o
u
g
h
p
u
t,
an
d
p
a
r
tial
-
tr
an
s
cr
ip
t
s
tab
ilit
y
wer
e
n
o
t
m
ea
s
u
r
e
d
.
C
o
n
s
eq
u
en
tly
,
th
e
p
r
esen
t
r
esu
lts
s
h
o
u
ld
n
o
t
b
e
in
ter
p
r
eted
as
ev
id
e
n
ce
o
f
r
ea
l
-
tim
e
o
r
s
tr
ea
m
in
g
r
ea
d
in
ess
.
T
o
f
ac
ilit
ate
ex
p
er
im
en
tal
r
ep
r
o
d
u
cib
ilit
y
,
th
is
s
tu
d
y
f
o
llo
ws
th
e
o
r
ig
in
al
o
p
en
-
s
o
u
r
ce
im
p
lem
en
tatio
n
o
f
GFD
r
ele
ased
b
y
Hsu
et
a
l.
[
2
7
]
wh
il
e
u
s
in
g
th
e
p
u
b
licly
av
ailab
l
e
W
h
is
p
er
-
lar
g
e
-
v3
ch
ec
k
p
o
i
n
t.
T
h
e
ev
alu
ated
FL
E
UR
S
-
I
D
an
d
Har
k
it
m
ed
ical
-
d
ictatio
n
d
atasets
wer
e
o
b
tain
ed
f
r
o
m
th
e
o
f
f
icial
r
ep
o
s
ito
r
ies
lis
ted
in
th
e
d
ata
av
ailab
ilit
y
s
tatem
en
t,
with
t
h
e
ev
alu
atio
n
s
u
b
s
ets
s
elec
ted
ac
co
r
d
in
g
to
th
e
cr
iter
ia
d
escr
ib
ed
in
s
ec
tio
n
2
.
2
.
All
m
o
d
el
p
ar
am
eter
s
r
em
ai
n
ed
f
r
o
ze
n
th
r
o
u
g
h
o
u
t
th
e
ex
p
er
im
en
ts
,
an
d
o
n
ly
d
ec
o
d
in
g
-
tim
e
co
n
f
ig
u
r
atio
n
s
—
in
clu
d
in
g
th
e
s
tatic
f
u
s
io
n
weig
h
t,
p
r
o
m
p
tin
g
co
n
d
itio
n
,
an
d
o
th
e
r
d
ec
o
d
in
g
p
ar
am
eter
s
d
escr
ib
ed
i
n
s
u
b
s
e
ctio
n
s
2
.
2
–
2
.
6
—
wer
e
v
a
r
ied
.
3.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
3
.
1
.
O
v
er
a
ll
perf
o
r
m
a
nce
a
c
ro
s
s
da
t
a
s
et
s
T
ab
le
2
r
ep
o
r
ts
co
r
p
u
s
-
lev
el
W
E
R
f
o
r
th
e
u
n
f
u
s
ed
W
h
is
p
er
-
lar
g
e
-
v
3
b
aselin
e
an
d
t
h
e
b
est
GFD
co
n
f
ig
u
r
atio
n
s
elec
ted
f
o
r
ea
c
h
d
ataset.
W
E
R
is
r
ep
o
r
ted
as
a
p
er
ce
n
tag
e,
Δ
W
E
R
d
en
o
tes
th
e
ab
s
o
lu
te
W
E
R
r
ed
u
ctio
n
in
p
er
ce
n
ta
g
e
p
o
in
t
s
(
p
p
)
,
an
d
W
E
R
R
d
en
o
tes
th
e
r
elativ
e
W
E
R
r
ed
u
ctio
n
.
L
o
wer
W
E
R
in
d
icate
s
b
etter
r
ec
o
g
n
itio
n
p
e
r
f
o
r
m
an
c
e.
T
ab
le
2
.
C
o
r
p
u
s
-
lev
el
W
E
R
o
f
W
h
is
p
er
-
lar
g
e
-
v
3
an
d
t
h
e
b
es
t G
FD c
o
n
f
ig
u
r
atio
n
s
o
n
th
e
h
eld
-
o
u
t te
s
t sets
D
a
t
a
s
e
t
W
h
i
s
p
e
r
-
l
a
r
g
e
-
v
3
W
ER
(
%)
B
e
st
G
F
D
W
ER
(
%)
ΔW
ER
(
p
p
)
W
ER
R
(
%)
B
e
st
G
F
D
c
o
n
f
i
g
u
r
a
t
i
o
n
F
LEU
R
S
-
I
D
(
g
e
n
e
r
a
l)
5
.
9
1
4
.
1
7
1
.
7
4
2
9
.
44
S
EA
-
LI
O
N
,
p
r
o
mp
t
e
d
,
r
=
0
.
7
5
H
a
r
k
i
t
(
me
d
i
c
a
l)
1
5
.
8
9
1
2
.
3
0
3
.
5
9
2
2
.
5
9
S
A
I
,
p
r
o
mp
t
e
d
,
r
=
0
.
5
0
On
FLE
UR
S
-
I
D,
th
e
lo
west
co
r
p
u
s
-
lev
el
W
E
R
is
o
b
tain
ed
b
y
SEA
-
L
I
ON
u
n
d
er
th
e
p
r
o
m
p
te
d
co
n
f
ig
u
r
atio
n
with
a
s
tatic
f
u
s
io
n
weig
h
t
o
f
r
=
0
.
7
5
,
r
e
d
u
cin
g
W
E
R
f
r
o
m
5
.
9
1
%
to
4
.
1
7
%
(
a
n
a
b
s
o
lu
te
r
ed
u
ctio
n
o
f
1
.
7
4
pp
an
d
a
r
elativ
e
r
ed
u
ctio
n
o
f
2
9
.
44
%).
On
Har
k
it,
th
e
lo
west c
o
r
p
u
s
-
lev
e
l WER
i
s
o
b
tain
ed
b
y
SAI
u
n
d
er
th
e
p
r
o
m
p
te
d
co
n
f
i
g
u
r
atio
n
with
r
=
0
.
5
0
,
r
e
d
u
cin
g
W
E
R
f
r
o
m
1
5
.
8
9
%
to
1
2
.
3
0
%
(
an
ab
s
o
lu
te
r
ed
u
ctio
n
o
f
3
.
5
9
pp
an
d
a
r
elativ
e
r
e
d
u
ctio
n
o
f
2
2
.
5
9
%).
T
h
ese
r
e
s
u
lts
id
en
tify
th
e
b
est
-
p
er
f
o
r
m
in
g
co
n
f
ig
u
r
atio
n
s
u
n
d
e
r
th
e
ev
al
u
ated
e
x
p
er
im
en
tal
s
ettin
g
s
an
d
e
s
tab
lis
h
co
n
s
is
ten
t
im
p
r
o
v
em
e
n
ts
o
v
er
t
h
e
u
n
f
u
s
e
d
W
h
is
p
er
-
lar
g
e
-
v
3
b
aselin
e.
Ov
er
all,
th
e
r
esu
lts
in
d
icate
th
at
GFD
g
en
er
alize
s
ac
r
o
s
s
b
o
th
ev
alu
ated
co
n
d
itio
n
s
.
Sti
ll,
it
s
p
er
f
o
r
m
an
ce
d
ep
e
n
d
s
o
n
th
e
in
ter
ac
tio
n
am
o
n
g
th
e
L
L
M
b
ac
k
en
d
,
th
e
p
r
o
m
p
tin
g
co
n
d
itio
n
,
th
e
s
p
ee
ch
d
o
m
ain
,
an
d
th
e
s
tatic
f
u
s
io
n
weig
h
t.
