Building Spoken Dialogue Systems for Embodied Agents Lecture 3
Slides for On the Usability of Spoken Dialogue …kom.aau.dk/~lbl/phd/Slides_for_On_the_Usability_of...
Transcript of Slides for On the Usability of Spoken Dialogue …kom.aau.dk/~lbl/phd/Slides_for_On_the_Usability_of...
On
the
Usa
bilit
y of
Spo
ken
Dia
logu
e Sy
stem
s
On
the
Usa
bilit
y of
Spo
ken
Dia
logu
e Sy
stem
s
Pre
sent
atio
nof
Ph.D
. the
sis
byLa
rs B
o La
rsen
Aal
borg
Uni
vers
ity, S
ep. 1
2, 2
003
Page
2of
33
On
the
Usa
bilit
y of
Spo
ken
Dia
logu
e Sy
stem
s
Ove
rvie
w
•In
trodu
ctio
n –
back
grou
nd•
Def
initi
on o
f usa
bilit
y•
The
OV
ID p
roje
ct•
Obj
ectiv
e m
easu
res
•Tu
rn-ta
king
and
use
r ini
tiativ
es•
Per
ceiv
ed a
nd o
bser
ved
task
suc
cess
•S
ubje
ctiv
e m
easu
res
•Q
uest
ionn
aire
s fo
r mea
surin
g us
er s
atis
fact
ion
•Fa
ctor
Ana
lysi
s•
Com
bine
d an
alys
is u
sing
the
Par
adis
e sc
hem
e•
Sum
mar
y an
d co
nclu
sion
s
Page
3of
33
On
the
Usa
bilit
y of
Spo
ken
Dia
logu
e Sy
stem
s
Bac
kgro
und
This
wor
k ha
s be
en c
arrie
d ou
t in
two
phas
es:
•Th
e fir
st p
hase
was
the
expe
rimen
tal w
ork
carr
ied
out i
n th
e E
sprit
OV
ID p
roje
ct in
199
6-7
•Th
is re
sulte
d in
a n
umbe
r of r
epor
ts a
nd p
ublic
atio
ns in
19
97-9
9•
Ano
ther
impo
rtant
resu
lt w
as a
fully
ann
otat
ed d
ialo
gue
corp
us•
The
seco
nd m
ore
rece
nt p
hase
was
in 2
002-
3, w
here
the
resu
lts w
ere
verif
ied
and
anal
ysed
from
a m
etho
dica
l poi
nt
of v
iew
, and
new
ana
lyse
s w
ere
carr
ied
out o
n th
e co
rpus
•M
ost n
otab
ly, a
new
par
adig
m (c
alle
d P
arad
ise)
had
be
en p
ropo
sed
sinc
e th
e or
igin
al O
VID
wor
k•
In b
etw
een
the
two
phas
es I
mai
nly
wor
ked
in th
e ar
ea o
f m
ulti
mod
al s
yste
ms
and
teac
hing
.
Page
4of
33
On
the
Usa
bilit
y of
Spo
ken
Dia
logu
e Sy
stem
s
The
Goa
ls o
f the
OVI
D P
roje
ct
OVI
D T
echn
ical
Ann
ex:
“The
par
tner
s in
tend
to a
ppro
ach
the
wor
k vi
a a
serie
s of
co
ntro
lled
usab
ility
tria
lsof
the
softw
are
in a
real
istic
ba
nkin
g se
rvic
e w
ith re
al b
ank
cust
omer
s.Th
e re
sults
will
be
an a
sses
smen
tof h
ow b
ank
cust
omer
s ar
e ab
le to
use
the
auto
mat
ed s
ervi
ce w
ithou
t tra
inin
gin
its
use,
to d
esig
n an
opt
imal
use
r int
erfa
ce d
ialo
gue
whi
ch c
an
acco
mm
odat
e th
e un
trai
ned
user
.”
Page
5of
33
On
the
Usa
bilit
y of
Spo
ken
Dia
logu
e Sy
stem
s
Usa
bilit
y
ISO
’s d
efin
ition
of u
sabi
lity:
•Ef
fect
iven
ess:
Acc
urac
y an
d co
mpl
eten
ess
with
w
hich
use
rs a
chie
ve s
peci
fied
goal
s.•
Effic
ienc
y:R
esou
rces
exp
ende
d in
rela
tion
to th
e ac
cura
cy a
nd c
ompl
eten
ess
with
whi
ch u
sers
ach
ieve
go
als.
•Sa
tisfa
ctio
n:Fr
eedo
m fr
om d
isco
mfo
rt, a
nd p
ositi
ve
attit
udes
tow
ards
the
use
of th
e pr
oduc
t.
Usa
bilit
y:ex
tent
to w
hich
a p
rodu
ct c
an b
e us
ed b
y sp
ecifi
ed u
sers
to a
chie
ve s
peci
fied
goal
s w
ith
effe
ctiv
enes
s, e
ffici
ency
, and
sat
isfa
ctio
n in
a s
peci
fied
cont
ext o
f use
.”
Page
6of
33
On
the
Usa
bilit
y of
Spo
ken
Dia
logu
e Sy
stem
s
Usa
bilit
y of
Spo
ken
Dia
logu
e Sy
stem
s (S
DS)
The
gene
ral d
efin
ition
of u
sabi
lity
and
the
asso
ciat
ed
attri
bute
s ar
e of
cou
rse
also
true
for t
he c
ase
of
SDS.
How
ever
:•
Due
to th
e co
mpl
exity
of t
he in
put p
roce
ssin
g an
d th
e no
n-pe
rsis
tenc
e of
spe
ech,
spe
cial
atte
ntio
n m
ust b
e pa
id to
:•
The
lear
nabi
lity
and
mem
orab
ility
of s
peec
h ba
sed
inte
rface
s•
The
trans
pare
ncy
and
erro
r-ha
ndlin
g ca
pabi
lity
For t
his
reas
on, t
he m
etho
ds th
at h
as b
een
deve
lope
d an
d ar
e w
ell-p
rove
n fo
r tra
ditio
nal i
nter
face
s ca
n no
t di
rect
ly b
e us
ed fo
r spe
ech.
Page
7of
33
On
the
Usa
bilit
y of
Spo
ken
Dia
logu
e Sy
stem
s
Usa
bilit
y M
easu
res
Two
orth
ogon
al c
ateg
orie
s of
usa
bilit
y m
easu
res
mus
t be
cap
ture
d si
mul
tane
ousl
y:
Obj
ectiv
e m
easu
res:
To
eval
uate
the
effe
ctiv
enes
s an
d ef
ficie
ncy
of th
e sy
stem
:•O
bser
ved
valu
es o
f e.g
. tim
e to
com
plet
e ta
sks,
task
suc
cess
ra
tes,
err
or ra
tes,
num
ber o
f hel
p m
essa
ges,
num
ber o
f use
r ba
rge-
ins,
etc
.•O
bjec
tive
mea
sure
s ar
e di
rect
ly o
bser
vabl
e
Subj
ectiv
e m
easu
res:
To
eval
uate
use
r pre
fere
nces
:•U
ser s
atis
fact
ion,
the
user
’s a
ttitu
des
tow
ards
the
over
all s
yste
m,
or p
artic
ular
asp
ects
of i
t•U
ser a
ttitu
des
cann
ot b
e ob
serv
ed d
irect
ly -
you
mus
t ask
the
user
s
Page
8of
33
On
the
Usa
bilit
y of
Spo
ken
Dia
logu
e Sy
stem
s
Req
uire
men
ts to
the
Dia
logu
e
Bas
ed o
n in
terv
iew
s w
ith b
anki
ng p
erso
nnel
, the
fu
nctio
nalit
y of
the
hom
e ba
nk w
as c
hose
n to
be:
•P
rovi
de b
alan
ce a
nd in
form
atio
n of
mov
emen
ts fo
r use
r ac
coun
ts.