No
s
in
g
le
b
ac
k
en
d
o
r
p
r
o
m
p
tin
g
co
n
f
ig
u
r
atio
n
is
u
n
if
o
r
m
l
y
o
p
tim
a
l
ac
r
o
s
s
th
e
two
d
atasets
,
h
ig
h
lig
h
tin
g
th
e
im
p
o
r
tan
ce
o
f
co
n
f
i
g
u
r
atio
n
-
s
p
ec
if
ic
ev
alu
atio
n
r
a
th
er
th
an
ap
p
ly
in
g
a
s
in
g
le
d
ec
o
d
in
g
s
tr
ateg
y
u
n
i
v
er
s
ally
.
3
.
2
.
Sens
it
iv
it
y
t
o
f
us
io
n we
ig
ht
(
r)
Fig
u
r
e
2
s
h
o
ws
co
r
p
u
s
-
lev
el
W
E
R
ac
r
o
s
s
th
e
ev
alu
ated
s
ta
tic
f
u
s
io
n
weig
h
ts
f
o
r
SEA
-
L
I
ON,
SAI
,
an
d
Mistra
l
u
n
d
er
p
r
o
m
p
ted
a
n
d
n
o
n
-
p
r
o
m
p
ted
d
ec
o
d
in
g
.
T
h
e
s
am
e
m
ar
k
er
id
e
n
tifie
s
ea
c
h
L
L
M
b
ac
k
e
n
d
in
b
o
th
p
an
els,
wh
ile
s
o
lid
an
d
d
ash
ed
lin
es
d
e
n
o
te
p
r
o
m
p
te
d
an
d
n
o
n
-
p
r
o
m
p
ted
d
e
co
d
in
g
,
r
esp
ec
tiv
ely
.
T
h
e
h
o
r
izo
n
tal
d
o
tted
lin
e
r
ep
r
es
en
ts
th
e
u
n
f
u
s
ed
W
h
is
p
er
-
lar
g
e
-
v
3
b
aselin
e
(
r
=
0
)
.
All
co
n
f
ig
u
r
atio
n
s
ar
e
co
m
p
ar
ed
u
s
in
g
id
en
tical
ca
n
d
id
ate
v
alu
es
o
f
r
,
en
ab
lin
g
d
ir
ec
t
ass
es
s
m
en
t
o
f
th
e
co
n
tr
ib
u
tio
n
o
f
ex
p
licit
in
s
tr
u
ctio
n
p
r
o
m
p
tin
g
b
ey
o
n
d
th
e
ex
ter
n
al
L
L
M
p
r
io
r
.
L
o
we
r
W
E
R
in
d
icate
s
b
etter
r
ec
o
g
n
itio
n
p
er
f
o
r
m
an
ce
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
E
mp
ir
ica
l e
va
lu
a
tio
n
o
f g
en
er
a
tive
fu
s
io
n
d
ec
o
d
in
g
fo
r
I
n
d
o
n
esia
n
g
en
era
l a
n
d
med
ic
a
l
…
(
A
s
r
il J
a
r
in
)
4899
Fig
u
r
e
2
(
a)
s
h
o
ws
FLE
UR
S
-
I
D,
th
e
lo
west
W
E
R
i
s
o
b
tain
ed
with
p
r
o
m
p
ted
SEA
-
L
I
ON
at
r
=
0
.
7
5
.
Acr
o
s
s
th
e
ev
alu
ated
co
n
f
ig
u
r
atio
n
s
,
im
p
r
o
v
em
e
n
ts
ar
e
g
en
er
ally
m
ain
tain
e
d
f
r
o
m
t
h
e
lo
w
-
to
-
m
o
d
er
ate
f
u
s
io
n
s
ettin
g
s
th
r
o
u
g
h
r
=
0
.
8
0
,
wh
er
ea
s
p
e
r
f
o
r
m
an
ce
d
eter
io
r
ates
as
r
ap
p
r
o
ac
h
es
0
.
9
0
.
T
h
is
p
atter
n
in
d
icate
s
th
at
GFD
b
e
n
ef
its
f
r
o
m
s
u
b
s
tan
tial
L
L
M
g
u
id
an
ce
in
t
h
e
clea
n
r
ea
d
-
s
p
ee
ch
co
n
d
itio
n
,
b
u
t
ex
ce
s
s
iv
e
weig
h
tin
g
o
f
th
e
L
L
M
s
co
r
e
e
v
en
tu
ally
wea
k
e
n
s
co
n
s
is
ten
cy
with
th
e
r
ec
o
g
n
itio
n
e
v
id
en
c
e.
Fig
u
r
e
2
(
b
)
s
h
o
ws
Har
k
it
m
e
d
ical
d
ictatio
n
,
th
e
o
p
tim
u
m
s
h
if
ts
to
r
=
0
.
5
0
,
with
p
r
o
m
p
ted
SAI
p
r
o
d
u
cin
g
th
e
lo
west
co
r
p
u
s
-
l
ev
el
W
E
R
.
T
h
is
r
esu
lt
in
d
icat
es
th
at
ter
m
in
o
lo
g
y
-
d
en
s
e
s
p
e
ec
h
b
en
e
f
its
f
r
o
m
L
L
M
g
u
id
a
n
ce
,
b
u
t
th
e
m
o
s
t
ef
f
ec
tiv
e
co
n
f
ig
u
r
atio
n
r
etain
s
a
m
o
r
e
b
alan
ce
d
in
ter
p
o
latio
n
b
etwe
en
th
e
ASR
an
d
L
L
M
s
co
r
es
th
an
th
e
FL
E
UR
S
-
I
D
o
p
tim
u
m
.
Per
f
o
r
m
an
ce
d
eter
io
r
atio
n
b
e
y
o
n
d
th
e
o
p
tim
u
m
s
u
g
g
ests
th
at
ex
ce
s
s
iv
e
lin
g
u
is
tic
b
ias
m
ay
p
r
o
d
u
ce
o
v
e
r
-
co
r
r
ec
tio
n
wh
en
th
e
L
L
M
p
r
io
r
co
n
f
licts
with
th
e
r
ec
o
g
n
itio
n
ev
id
en
ce
.
I
n
s
tr
u
ctio
n
p
r
o
m
p
t
in
g
p
r
o
v
id
es
ad
d
itio
n
al
d
o
m
ain
co
n
d
itio
n
in
g
f
o
r
s
ev
er
al
b
ac
k
en
d
–
weig
h
t
co
m
b
in
atio
n
s
,
b
u
t
its
b
en
ef
it
is
n
o
t
u
n
if
o
r
m
ac
r
o
s
s
all
co
n
f
ig
u
r
atio
n
s
an
d
d
o
es
n
o
t
eli
m
in
ate
th
e
n
ee
d
t
o
p
r
eser
v
e
ad
e
q
u
ate
ASR
g
r
o
u
n
d
in
g
.
(
a)
(
b
)
Fig
u
r
e
2
.
Sen
s
itiv
ity
o
f
co
r
p
u
s
-
lev
el
W
E
R
to
th
e
s
tatic
f
u
s
io
n
weig
h
t r
o
n
(
a
)
FLE
UR
S
-
I
D
an
d
(
b
)
Har
k
it
T
a
k
e
n
t
o
g
e
t
h
e
r
,
t
h
e
s
e
r
e
s
u
l
t
s
s
h
o
w
t
h
a
t
r
i
s
a
d
o
m
a
i
n
-
s
e
n
s
i
t
iv
e
i
n
t
e
r
p
o
l
a
t
i
o
n
c
o
e
f
f
i
c
i
e
n
t
r
a
th
e
r
t
h
a
n
a
u
n
i
v
e
r
s
a
l
c
o
n
s
t
a
n
t
.
F
L
E
UR
S
-
I
D
r
e
a
c
h
e
s
i
ts
o
p
t
i
m
u
m
a
t
r
=
0
.
7
5
,
w
h
e
r
e
a
s
H
a
r
k
it
r
e
a
c
h
es
i
ts
o
p
t
i
m
u
m
a
t
r
=
0
.
5
0
.