•U
ser m
ust p
rovi
de Id
and
PIN
cod
es fo
r acc
ess
•Th
e us
er m
ust b
e (o
r fee
l) in
con
trol o
f the
dia
logu
e•
The
serv
ice
mus
t be
equa
lly a
ccep
tabl
e to
use
rs
rega
rdle
ss o
f gen
der,
age
and
acce
ntFu
rther
mor
e, th
e di
alog
ue m
ust a
ccom
mod
ate
the
untra
ined
use
r
Page
9of
33
On
the
Usa
bilit
y of
Spo
ken
Dia
logu
e Sy
stem
s
The
Ove
rall
Dia
logu
e M
odel
Id-n
umbe
r
Acc
ess C
ode
Mai
n
Bal
ance
Min
i Sta
t.
Pag
e 10
of 3
3
On
the
Usa
bilit
y of
Spo
ken
Dia
logu
e Sy
stem
s
Bal
ance
or
Min
i Sta
t ?
For
whi
chA
ccou
nt ?
Prov
ide
Bal
ance
Mor
eA
ccou
nts ?
For
whi
chA
ccou
nt ?
Prov
ide
Min
i-Sta
t
Mor
eA
ccou
nts ?
Whi
sh to
Con
tinue
?
Dia
logu
e T
ask
Stru
ctur
eM
ain
Bal
ance
Min
i Sta
t
Syst
em d
irec
ted
Tra
nsiti
ons:
Use
r In
itiat
ed
A M
ixed
-Initi
ativ
e D
ialo
gue
Mod
el w
ith S
hort
-Cut
s
Pag
e 11
of 3
3
On
the
Usa
bilit
y of
Spo
ken
Dia
logu
e Sy
stem
s
Obj
ectiv
e M
easu
res
on th
e O
VID
Cor
pus
The
Cor
pus:
•70
0 tra
nscr
ibed
dia
logu
es fo
r 310
use
rs w
ho w
ere
requ
este
d to
car
ry o
ut tw
o pr
e-de
fined
sce
nario
s.Sp
eech
I/O
Qua
lity:
•S
peec
h (c
once
pt) r
ecog
nitio
n pe
rform
ance
Dia
logu
e Sy
mm
etry
•Tu
rn-ta
king
stra
tegy
, in
parti
cula
r how
and
whe
n us
ers
took
the
initi
ativ
e in
the
dial
ogue
Com
mun
icat
ion
Effe
cien
cy•
Tim
ing
para
met
ers
for o
vera
ll an
d su
btas
k pe
rform
ance
Task
Effe
ctiv
enes
s:•
Task
suc
cess
rate
s
Pag
e 12
of 3
3
On
the
Usa
bilit
y of
Spo
ken
Dia
logu
e Sy
stem
s
Obj
ectiv
eM
easu
res
–Ti
min
g
All
user
s w
ere
requ
ired
to c
arry
out
two
scen
ario
s, A
and
B
The
tabl
e sh
ows
the
aver
age
time
spen
t in
the
logi
n su
btas
ks fo
r the
firs
t (A
1,B
1) a
nd s
econ
d (A
2,B
2) d
ialo
gues
A p
aire
d, tw
o ta
iled
t-tes
t rev
eale
d a
sign
ifica
nt re
duct
ion
of th
e tim
e sp
ent
in th
e “Id
_num
ber”
sub
task
whe
n co
mpa
ring
the
first
to th
e se
cond
di
alog
ue (p
= 0
.03)
Pag
e 13
of 3
3
On
the
Usa
bilit
y of
Spo
ken
Dia
logu
e Sy
stem
s
Obj
ectiv
eM
easu
res
–Tu
rnta
king
Ana
lysi
ngth
e tu
rn-ta
king
stra
tegi
es u
ncov
ered
a s
imila
r tre
nd –
user
s co
mpl
eted
the
dial
ogue
s w
ith a
sm
alle
r num
ber o
f tur
ns in
th
e se
cond
dia
logu
eIn
par
ticul
ar, u
sers
wer
e m
ore
will
ing
to ta
ke th
e in
itiat
ive
in th
e di
alog
ue,
the
mor
e ex
perie
nced
th
ey b
ecam
e.