B
e
c
a
u
s
e
t
h
e
AS
R
a
n
d
L
L
M
s
c
o
r
e
d
i
s
t
r
i
b
u
t
i
o
n
s
a
r
e
n
o
t
n
e
c
e
s
s
a
r
i
l
y
c
al
i
b
r
a
t
e
d
t
o
a
co
m
m
o
n
s
c
a
l
e
,
t
h
e
s
e
n
u
m
e
r
i
c
a
l
v
a
l
u
e
s
s
h
o
u
l
d
n
o
t
b
e
i
n
t
e
r
p
r
e
t
e
d
a
s
a
b
s
o
l
u
t
e
m
ea
s
u
r
e
s
o
f
l
i
n
g
u
is
t
ic
i
n
f
l
u
e
n
c
e
a
c
r
o
s
s
d
o
m
a
i
n
s
o
r
b
a
c
k
e
n
d
s
.
I
n
s
t
e
a
d
,
t
h
e
d
i
f
f
e
r
i
n
g
o
p
t
i
m
a
i
n
d
i
c
a
t
e
t
h
a
t
t
h
e
e
f
f
e
c
t
i
v
e
b
a
l
a
n
c
e
d
e
p
e
n
d
s
j
o
i
n
t
l
y
o
n
a
c
o
u
s
t
i
c
r
e
l
i
a
b
il
i
t
y
,
l
e
x
i
c
a
l
u
n
c
e
r
t
a
i
n
t
y
,
p
r
o
m
p
t
i
n
g
c
o
n
d
i
t
i
o
n
,
a
n
d
t
h
e
a
l
i
g
n
m
e
n
t
o
f
t
h
e
L
L
M
p
r
i
o
r
w
it
h
t
h
e
t
a
r
g
et
s
p
e
ec
h
.
A
s
in
g
le
d
ataset
-
lev
el
v
alu
e
m
a
y
th
er
e
f
o
r
e
b
e
s
u
b
o
p
tim
al
f
o
r
in
d
iv
i
d
u
al
u
tter
an
ce
s
o
r
d
ec
o
d
in
g
s
tep
s
.
T
h
is
o
b
s
er
v
atio
n
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
15
,
No
.
5
,
Octo
b
er
20
26
:
4
8
9
3
-
490
6
4900
m
o
tiv
ates
f
u
tu
r
e
ad
a
p
tiv
e
f
u
s
io
n
m
ec
h
an
is
m
s
th
at
ad
ju
s
t
r
ac
co
r
d
in
g
to
lo
ca
l
r
ec
o
g
n
itio
n
co
n
f
id
en
ce
,
ASR
–
L
L
M
ag
r
ee
m
en
t,
o
r
d
ete
cted
d
o
m
ain
ch
ar
ac
ter
is
tics
.
3
.
3
.
St
a
t
is
t
ica
l sig
nifica
nce
a
nd
ef
f
ec
t
s
ize
T
ab
le
3
co
m
p
lem
en
ts
th
e
co
r
p
u
s
-
lev
el
r
esu
lts
in
T
a
b
le
2
with
p
air
e
d
u
tter
an
ce
-
lev
el
s
tatis
t
ical
test
s
.
T
h
e
W
E
R
v
alu
es
ar
e
r
ep
o
r
ted
at
t
h
e
c
o
r
p
u
s
lev
el,
Δ
W
E
R
d
en
o
tes
th
e
a
b
s
o
lu
t
e
r
ed
u
ctio
n
i
n
pp
,
an
d
th
e
p
-
v
alu
es
an
d
C
o
h
en
'
s
d
ef
f
ec
t
s
izes
ar
e
d
er
iv
ed
f
r
o
m
p
air
ed
u
tter
a
n
ce
-
lev
el
o
b
s
er
v
atio
n
s
.
B
o
th
d
atasets
s
h
o
w
s
tati
s
tically
s
i
g
n
if
ican
t
im
p
r
o
v
em
en
ts
(
p
<
0
.
0
1
)
,
co
n
f
ir
m
in
g
th
at
th
e
g
ain
s
d
eliv
er
ed
b
y
GFD
ar
e
n
o
t
attr
ib
u
ta
b
le
to
s
tatis
tical
v
ar
ian
ce
o
r
r
a
n
d
o
m
f
lu
ctu
atio
n
s
in
d
ec
o
d
i
n
g
.
T
h
e
ef
f
ec
t
s
ize
v
alu
es
(
C
o
h
en
'
s
d
≈
0
.
3
0
–
0
.
3
6
)
f
all
with
in
th
e
s
m
all
ef
f
ec
t
s
izes
r
an
g
e,
as
ex
p
ec
ted
,
g
iv
en
th
at
W
h
is
p
er
-
lar
g
e
-
v
3
is
alr
ea
d
y
a
r
o
b
u
s
t
m
u
ltil
in
g
u
al
b
aselin
e
f
o
r
I
n
d
o
n
esian
ASR
.
I
n
s
u
ch
s
tr
o
n
g
-
b
aselin
e
r
eg
i
m
es,
ev
en
m
o
d
er
ate
ef
f
ec
t
s
izes
ca
n
b
e
co
n
s
id
er
ed
m
ater
ially
m
ea
n
in
g
f
u
l
d
u
e
to
d
im
in
is
h
in
g
r
et
u
r
n
s
b
eh
av
io
r
t
y
p
ical
in
l
o
w
-
W
E
R
r
eg
io
n
s
.
M
o
r
e
im
p
o
r
ta
n
tly
,
th
e
s
tatis
tical
r
esu
lts
d
em
o
n
s
tr
ate
th
at
GFD
b
en
e
f
its
g
e
n
er
al
-
d
o
m
ain
a
n
d
d
o
m
ain
-
s
p
ec
if
ic
ASR
co
n
s
is
ten
tly
,
with
o
u
t
r
eq
u
ir
i
n
g
m
o
d
el
r
etr
ain
in
g
.
I
n
p
r
ac
tice,
t
h
is
in
d
icate
s
th
at
d
ec
o
d
in
g
-
lev
el
g
en
er
ativ
e
f
u
s
i
o
n
o
p
er
ates
as
a
r
o
b
u
s
t
im
p
r
o
v
em
en
t
lay
e
r
th
at
ca
n
b
e
e
x
p
ec
ted
to
g
en
er
alize
ac
r
o
s
s
task
co
n
d
itio
n
s
—
ev
en
wh
en
th
e
p
er
f
o
r
m
an
ce
g
ap
b
etwe
en
d
o
m
ain
s
d
if
f
e
r
s
s
u
b
s
tan
tially
in
ab
s
o
lu
te
W
E
R
m
ag
n
itu
d
e.
T
ab
le
3
.
Statis
tical
co
m
p
ar
is
o
n
o
f
W
h
is
p
er
-
lar
g
e
-
v
3
an
d
th
e
b
est GFD
co
n
f
ig
u
r
atio
n
s
D
a
t
a
s
e
t
W
h
i
s
p
e
r
W
ER
(
%)
B
e
st
G
F
D
W
E
R
(
%)
ΔW
ER
(
p
p
)
p
-
v
a
l
u
e
C
o
h
e
n
's
d
F
LEU
R
S
-
ID
5
.
9
1
4
.
1
7
1
.
7
4
1
.
6
3
×
1
0
⁻
¹⁹
0
.
3
6
H
a
r
k
i
t
(
me
d
i
c
a
l
)
1
5
.
8
9
1
2
.
3
0
3
.
5
9
3
.
5
0
×
1
0
⁻
¹³
0
.
3
1
3
.
4
.
L
L
M
-
prio
r
a
nd
pro
m
pt
-
co
nd
it
io
nin
g
a
bla
t
io
n
Fig
u
r
e
3
co
m
p
ar
es
u
tter
an
c
e
-
lev
el
W
E
R
d
is
tr
ib
u
tio
n
s
ac
r
o
s
s
L
L
M
b
ac
k
en
d
s
an
d
p
r
o
m
p
tin
g
co
n
d
itio
n
s
u
s
in
g
t
h
e
d
ev
el
o
p
m
en
t
-
s
elec
ted
s
tatic
r
f
o
r
e
ac
h
b
ac
k
en
d
–
p
r
o
m
p
t
co
m
b
i
n
atio
n
.