Ave
rage
num
ber o
f use
r ini
tiativ
es
per d
ialo
gue
for t
he ”A
” and
”B”
scen
ario
s, fo
r the
firs
t and
sec
ond
dial
ogue
s.
An
unpa
ired
two-
taile
d t-t
est
show
s a
sign
ifica
nt (p
= 0
.02)
in
crea
se in
the
num
ber o
f use
r in
itiat
ives
rela
tive
to th
e to
tal
num
ber o
f tur
ns fo
r sce
nario
B2
com
pare
d to
B1.
Pag
e 14
of 3
3
On
the
Usa
bilit
y of
Spo
ken
Dia
logu
e Sy
stem
s
Obj
ectiv
e M
easu
res
–Ta
sk C
ompl
etio
n
Per
ceiv
ed v
ersu
s ob
serv
ed ta
sk s
ucce
ss:
•A
lthou
gh 9
6% o
f the
use
rs b
elie
ved
that
they
had
co
mpl
eted
bot
h sc
enar
ios,
onl
y 74
% a
ctua
lly d
id s
o
Ther
e is
a re
duct
ion
of
alm
ost 5
0% o
f the
faile
d di
alog
ues
from
the
first
to
the
seco
nd c
all (
a 25
%
redu
ctio
n is
sig
nific
ant a
t th
e 95
% c
onf.
leve
l)
Pag
e 15
of 3
3
On
the
Usa
bilit
y of
Spo
ken
Dia
logu
e Sy
stem
s
Obj
ectiv
eM
easu
res
–Sp
eech
Rec
ogni
tion
The
inte
rval
s sh
ow th
esp
eech
reco
gniti
onpe
rform
ance
ex
perie
nced
by a
par
ticul
arpr
opor
tion
ofus
ers
A re
cogn
ition
pe
rform
ance
of
90%
ro
ughl
y co
rres
pond
s to
one
erro
r pe
r dia
logu
e
Pag
e 16
of 3
3
On
the
Usa
bilit
y of
Spo
ken
Dia
logu
e Sy
stem
s
Obj
ectu
reM
easu
res
-Con
clus
ions
A s
igni
fican
t red
uctio
n of
tim
e fo
r the
ID_n
umbe
rsub
task
was
ob
serv
ed w
hen
com
parin
g du
ratio
ns o
f the
firs
t and
sec
ond
dial
ogue
s.A
naly
ses
of th
e us
ers’
turn
-tak
ing
stra
tegy
for t
he fi
rst a
nd
seco
nd c
alls
reve
al a
sig
nific
ant i
ncre
ase
in th
e us
ers’
te
nden
cy to
take
the
initi
ativ
e in
the
dial
ogue
.Ta
sk c
ompl
etio
n ra
tes
also
sho
wed
a s
igni
fican
t inc
reas
e fro
m
the
first
to th
e se
cond
dia
logu
e.Th
ese
findi
ngs
are
inte
rpre
ted
as s
igns
of s
yste
m le
arna
bilit
y.--
----
----
----
----
----
----
----
----
----
No
diffe
renc
es w
ere
iden
tifie
d fo
r use
rs fr
om d
iffer
ent
dem
ogra
phic
gro
ups
(gen
der,
regi
on, a
ge)
Pag
e 17
of 3
3
On
the
Usa
bilit
y of
Spo
ken
Dia
logu
e Sy
stem
s
Subj
ectiv
e M
easu
res
The
term
use
r sat
isfa
ctio
nis
use
d to
den
ote
the
degr
ee to
whi
ch
the
user
s ar
e sa
tisfie
d w
ith, o
r acc
ept t
he s
yste
m p
erfo
rman
ce.