E
ac
h
v
io
lin
r
ep
r
esen
ts
th
e
d
is
tr
ib
u
tio
n
ac
r
o
s
s
test
u
tter
an
ce
s
,
with
th
e
b
lack
m
ar
k
e
r
an
d
a
d
jace
n
t
v
alu
e
in
d
icatin
g
t
h
e
m
ed
ian
W
E
R
.
Un
lik
e
th
e
c
o
r
p
u
s
-
lev
el
W
E
R
in
T
ab
le
2
,
Fig
u
r
e
3
s
u
m
m
ar
izes
ty
p
ical
u
tter
an
ce
-
lev
el
b
eh
av
io
r
;
t
h
er
ef
o
r
e,
th
e
two
m
ea
s
u
r
es
ar
e
co
m
p
lem
e
n
tar
y
r
ath
er
th
an
n
u
m
e
r
ically
eq
u
iv
alen
t.
B
ec
au
s
e
ea
ch
co
n
f
ig
u
r
atio
n
u
s
e
s
its
o
wn
s
elec
ted
r
,
Fig
u
r
e
3
c
o
m
p
ar
es
th
e
b
est
-
p
er
f
o
r
m
in
g
c
o
n
f
ig
u
r
atio
n
s
r
ath
er
th
an
is
o
latin
g
th
e
ef
f
ec
t o
f
p
r
o
m
p
ti
n
g
; d
ir
ec
t sam
e
-
r
p
r
o
m
p
t c
o
m
p
ar
is
o
n
s
ar
e
p
r
o
v
id
e
d
in
Fig
u
r
e
2
.
Fig
u
r
e
3
(
a)
s
h
o
ws
th
at
SE
A
-
L
I
ON
ac
h
iev
es
t
h
e
lo
wes
t
m
ed
ian
u
tter
an
ce
-
lev
el
W
E
R
o
n
t
h
e
FLE
UR
S
-
I
D.
F
ig
u
r
e
3
(
b
)
s
h
o
ws
th
at
p
r
o
m
p
ted
Mistra
l
ac
h
iev
es
th
e
lo
west
m
ed
ian
o
n
Har
k
it,
wh
er
ea
s
p
r
o
m
p
te
d
SAI
,
d
esp
ite
h
a
v
i
n
g
a
h
ig
h
er
m
ed
ian
an
d
b
r
o
ad
er
d
is
tr
ib
u
tio
n
,
ac
h
iev
es
th
e
lo
west
Har
k
it
co
r
p
u
s
-
lev
el
W
E
R
r
ep
o
r
ted
in
T
ab
le
2
.
T
h
is
d
if
f
er
e
n
ce
r
ef
le
cts th
e
ag
g
r
eg
atio
n
m
eth
o
d
s
: Fig
u
r
e
3
em
p
h
asizes
ty
p
ical
u
tter
an
ce
-
lev
el
b
e
h
av
i
o
r
th
r
o
u
g
h
th
e
m
e
d
ian
,
wh
e
r
e
as
T
ab
le
2
ag
g
r
eg
ates
er
r
o
r
s
ac
r
o
s
s
all
r
ef
er
en
ce
wo
r
d
s
.
Ov
er
all,
th
e
r
esu
lts
in
d
icate
th
at
b
ac
k
en
d
p
er
f
o
r
m
an
c
e
d
ep
en
d
s
o
n
th
e
s
p
ee
ch
co
n
d
itio
n
an
d
e
v
alu
atio
n
m
etr
ic,
wh
ile
t
h
e
ex
te
r
n
al
L
L
M
p
r
io
r
p
r
o
v
i
d
es
a
m
o
r
e
c
o
n
s
is
ten
t
im
p
r
o
v
em
en
t
th
an
ex
p
licit
in
s
tr
u
ctio
n
p
r
o
m
p
tin
g
,
wh
o
s
e
ef
f
ec
t r
em
ai
n
s
co
n
f
ig
u
r
atio
n
-
d
ep
en
d
en
t.
(
a)
(
b
)
Fig
u
r
e
3
.
Utter
an
ce
-
lev
el
W
E
R
d
is
tr
ib
u
tio
n
s
f
o
r
th
e
d
ev
elo
p
m
en
t
-
s
elec
ted
GFD
co
n
f
ig
u
r
atio
n
s
o
n
(
a)
FLE
UR
S
-
I
D
an
d
(
b
)
Har
k
i
t
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
E
mp
ir
ica
l e
va
lu
a
tio
n
o
f g
en
er
a
tive
fu
s
io
n
d
ec
o
d
in
g
fo
r
I
n
d
o
n
esia
n
g
en
era
l a
n
d
med
ic
a
l
…
(
A
s
r
il J
a
r
in
)
4901
3
.
5
.
Dis
cus
s
io
n
Acr
o
s
s
th
e
g
en
er
al
-
d
o
m
ai
n
(
FLE
UR
S
-
I
D)
an
d
m
ed
ical
-
d
o
m
ain
(
Har
k
it)
ev
al
u
atio
n
s
,
th
e
r
esu
lt
s
co
llectiv
ely
co
n
f
ir
m
th
at
GFD
im
p
r
o
v
es
I
n
d
o
n
esian
ASR
p
e
r
f
o
r
m
a
n
ce
r
elativ
e
to
a
s
tr
o
n
g
W
h
is
p
er
b
aselin
e
(
R
Q1
)
.
T
h
e
im
p
r
o
v
e
m
en
ts
ar
e
o
b
tain
ed
with
o
u
t
r
etr
ain
in
g
,
p
u
r
ely
th
r
o
u
g
h
p
r
o
b
ab
ilis
tic
s
t
ee
r
in
g
at
d
ec
o
d
in
g
tim
e.
T
h
is
f
in
d
in
g
estab
lis
h
es
GFD
as
a
p
r
ac
tical,
s
ca
lab
le
m
ec
h
an
is
m
f
o
r
ex
p
lo
itin
g
lin
g
u
is
tic
p
r
io
r
s
in
u
n
d
er
r
ep
r
esen
ted
la
n
g
u
a
g
es su
ch
as I
n
d
o
n
esian
.
Seco
n
d
,
f
u
s
io
n
weig
h
t
a
n
aly
s
is
d
em
o
n
s
tr
ates
th
at
is
n
o
t
a
u
n
iv
e
r
s
al
co
n
s
tan
t
(
R
Q2
)
.
I
n
s
tead
,
it
en
co
d
es
a
d
o
m
ain
tr
ad
e
-
o
f
f
:
h
ig
h
er
v
alu
es
(
≈
0
.
7
0
–
0
.
8
0
)
a
r
e
o
p
tim
al
f
o
r
g
en
e
r
al
-
d
o
m
ain
I
n
d
o
n
esian
s
p
ee
ch
,
wh
er
e
ac
o
u
s
tic
ev
id
e
n
ce
is
clea
n
,
an
d
W
h
is
p
er
alr
ea
d
y
m
o
d
els
co
r
e
s
tr
u
ctu
r
e
w
ell;
wh
er
ea
s
m
ed
ical
s
p
ee
ch
b
en
ef
its
m
o
s
t
at
m
o
d
e
r
ate
f
u
s
io
n
(
≈
0
.
5
0
)
,
wh
er
e
ter
m
in
o
lo
g
y
d
en
s
ity
r
e
q
u
ir
es
lin
g
u
is
tic
p
r
io
r
s
u
p
p
o
r
t
b
u
t
ex
ce
s
s
iv
e
L
L
M
d
o
m
in
a
n
ce
lead
s
to
o
v
er
-
co
r
r
ec
tio
n
.
T
h
e
o
b
s
er
v
e
d
d
o
m
ain
-
d
e
p
en
d
en
t
o
p
tim
u
m
d
em
o
n
s
tr
ates
em
p
ir
ically
th
at
a
f
ix
ed
co
ef
f
icien
t
is
u
n
lik
e
ly
to
b
e
u
n
iv
e
r
s
ally
o
p
tim
al
ac
r
o
s
s
I
n
d
o
n
esian
ASR
co
n
d
itio
n
s
[
3
3
]
.
An
ad
d
itio
n
al
lim
itatio
n
co
n
c
er
n
s
th
e
u
s
e
o
f
a
s
tatic
f
u
s
io
n
weig
h
t.
I
n
th
e
p
r
esen
t
s
tu
d
y
,
a
s
in
g
le
v
alu
e
o
f
r
,
s
elec
ted
u
s
in
g
d
ev
elo
p
m
en
t
d
ata,
is
ap
p
lied
u
n
if
o
r
m
ly
th
r
o
u
g
h
o
u
t
d
ec
o
d
in
g
,
ev
e
n
th
o
u
g
h
ac
o
u
s
tic
r
eliab
ilit
y
an
d
lex
ica
l
am
b
ig
u
ity
m
ay
v
a
r
y
s
u
b
s
tan
tially
ac
r
o
s
s
u
tter
an
ce
s
an
d
d
ec
o
d
in
g
p
o
s
itio
n
s
.
T
h
e
o
b
s
er
v
ed
d
o
m
ai
n
-
d
e
p
en
d
en
t
o
p
tim
u
m
th
er
ef
o
r
e
s
u
g
g
ests
th
at
a
f
ix
e
d
co
e
f
f
icien
t
is
u
n
lik
ely
to
b
e
u
n
iv
er
s
ally
o
p
tim
al.
Fu
tu
r
e
w
o
r
k
m
a
y
in
s
tead
i
n
v
esti
g
ate
a
d
ap
tiv
e
f
u
s
io
n
m
ec
h
an
is
m
s
o
p
er
atin
g
at
m
u
ltip
le
lev
els.
At
th
e
u
tter
an
ce
lev
el,
r
co
u
ld
b
e
co
n
d
itio
n
ed
o
n
g
lo
b
al
ASR
co
n
f
id
en
ce
,
esti
m
ated
ac
o
u
s
tic
q
u
ality
,
o
r
d
etec
ted
d
o
m
ai
n
cu
es.
At
th
e
to
k
en
o
r
d
ec
o
d
in
g
-
s
tep
lev
el,
it
co
u
ld
b
e
a
d
ju
s
ted
ac
co
r
d
in
g
to
p
o
s
ter
io
r
en
tr
o
p
y
,
th
e
s
co
r
e
m
ar
g
in
b
et
wee
n
co
m
p
etin
g
b
ea
m
h
y
p
o
th
eses
,
ASR
–
L
L
M
d
is
ag
r
ee
m
en
t,
o
r
u
n
ce
r
tain
ty
in
n
ex
t
-
to
k
e
n
p
r
ed
ictio
n
.
Su
ch
a
s
tr
ateg
y
co
u
ld
in
cr
ea
s
e
L
L
M
in
f
lu
en
ce
wh
en
ac
o
u
s
tic
ev
id
en
ce
is
u
n
ce
r
tain
b
u
t
lin
g
u
is
tic
ev
id
en
ce
is
r
eli
ab
le,
wh
ile
p
r
eser
v
in
g
s
tr
o
n
g
er
ASR
g
u
id
an
ce
wh
en
th
e
l
an
g
u
ag
e
p
r
io
r
r
is
k
s
o
v
er
-
c
o
r
r
ec
tio
n
.
I
m
p
lem
en
tin
g
ad
a
p
tiv
e
f
u
s
io
n
wo
u
ld
r
eq
u
ir
e
ca
r
ef
u
l
ca
lib
r
a
tio
n
to
e
n
s
u
r
e
s
tab
le
d
ec
o
d
in
g
b
eh
av
i
o
r
.
I
n
p
ar
ticu
lar
,
th
e
f
u
s
io
n
co
ef
f
icien
t
s
h
o
u
ld
b
e
s
m
o
o
th
ed
a
n
d
b
o
u
n
d
ed
to
av
o
id
ab
r
u
p
t
f
lu
ctu
atio
n
s
d
u
r
in
g
b
ea
m
s
ea
r
ch
.
All
ad
ap
tatio
n
r
u
les
s
h
o
u
ld
b
e
tu
n
ed
e
x
clu
s
iv
ely
o
n
d
e
v
elo
p
m
e
n
t
d
ata
b
e
f
o
r
e
e
v
alu
atio
n
o
n
h
eld
-
o
u
t
test
d
ata
to
p
r
e
v
en
t
s
elec
tio
n
b
ias.
T
h
e
p
r
esen
t
s
tu
d
y
d
o
es
n
o
t
im
p
lem
en
t
s
u
ch
an
ad
a
p
tiv
e
m
ec
h
an
is
m
;
in
s
tead
,
th
e
em
p
i
r
ical
o
b
s
er
v
atio
n
th
at
d
if
f
er
en
t
d
o
m
ain
s
f
av
o
r
d
if
f
e
r
en
t
v
al
u
es
o
f
r
p
r
o
v
id
es
a
m
o
tiv
atio
n
f
o
r
f
u
tu
r
e
c
o
n
f
id
e
n
ce
-
an
d
d
o
m
ain
-
co
n
d
itio
n
ed
f
u
s
io
n
s
tr
ateg
ies.
T
h
e
p
r
esen
t
r
esu
lts
s
h
o
u
ld
n
o
t
b
e
in
ter
p
r
eted
as
d
em
o
n
s
tr
at
in
g
th
at
GFD
o
u
tp
er
f
o
r
m
s
co
n
v
en
tio
n
al
s
h
allo
w
f
u
s
io
n
,
co
ld
f
u
s
io
n
,
o
r
C
L
AS,
b
ec
au
s
e
th
ese
s
y
s
tem
s
wer
e
n
o
t
im
p
lem
en
t
ed
in
th
e
c
u
r
r
en
t
ex
p
er
im
en
ts
.
T
h
e
em
p
i
r
ical
co
m
p
ar
is
o
n
in
th
is
s
tu
d
y
is
s
p
ec
if
ically
b
etwe
en
GFD
an
d
th
e
u
n
c
h
an
g
e
d
W
h
is
p
er
-
lar
g
e
-
v
3
d
ec
o
d
er
.
T
h
e
alter
n
ativ
e
m
eth
o
d
s
al
s
o
r
ep
r
esen
t
d
if
f
er
en
t
in
teg
r
atio
n
ass
u
m
p
tio
n
s
.
C
o
n
v
en
tio
n
al
s
h
allo
w
f
u
s
io
n
co
m
b
in
es
in
d
e
p
en
d
e
n
tly
tr
ain
ed
ASR
an
d
L
M
s
co
r
es
d
u
r
in
g
b
ea
m
s
ea
r
ch
[
2
2
]
.
C
o
ld
f
u
s
io
n
in
c
o
r
p
o
r
ates
a
f
ix
ed
p
r
etr
ai
n
ed
L
M
d
u
r
i
n
g
r
ec
o
g
n
izer
tr
ain
i
n
g
[
2
3
]
,
wh
er
ea
s
C
L
AS
jo
in
tly
lear
n
s
co
n
tex
tu
al
-
p
h
r
ase
r
ep
r
esen
tati
o
n
s
with
th
e
AS
R
s
y
s
tem
[
2
4
]
.
B
y
co
n
tr
ast,
GFD
u
s
es
b
y
te
-
lev
el
lik
elih
o
o
d
m
ap
p
in
g
t
o
co
m
b
in
e
a
f
r
o
ze
n
r
ec
o
g
n
izer
an
d
a
f
r
o
ze
n
a
u
to
r
eg
r
ess
iv
e
L
L
M
d
u
r
in
g
d
ec
o
d
in
g
[
2
7
]
.