Con
trary
to (m
ost)
obje
ctiv
e m
easu
res,
info
rmat
ion
of u
ser
satis
fact
ion
is n
ot d
irect
ly o
bser
vabl
e, b
ut m
ust b
e ob
tain
ed b
yas
king
the
user
sO
ften
the
user
is a
sked
to e
xpre
ss h
is/h
er a
ttitu
de to
war
ds a
nu
mbe
r of s
tate
men
ts a
bout
the
syst
em, f
or e
xam
ple
usin
g a
so-
calle
d Li
kert
attit
ude
ques
tionn
aire
Pag
e 18
of 3
3
On
the
Usa
bilit
y of
Spo
ken
Dia
logu
e Sy
stem
s
The
OVI
D Q
uest
ionn
aire
with
25
Stat
emen
tsA
vera
ge U
ser
Att
itude
s with
98%
con
fiden
ce in
terv
als
1234567
easy to use
knew what to do
friendliness
confusing
use again
reliability
out of control
like voice
concentration
effeciency
flustered
too fast
under stress
voice clear
frustation
prefer human
too complicated
enjoyment
needs improvement
politeness
security
convenient
confidentiality
remember too much
good value
Average
Cat
egor
y
Attitude
Dom
ain
Dep
.
Pag
e 19
of 3
3
On
the
Usa
bilit
y of
Spo
ken
Dia
logu
e Sy
stem
s
Fact
or A
naly
sis
Fact
or A
naly
sis
(FA
) is
used
to id
entif
y th
e un
derly
ing
rela
tions
hips
bet
wee
n th
e st
atem
ents
•M
athe
mat
ical
ly, F
A re
sem
bles
Prin
cipa
l Com
pone
nts
Ana
lysi
s (P
CA
), bu
t:•
In F
A, t
he fa
ctor
s ar
e pe
rcei
ved
as th
e ca
use
of th
e ob
serv
ed
varia
ble
scor
es, i
.e. i
t is
the
unde
rlyin
g fa
ctor
stru
ctur
e th
atha
s pr
oduc
ed (o
r cau
sed)
the
obse
rved
var
iabl
e sc
ores
.•
In c
ontra
st, f
or P
CA
, the
com
pone
nts
are
just
per
ceiv
ed a
s ag
greg
ates
of t
he o
bser
ved
varia
ble
scor
es•
Furth
erm
ore,
in P
CA
all
varia
nce
is m
odel
led,
whe
reas
in F
A o
nly
the
varia
nce
the
varia
bles
hav
e in
com
mon
(com
mun
aliti
es) a
re
cons
ider
ed•
FA h
as a
n el
emen
t of s
ubje
ctiv
e ju
dgm
ent,
sinc
e th
e go
al
is to
arr
ive
at a
fact
or s
et th
at w
ill p
rovi
de a
n in
terp
reta
tion
of th
e ob
serv
ed d
ata
Pag
e 20
of 3
3
On
the
Usa
bilit
y of
Spo
ken
Dia
logu
e Sy
stem
s
Verif
icat
ion
of O
VID
Fac
tors
OVI
D F
acto
rsVa
rianc
eF1
: Qua
lity
of in
terfa
ce/
19%
perfo
rman
ceU
se A
gain
, Rel
iabi
lity,
Effi
cien
cy,
pref
er H
uman
, Enj
oym
ent
Nee
ds Im
prov
emen
tF2
: Cog
nitiv
e lo
ad13
%C
once
ntra
tion,
Spe
ed,
Und
er S
tress
F3: C
ontro
l/Con
fusi
on9%
Kno
w w
hat w
as e
xpec
ted,
perc
eive
d co
ntro
l, C
onfu
sion
,Fl
uste
red,
/Too
Com
plic
ated
F4: F
riend
lines
s8%
Frie
ndly
, Pol
iteF5
: Voi
ce8%
Like
d V
oice
, Voi
ce c
lear
Tota
l Exp
lain
ed V
aria
nce
57%
Orig
inal
CC
IR F
acto
rsVa
rianc
eF1
:Qua
lity
of in
terfa
ce/
21%
perfo
rman
ceU
se A
gain
, Effi
cien
cy, R
elia
bilit
yN
eeds
Impr
ovem
ent
F2: C
ogni
tive
effo
rt an
d S
tress
, 17%
Spe
ed, U
nder
Stre
ss, C
once
ntra
tion,
P
erce
ived
con
trol
F3: C
onve
rsat
iona
l mod
elV
oice
, Ton
e pr
ompt
s, F
riend
lines
s
F4: F
luen
cyV
oice
cla
rity,
Pol
itene
ss,
Kno
w w
hat w
as e
xpec
ted
F5: T
rans
pare
ncy
Eas
e of
use
, Pro
mpt
hel
pful
ness
Deg
ree
of
flust
er
Tota
l Exp
lain
ed V
aria
nce
74%
Pag
e 21
of 3
3
On
the
Usa
bilit
y of
Spo
ken
Dia
logu
e Sy
stem
s
Six-
Fact
or S
truc
ture
Whe
n th
e fiv
e do
mai
n de
pend
ent
stat
emen
ts
are
adde
d, a
si
xth
fact
or
emer
ges.