T
h
is
tr
ain
in
g
-
f
r
ee
p
r
o
p
er
ty
m
ak
es
GFD
p
ar
ticu
lar
ly
r
elev
an
t
wh
en
r
etr
ain
in
g
d
ata
o
r
m
o
d
el
ac
ce
s
s
i
s
lim
ited
.
St
ill,
it d
o
es n
o
t b
y
its
elf
estab
lis
h
s
u
p
er
io
r
ity
o
v
er
tr
ain
in
g
-
b
ased
o
r
co
n
te
x
t
-
s
p
ec
if
ic
alter
n
ativ
e
s
.
A
d
ir
ec
t
co
m
p
ar
ativ
e
s
tu
d
y
s
h
o
u
ld
im
p
le
m
en
t
th
ese
m
eth
o
d
s
u
n
d
er
m
atch
ed
d
ata
p
ar
ti
tio
n
s
,
tex
t
n
o
r
m
aliza
tio
n
,
b
ea
m
wid
th
,
s
ea
r
ch
b
u
d
g
et,
lan
g
u
ag
e
-
m
o
d
el
ca
p
ac
ity
,
co
n
tex
t
u
a
l
r
eso
u
r
ce
s
,
a
n
d
h
y
p
er
p
ar
am
eter
-
s
elec
tio
n
p
r
o
ce
d
u
r
es.
I
n
a
d
d
itio
n
to
W
E
R
,
s
u
ch
ev
alu
atio
n
s
h
o
u
l
d
r
ep
o
r
t
tr
ai
n
in
g
r
eq
u
ir
em
e
n
ts
,
in
f
er
en
ce
laten
c
y
,
m
em
o
r
y
co
n
s
u
m
p
tio
n
,
an
d
co
m
p
u
tatio
n
al
c
o
s
t.
T
h
is
co
n
t
r
o
lled
co
m
p
ar
is
o
n
is
lef
t f
o
r
f
u
tu
r
e
wo
r
k
.
T
h
e
ab
latio
n
an
aly
s
is
d
is
tin
g
u
is
h
es
th
e
co
n
tr
ib
u
tio
n
o
f
th
e
e
x
ter
n
al
L
L
M
p
r
io
r
f
r
o
m
th
at
o
f
ex
p
licit
in
s
tr
u
ctio
n
p
r
o
m
p
tin
g
(
R
Q3
)
.
T
h
e
W
h
is
p
er
-
o
n
ly
b
aselin
e
(
r
=
0
)
e
x
clu
d
es
e
x
ter
n
a
l
L
L
M
g
u
id
a
n
ce
;
non
-
p
r
o
m
p
ted
GFD
in
tr
o
d
u
c
es
th
e
p
r
etr
ain
ed
L
L
M
p
r
io
r
with
o
u
t
task
in
s
tr
u
ctio
n
s
,
an
d
p
r
o
m
p
ted
GFD
ad
d
itio
n
ally
in
co
r
p
o
r
ates
a
f
ix
ed
d
ataset
-
s
p
ec
if
ic
in
s
tr
u
ctio
n
.
B
ec
au
s
e
s
ev
er
al
n
o
n
-
p
r
o
m
p
ted
co
n
f
ig
u
r
atio
n
s
alr
ea
d
y
im
p
r
o
v
e
o
v
er
th
e
W
h
is
p
er
b
aselin
e,
th
e
o
b
s
er
v
ed
g
ain
s
ca
n
n
o
t
b
e
attr
ib
u
ted
s
o
lely
to
p
r
o
m
p
tin
g
.
I
n
s
tead
,
th
e
ex
ter
n
al
L
L
M
p
r
io
r
p
r
o
v
id
es
th
e
m
o
r
e
co
n
s
is
ten
tly
o
b
s
er
v
ed
c
o
n
tr
ib
u
tio
n
.
I
n
co
n
t
r
ast,
p
r
o
m
p
t
co
n
d
itio
n
in
g
p
r
o
d
u
ce
s
s
m
aller
,
co
n
f
ig
u
r
atio
n
-
d
ep
en
d
en
t
ch
an
g
es
wh
o
s
e
ef
f
ec
tiv
en
ess
v
ar
ies
with
th
e
L
L
M
b
ac
k
en
d
,
s
p
ee
ch
d
o
m
ain
,
a
n
d
s
elec
ted
s
tatic
f
u
s
io
n
weig
h
t.
B
ac
k
en
d
co
m
p
ar
is
o
n
s
lik
ewise
d
o
n
o
t
s
u
p
p
o
r
t
a
u
n
iv
e
r
s
al
p
er
f
o
r
m
an
ce
r
a
n
k
in
g
.
SEA
-
L
I
ON
ac
h
iev
es
th
e
lo
west
m
ed
ian
u
tter
an
ce
-
lev
el
W
E
R
o
n
FLE
UR
S
-
I
D,
p
r
o
m
p
ted
Mistra
l
attain
s
th
e
lo
west
m
ed
ian
o
n
Har
k
it,
a
n
d
p
r
o
m
p
ted
SAI
p
r
o
d
u
ce
s
th
e
l
o
west
Har
k
it
co
r
p
u
s
-
lev
el
W
E
R
.
T
h
ese
co
m
p
lem
en
tar
y
o
b
s
er
v
atio
n
s
in
d
icate
th
at
b
ac
k
en
d
ef
f
ec
tiv
en
ess
d
ep
en
d
s
o
n
b
o
th
t
h
e
s
p
ee
ch
co
n
d
itio
n
a
n
d
th
e
ag
g
r
eg
atio
n
m
etr
ic.
B
ec
au
s
e
th
e
p
r
esen
t
ex
p
er
im
e
n
ts
d
o
n
o
t
is
o
late
th
e
in
d
iv
i
d
u
al
ef
f
ec
ts
o
f
p
r
e
-
tr
ain
in
g
co
r
p
u
s
co
m
p
o
s
itio
n
,
to
k
e
n
izer
d
esig
n
,
in
s
tr
u
ctio
n
tu
n
in
g
,
o
r
m
o
d
el
ar
ch
itectu
r
e,
attr
ib
u
tin
g
th
e
o
b
s
er
v
ed
d
if
f
er
en
ce
s
to
an
y
s
in
g
le
b
ac
k
en
d
ch
ar
ac
te
r
is
tic
r
em
ain
s
o
u
ts
id
e
th
e
s
co
p
e
o
f
th
is
s
tu
d
y
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
15
,
No
.
5
,
Octo
b
er
20
26
:
4
8
9
3
-
490
6
4902
B
ey
o
n
d
th
ese
m
ec
h
an
is
tic
in
ter
p
r
etatio
n
s
,
th
e
p
r
esen
t
f
in
d
i
n
g
s
s
h
o
u
ld
also
b
e
in
ter
p
r
eted
with
in
th
e
b
o
u
n
d
ar
ies
o
f
th
e
ev
al
u
ated
d
ata
an
d
d
ec
o
d
i
n
g
c
o
n
d
itio
n
s
.
FLE
UR
S
-
I
D
r
ep
r
esen
ts
co
n
t
r
o
lled
r
ea
d
s
p
ee
ch
,
wh
er
ea
s
Har
k
it r
ep
r
esen
ts
a
co
n
tr
o
lled
,
ter
m
in
o
lo
g
y
-
d
e
n
s
e
m
ed
ical
-
d
ictatio
n
co
n
d
itio
n
.
T
h
e
Har
k
it
ev
alu
atio
n
d
o
es
n
o
t
co
v
er
th
e
f
u
ll
b
r
ea
d
th
o
f
I
n
d
o
n
esian
clin
ical
s
p
ee
ch
,
in
clu
d
in
g
s
p
o
n
tan
e
o
u
s
d
o
cto
r
–
p
atien
t
co
n
v
er
s
atio
n
s
,
m
u
ltip
ar
ty
c
lin
ical
in
ter
ac
tio
n
s
,
d
iv
er
s
e
m
ed
ical
s
p
ec
ialties
,
v
ar
y
i
n
g
d
o
cu
m
en
tatio
n
wo
r
k
f
lo
ws,
en
v
ir
o
n
m
en
tal
n
o
is
e,
f
ar
-
f
ield
r
ec
o
r
d
in
g
,
ac
ce
n
t
an
d
s
p
ea
k
in
g
-
s
ty
le
d
iv
er
s
i
ty
,
o
r
co
n
tin
u
o
u
s
ly
ch
an
g
in
g
clin
ical
co
n
tex
ts
[
2
4
]
.