Pag
e 22
of 3
3
On
the
Usa
bilit
y of
Spo
ken
Dia
logu
e Sy
stem
s
Cor
rela
ting
Stat
emen
t Sco
res
0
0,250,5
0,75
knew
wha
t to
do too
fast co
nfus
ing
polit
e like
voic
e
voic
e cl
ear
conc
entra
te
frie
ndly
rem
embe
r muc
h
stre
ssco
ntro
lco
nfid
entia
lity
secu
rity
flust
ered
need
s im
prov
emen
tpr
efer
hum
an
relia
bilit
y
com
plic
ated
easy
to u
se
Effic
ienc
y
Frus
tratio
n
good
val
ueen
joym
ent
conv
enie
nt
Cor
rela
tion
with
use
r atti
tude
to “
Use
aga
in”
F3: C
onve
nien
ce
Pag
e 23
of 3
3
On
the
Usa
bilit
y of
Spo
ken
Dia
logu
e Sy
stem
s
Subj
ectiv
e M
easu
res
-Con
clus
ions
The
ques
tionn
aire
use
d fo
r the
sub
ject
ive
mea
sure
s w
ere
show
n to
be
valid
and
pro
duce
a fa
ctor
stru
ctur
e si
mila
r to
that
of t
he
orig
inal
CC
IR q
uest
ionn
aire
Whe
n th
e do
mai
n de
pend
ent s
tate
men
ts w
ere
incl
uded
, the
fa
ctor
stru
ctur
e ch
ange
d an
d ne
w fa
ctor
“con
fiden
tialit
y”
emer
ged
Gen
eral
ly, t
he u
sers
wer
e po
sitiv
e to
war
ds th
e O
VID
hom
e ba
nk
serv
ice
(ave
rage
sco
re w
as 5
.6 –
i.ebe
twee
n “a
gree
” and
“s
trong
ly a
gree
”)S
imila
r to
the
obje
ctiv
e m
easu
res,
no
sign
ifica
nt d
iffer
ence
s be
twee
n th
e de
mog
raph
ic g
roup
s w
ere
foun
d
Pag
e 24
of 3
3
On
the
Usa
bilit
y of
Spo
ken
Dia
logu
e Sy
stem
s
Com
bini
ng O
bjec
tive
and
Subj
ectiv
e M
easu
res
•Th
e P
AR
AD
ISE
(Par
adig
m fo
r Dia
logu
e S
yste
m E
valu
atio
n)
sche
me
(pro
pose
d by
Wal
ker e
t al f
rom
AT&
T in
199
7)
atte
mpt
s to
com
bine
the
subj
ectiv
e an
d ob
ject
ive
mea
sure
s.•
This
is d
one
by e
stim
atin
g a
perfo
rman
ce fu
nctio
n w
ith
“usa
bilit
y” a
s th
e in
depe
nden
t var
iabl
e an
d th
e ob
ject
ive
mea
sure
s as
the
depe
nden
t var
iabl
es•
The
perfo
rman
ce e
quat
ion
is m
odel
ed u
sing
Mul
tiple
Lin
ear
Reg
ress
ion
(MLR
)
Pag
e 25
of 3
3
On
the
Usa
bilit
y of
Spo
ken
Dia
logu
e Sy
stem
s
The
PAR
AD
ISE
Mod
el
Kap
pa a
ttem
pts
to c
ompe
nsat
e fo
r diff
eren
ces
in
the
com
plex
ity o
f th
e di
alog
ues
Pag
e 26
of 3
3
On
the
Usa
bilit
y of
Spo
ken
Dia
logu
e Sy
stem
s
App
lyin
g PA
RA
DIS
E to
the
OVI
D C
orpu
s
Che
ckin
g th
e co
rrel
atio
n of
the
inde
pend
ent v
aria
bles
bef
ore
appl
ying
MLR
. Onl
y th
e sp
eech
reco
gniti
on a
nd ta
sk s
ucce
ss
para
met
ers
turn
ed o
ut to
be
sign
ifica
nt p
redi
ctor
s of
usa
bilit
y (r
epre
sent
ed a
s th
e F1
-fact
or g
roup
)
Pag
e 27
of 3
3
On
the
Usa
bilit
y of
Spo
ken
Dia
logu
e Sy
stem
s
The
Reg
resi
son
Pag
e 28
of 3
3
On
the
Usa
bilit
y of
Spo
ken
Dia
logu
e Sy
stem
s
The
Res
ultin
g Pe
rfor
man
ce F
unct
ion
The
resu
lting
per
form
ance
func
tion
for t
he O
VID
exp
erim
ent,
com
pare
d w
ith s
imila
r PA
RA
DIS
E a
naly
ses
by W
alke
r et a
l at
AT&
T.
Pag
e 29
of 3
3
On
the
Usa
bilit
y of
Spo
ken
Dia
logu
e Sy
stem
s
Estim
atio
n of
the
user
atti
tude
sO
bser
ved
and
Est
imat
ed U
ser A
ttitu
des
Use
rs
User Attitude (F1)
↓ +
95%
Con
f.
Obs
erve
d
Esi
tmat
ed
← -
95%
Con
f.
510
1520
2530
1234567
For v
erifi
catio
n of
the
mod
el, t
he id
entif
ied
para
met
ers
are
used
to
estim
ate
the
user
sat
isfa
ctio
n (r
ed li
ne).
It is
cle
ar th
at o
nly
half
of
the
varia
nce
of th
e ob
serv
ed (b
lue)
is c
aptu
red.
Pag
e 30
of 3
3
On
the
Usa
bilit
y of
Spo
ken
Dia
logu
e Sy
stem
s
Con
clus
ion
on th
e PA
RA
DIS
E R
esul
tsTh
e im
porta
nt q
uest
ion
is o
f cou
rse
whe
ther
any
new
in
form
atio
n w
as re
veal
ed.
•It
is h
ardl
y su
rpris
ing
that
a re
latio
nshi
p be
twee
n A
SR
pe
rform
ance
, tas
k su
cces
s an
d us
er s
atis
fact
ion
can
be
obse
rved
.•
Kap
pa p
rove
d to
be
a be
tter p
redi
ctor
of u
sabi
lity
than
a m
ore
sim
ple
ratio
of c
ompl
eted
sub
-goa
ls. T
he m
ain
func
tion
of K
appa
is
to n
orm
alis
efo
r tas
k co
mpl
exity
, whi
ch in
this
cas
e it
seem
s to
ha
ve d
one.