Mo
r
e
o
v
er
,
th
e
r
elativ
ely
lim
ited
d
u
r
atio
n
o
f
th
e
two
ev
alu
atio
n
s
ets
co
n
s
tr
ain
s
b
r
o
ad
d
ep
lo
y
m
e
n
t
-
lev
el
g
e
n
er
aliza
tio
n
[
3
4
]
.
Acc
o
r
d
i
n
g
ly
,
th
e
r
esu
lts
p
r
o
v
id
e
co
n
tr
o
lled
ev
id
en
ce
th
at
GFD
ca
n
im
p
r
o
v
e
I
n
d
o
n
esian
ASR
u
n
d
er
th
e
ev
alu
ated
c
o
n
d
itio
n
s
,
in
clu
d
in
g
th
e
Ha
r
k
it
m
ed
ical
-
d
ictatio
n
d
ataset.
Sti
ll
,
th
ey
d
o
n
o
t
estab
lis
h
r
o
b
u
s
tn
ess
ac
r
o
s
s
th
e
b
r
o
ad
e
r
I
n
d
o
n
es
ian
clin
ical
-
s
p
ee
ch
d
o
m
ain
o
r
u
n
co
n
s
tr
ain
ed
r
ea
l
-
wo
r
ld
en
v
ir
o
n
m
e
n
ts
.
Fro
m
a
d
ep
lo
y
m
en
t
p
er
s
p
ec
ti
v
e,
th
e
ev
alu
ate
d
c
o
n
f
ig
u
r
atio
n
is
th
er
e
f
o
r
e
m
o
s
t
d
ir
ec
tly
a
p
p
licab
le
to
o
f
f
lin
e,
s
er
v
e
r
-
s
id
e
tr
an
s
cr
ip
ti
o
n
wo
r
k
f
lo
ws,
s
u
ch
as b
atch
tr
an
s
cr
ip
tio
n
o
f
p
r
er
ec
o
r
d
e
d
I
n
d
o
n
esian
s
p
ee
ch
an
d
p
o
s
t
-
en
co
u
n
ter
tr
an
s
cr
ip
tio
n
o
f
co
n
tr
o
lled
m
ed
ical
d
ictatio
n
.
I
n
th
e
latter
s
ce
n
ar
io
,
th
e
g
en
er
ated
tr
a
n
s
cr
ip
t
s
h
o
u
ld
b
e
r
e
v
iewe
d
a
n
d
co
r
r
e
cted
b
y
a
q
u
alif
ied
u
s
er
b
ef
o
r
e
it
is
in
co
r
p
o
r
ate
d
in
t
o
a
clin
ical
r
ec
o
r
d
.
L
a
r
g
er
an
d
m
o
r
e
d
iv
er
s
e
I
n
d
o
n
esian
c
o
r
p
o
r
a
c
o
v
er
i
n
g
m
u
ltip
le
m
ed
i
ca
l
s
p
ec
ialties
,
s
p
o
n
t
an
eo
u
s
an
d
d
ictated
s
p
ee
ch
,
d
iv
er
s
e
s
p
ea
k
er
s
,
in
ter
ac
tio
n
ty
p
es,
d
o
c
u
m
en
tatio
n
wo
r
k
f
lo
ws,
an
d
ac
o
u
s
tic
co
n
d
itio
n
s
ar
e
r
eq
u
ir
e
d
f
o
r
b
r
o
ad
e
r
v
alid
atio
n
.
A
f
u
r
th
er
lim
itatio
n
co
n
ce
r
n
s
th
e
d
ec
o
d
in
g
p
r
o
to
co
l.
Alth
o
u
g
h
GFD
in
teg
r
ates
L
L
M
p
r
o
b
ab
ilit
ies
at
ea
ch
au
to
r
eg
r
ess
iv
e
d
ec
o
d
in
g
s
tep
,
th
e
p
r
esen
t
im
p
l
em
en
tatio
n
was
ev
alu
ated
o
n
ly
u
n
d
er
o
f
f
lin
e
f
u
ll
-
u
tter
an
ce
d
ec
o
d
in
g
an
d
d
o
es
n
o
t
p
r
o
ce
s
s
in
co
m
in
g
s
p
ee
ch
in
s
tr
ea
m
in
g
ch
u
n
k
s
o
r
p
r
o
d
u
ce
co
n
tin
u
o
u
s
ly
u
p
d
ated
p
ar
tial
tr
an
s
cr
ip
ts
.
C
o
n
s
eq
u
en
tly
,
th
e
o
b
s
er
v
ed
W
E
R
im
p
r
o
v
em
en
ts
s
h
o
u
ld
n
o
t
b
e
in
ter
p
r
eted
as
ev
id
en
ce
o
f
r
ea
l
-
tim
e
ap
p
licab
ilit
y
,
s
tr
ea
m
in
g
r
ea
d
in
ess
,
o
n
-
d
ev
ice
d
ep
l
o
y
m
en
t,
o
r
in
ter
ac
tiv
e
co
n
v
er
s
atio
n
al
ASR
.
Dep
lo
y
m
en
t
in
a
n
ew
ap
p
licatio
n
d
o
m
ain
wo
u
ld
also
r
eq
u
ir
e
a
r
ep
r
esen
tativ
e
d
ev
elo
p
m
e
n
t
s
et
to
s
elec
t
th
e
s
tatic
f
u
s
io
n
weig
h
t
an
d
p
r
o
m
p
tin
g
c
o
n
d
itio
n
,
a
f
ter
wh
ich
th
e
s
elec
ted
co
n
f
ig
u
r
atio
n
wo
u
l
d
r
em
ain
f
ix
ed
d
u
r
in
g
o
p
er
atio
n
.
Fu
tu
r
e
d
ep
lo
y
m
en
t
-
o
r
ien
ted
ev
alu
a
tio
n
s
s
h
o
u
ld
th
er
ef
o
r
e
e
x
am
i
n
e
r
ea
l
-
tim
e
f
ac
to
r
,
f
ir
s
t
-
to
k
en
a
n
d
en
d
-
to
-
en
d
l
aten
cy
,
m
e
m
o
r
y
a
n
d
c
o
m
p
u
tatio
n
al
o
v
er
h
ea
d
,
c
o
m
p
u
tatio
n
al
th
r
o
u
g
h
p
u
t
,
en
d
p
o
i
n
tin
g
b
eh
a
v
io
r
,
an
d
p
ar
tial
-
h
y
p
o
t
h
esis
s
tab
ili
ty
.
Su
ch
s
tu
d
ies
s
h
o
u
ld
also
d
eter
m
in
e
wh
eth
er
th
e
ad
d
itio
n
al
L
L
M
s
co
r
in
g
co
s
t
ca
n
b
e
co
n
tr
o
lled
with
o
u
t
s
ac
r
if
icin
g
th
e
ac
cu
r
ac
y
g
ai
n
s
o
b
s
er
v
ed
u
n
d
e
r
o
f
f
lin
e
d
ec
o
d
in
g
.
T
h
e
in
teg
r
atio
n
o
f
ASR
an
d
ex
ter
n
al
L
L
Ms
also
r
aises
eth
ical
an
d
p
r
i
v
ac
y
co
n
s
id
er
atio
n
s
th
at
ar
e
n
o
t
ca
p
tu
r
ed
b
y
ag
g
r
e
g
at
e
W
E
R
.