•Th
ere
is a
n (a
lmos
t sur
pris
ingl
y) g
ood
corr
espo
nden
ce b
etw
een
the
OV
ID re
sults
and
thos
e ob
tain
ed b
y A
T&T
•P
AR
AD
ISE
is li
mite
d by
the
requ
irem
ent f
or w
ell-d
efin
ed
scen
ario
bas
ed d
ialo
gues
and
a li
near
rela
tions
hip
betw
een
perfo
rman
ce m
easu
res
and
usab
ility
Pag
e 31
of 3
3
On
the
Usa
bilit
y of
Spo
ken
Dia
logu
e Sy
stem
s
Sum
mar
y
•C
erta
in u
sabi
lity
aspe
cts
mus
t rec
eive
spe
cial
atte
ntio
n, d
ue to
the
natu
re o
f spe
ech,
mos
t not
ably
tran
spar
ency
and
lear
nabi
lity
•Th
e re
quire
men
ts s
et u
p fo
r the
OV
ID d
ialo
gue
has
to a
larg
e de
gree
bee
n m
et. (
Exc
eptio
n: S
peed
) •
The
lear
nabi
lity
of th
e O
VID
dia
logu
e ha
s be
en d
emon
stra
ted
thro
ugh
mea
sure
s of
the
timin
g an
d tu
rn-ta
king
stra
tegy
•Th
e va
lidity
of t
he q
uest
ionn
aire
use
d fo
r OV
ID h
as b
een
esta
blis
hed
thro
ugh
fact
or a
naly
sis
•A
PA
RA
DIS
E a
naly
sis
conf
irmed
that
spe
ech
reco
gniti
on a
nd
task
suc
cess
are
impo
rtant
for u
ser s
atis
fact
ion,
and
a h
igh
corr
espo
nden
ce w
ith re
sults
obt
aine
d el
swhe
reis
sho
wn.
•Th
e im
porta
nt to
pic
of m
ulti
mod
al u
ser i
nter
actio
n an
d th
e is
sue
of m
emor
abili
tyha
ve n
ot b
een
addr
esse
d in
this
wor
k
Pag
e 32
of 3
3
On
the
Usa
bilit
y of
Spo
ken
Dia
logu
e Sy
stem
s
Wha
t is
the
in th
e Fu
ture
for S
peec
h?
•S
peec
h as
a m
odal
ity is
in a
hig
hly
com
petit
ive
“mar
ket”,
and
mus
t si
mpl
y be
bet
ter t
han
any
othe
r opt
ion
for p
eopl
e to
use
it•
Man
y en
visi
oned
“kill
er a
pplic
atio
ns” a
s e.
g. p
hone
-bas
ed h
ome
bank
ing
has
been
take
n ov
er b
y th
e W
eb (e
.g. 3
8% o
f Dan
ish
inte
rnet
use
rs u
sed
inte
rnet
hom
e ba
nkin
g re
gula
rly b
y 20
02, w
hile
no
ne u
sed
spee
ch)
•Th
e m
etho
ds fo
r mea
surin
g us
er s
atis
fact
ion
has
to a
larg
e de
gree
be
en o
verlo
oked
by
the
spee
ch c
omm
unity
and
mus
t rec
eive
mor
e at
tent
ion
if sp
eech
bas
ed in
terfa
ces
are
to b
e su
cces
sful
•M
uch
focu
s ha
s be
en o
n na
tura
lnes
s an
d us
er c
ontro
l, bu
t rea
lly
with
out a
ny h
ard
proo
f tha
t thi
s ac
tual
ly le
ads
to h
ighe
r use
r sa
tisfa
ctio
n –
lear
nabi
lity
mig
ht b
e ju
st a
s im
porta
nt•
The
focu
s on
mob
ility
mig
ht p
rovi
de a
bre
akth
roug
h fo
r spe
ech,
es
peci
ally
in c
ombi
natio
n w
ith o
ther
mod
aliti
es
Pag
e 33
of 3
3
On
the
Usa
bilit
y of
Spo
ken
Dia
logu
e Sy
stem
s
and
Fina
lly…
I wis
h to
than
k al
l tho
se o
f my
colle
ague
s at
CP
K a
nd th
e O
VID
team
who
hav
e he
lped
me
in th
is w
ork,
eith
er
dire
ctly
or b
y ta
king
ove
r som
e of
my
othe
r tas
ks.
I als
o w
ish
to th
ank
my
fam
ily fo
r the
ir su
ppor
t
Last
, I w
ish
to th
ank
you
all f
or c
omin
g he
re a
nd li
sten
to
wha
t I h
ad to
say
____
____
____
_