T
h
e
p
r
esen
t
s
tu
d
y
u
s
es
ex
is
tin
g
d
ataset
s
o
b
tain
ed
f
r
o
m
th
e
o
f
f
icial
r
ep
o
s
ito
r
ies
id
en
tifie
d
in
t
h
e
d
ata
av
ailab
ilit
y
s
tatem
en
t
an
d
d
o
es
n
o
t
in
v
o
l
v
e
n
ew
p
a
r
ticip
an
t
r
ec
r
u
itm
e
n
t
o
r
n
ew
s
p
ee
ch
-
d
ata
co
llectio
n
.
Nev
er
th
eless
,
p
r
o
s
p
ec
tiv
e
u
s
e
with
r
ea
l
clin
ical
s
p
ee
ch
m
ay
in
v
o
lv
e
s
en
s
itiv
e
p
er
s
o
n
al
an
d
m
ed
i
ca
l
in
f
o
r
m
atio
n
.
Op
er
atio
n
al
im
p
lem
en
tatio
n
wo
u
ld
th
er
ef
o
r
e
r
e
q
u
ir
e
d
at
a
m
in
im
izatio
n
,
co
n
tr
o
lled
ac
ce
s
s
,
s
ec
u
r
e
s
to
r
ag
e
an
d
tr
an
s
m
is
s
io
n
,
ap
p
r
o
p
r
iate
r
eten
tio
n
lim
its
,
au
d
itab
ilit
y
,
an
d
co
m
p
lian
ce
with
ap
p
lica
b
le
in
s
titu
tio
n
al,
eth
ical,
an
d
leg
al
r
eq
u
ir
em
en
ts
.
GFD
s
h
o
u
ld
b
e
r
eg
a
r
d
ed
as
a
tr
an
s
cr
ip
tio
n
-
s
u
p
p
o
r
t
m
eth
o
d
r
ath
e
r
th
an
a
d
iag
n
o
s
tic,
tr
ea
tm
en
t
-
r
ec
o
m
m
en
d
atio
n
,
o
r
a
u
to
n
o
m
o
u
s
clin
ical
-
d
o
cu
m
e
n
tatio
n
s
y
s
tem
.
Fair
n
ess
ac
r
o
s
s
s
p
ea
k
er
g
r
o
u
p
s
was
n
o
t
ev
alu
ated
in
th
e
p
r
e
s
en
t
ex
p
er
im
en
ts
.
W
h
is
p
er
-
lar
g
e
-
v
3
a
n
d
th
e
ex
ter
n
al
L
L
M
b
ac
k
en
d
s
m
ay
in
h
e
r
it
r
ep
r
esen
tatio
n
a
l
b
iases
f
r
o
m
th
eir
p
r
e
-
tr
ain
in
g
d
ata,
in
cl
u
d
in
g
u
n
eq
u
al
c
o
v
er
a
g
e
o
f
I
n
d
o
n
esian
ac
ce
n
ts
,
r
eg
io
n
al
la
n
g
u
ag
e
v
ar
ieties,
s
p
ea
k
in
g
s
ty
les,
d
em
o
g
r
a
p
h
ic
g
r
o
u
p
s
,
s
p
ee
ch
ch
ar
ac
ter
is
tics
,
m
ed
ical
s
p
ec
ialties
,
an
d
l
ex
ical
co
n
v
en
tio
n
s
.
Dec
o
d
in
g
-
tim
e
f
u
s
io
n
m
ay
co
r
r
ec
t
s
o
m
e
ac
o
u
s
tically
u
n
ce
r
tain
wo
r
d
s
wh
ile
also
f
av
o
r
i
n
g
m
o
r
e
f
r
eq
u
en
t
o
r
s
tan
d
ar
d
i
ze
d
lan
g
u
a
g
e
f
o
r
m
s
.
T
h
is
b
eh
av
io
r
co
u
l
d
d
is
ad
v
a
n
tag
e
less
r
ep
r
esen
ted
p
r
o
n
u
n
ciatio
n
s
,
p
er
s
o
n
al
n
am
es,
r
e
g
io
n
al
ex
p
r
ess
io
n
s
,
s
p
ee
ch
im
p
air
m
en
ts
,
o
r
s
p
ec
i
alize
d
ter
m
in
o
lo
g
y
.
T
h
er
e
f
o
r
e
,
lo
wer
ag
g
r
e
g
ate
co
r
p
u
s
-
lev
e
l
o
r
u
tter
an
ce
-
lev
el
W
E
R
s
h
o
u
ld
n
o
t
b
e
i
n
ter
p
r
eted
as
ev
id
en
ce
t
h
at
er
r
o
r
s
ar
e
d
is
tr
ib
u
ted
eq
u
ally
ac
r
o
s
s
u
s
er
s
o
r
th
at
th
e
s
y
s
tem
is
f
r
ee
f
r
o
m
lin
g
u
is
tic,
d
em
o
g
r
ap
h
ic,
o
r
d
o
m
ain
-
r
elate
d
b
ias.
Fu
tu
r
e
wo
r
k
s
h
o
u
ld
in
clu
d
e
s
u
b
g
r
o
u
p
-
lev
el
p
er
f
o
r
m
an
ce
a
n
aly
s
is
,
er
r
o
r
-
r
ate
d
is
p
ar
ity
as
s
ess
m
en
t,
ex
am
in
atio
n
o
f
clin
i
ca
lly
co
n
s
eq
u
e
n
tial
s
u
b
s
titu
tio
n
s
an
d
o
m
is
s
io
n
s
,
an
d
ev
alu
atio
n
o
f
th
e
ef
f
ec
ts
o
f
f
u
s
io
n
o
n
u
n
d
er
r
ep
r
ese
n
ted
s
p
ea
k
er
s
an
d
lan
g
u
ag
e
v
ar
ieties.
T
ak
en
to
g
eth
e
r
,
th
ese
f
in
d
i
n
g
s
d
em
o
n
s
tr
ate
th
at
GFD
b
eh
av
es
as
a
s
tr
u
ctu
r
ed
in
ter
f
a
ce
b
etwe
en
ac
o
u
s
tic
ev
id
en
ce
an
d
lan
g
u
a
g
e
p
r
io
r
k
n
o
wled
g
e,
wh
er
e
t
h
e
m
ag
n
itu
d
e
o
f
b
e
n
ef
it
is
d
eter
m
in
ed
jo
in
tly
b
y
i)
th
e
s
tr
en
g
t
h
o
f
lin
g
u
is
tic
alig
n
m
en
t,
a
n
d
ii)
th
e
m
atch
b
etwe
en
r
an
d
d
o
m
ain
s
ig
n
al
ch
ar
ac
ter
is
tics
.
T
h
e
g
en
er
al
p
r
i
n
cip
le
em
er
g
in
g
is
th
at
ASR
–
L
L
M
f
u
s
io
n
f
o
r
I
n
d
o
n
esian
is
n
o
t
s
im
p
ly
a
f
u
n
ctio
n
o
f
“
m
o
r
e
I
n
d
o
n
esian
tu
n
in
g
”
,
b
u
t
d
e
p
en
d
s
o
n
h
o
w
lin
g
u
is
tic
d
en
s
ity
,
d
o
m
ain
s
p
ec
ializatio
n
,
an
d
d
ec
o
d
in
g
-
tim
e
weig
h
tin
g
in
te
r
ac
t.
T
h
is
in
s
ig
h
t
p
r
o
v
id
es
a
p
r
in
cip
led
f
o
u
n
d
atio
n
f
o
r
f
u
tu
r
e
d
o
m
ain
-
s
en
s
itiv
e
d
ec
o
d
in
g
s
tr
ateg
ies
in
I
n
d
o
n
esian
AS
R
.
At
th
e
s
am
e
tim
e,
d
ep
lo
y
m
en
t
in
o
p
er
atio
n
al
en
v
ir
o
n
m
en
ts
will
r
eq
u
ir
e
Evaluation Warning : The document was created with Spire.PDF for Python.