TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.5, May 2014, pp
. 3902 ~ 39
1
0
DOI: http://dx.doi.org/10.11591/telkomni
ka.v12i5.4816
3902
Re
cei
v
ed O
c
t
ober 2
0
, 201
3; Revi
se
d Decem
b
e
r
5, 2013; Accepte
d
De
cem
ber
31, 2013
Acoustic Source Localization Based on Iterative
Unscented Particle Filt
er
Jie Cao, Jia-qi Liu*, Di Wu, Jin-hua Wang
Coll
eg
e of Elec
trical an
d Information En
gi
ne
erin
g, Lanz
hou
Universit
y
of T
e
chn
o
lo
g
y
,
Lanz
ho
u 730
0
50, P.R. China
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: 4509
71
614
@
qq.com
A
b
st
r
a
ct
T
o
solve the
p
r
obl
em
of tracking a
n
aco
u
st
ic s
ource i
n
n
o
i
se an
d rever
b
eratio
n env
iron
me
nt, a
new
metho
d
is
prop
ose
d
i
n
p
u
rsuit of
hig
h
e
r
accura
cy. F
i
r
s
t, this pap
er i
m
pr
oves th
e u
n
scente
d
p
a
rti
c
l
e
filter, w
h
ich ca
n add th
e late
st meas
ure
m
e
n
t infor
m
at
io
n
to opti
m
i
z
e
th
e
propos
al
d
i
stri
butio
n. T
hen, the
likel
ih
ood
funct
i
on
is c
onstruct
ed
by c
a
lcu
l
ati
ng th
e
microp
h
one
arrays
’
out
put e
ner
gy i
n
t
he fra
m
ew
ork
of
the i
m
prove
d
a
l
gorit
hm. F
i
na
ll
y, t
he ex
peri
m
ent res
u
lts i
ndi
cate that th
e
p
r
opos
ed
loca
li
zation
metho
d
c
a
n
not o
n
ly i
m
pro
v
es the
accur
a
cy of loc
a
tio
n
e
s
timati
on,
but
also c
an
en
ha
nce th
e a
b
il
ity to resist
nois
e
an
d
reverberation in the acoustic
source locali
z
a
tion system
.
Ke
y
w
ords
:
co
mp
uter ap
pl
ica
t
ion, aco
u
stic
source l
o
cal
i
z
a
tion, iterativ
e un
sce
nted p
a
rticle filter (IUP
F),
micr
oph
on
e arr
a
ys, propos
al d
i
stributi
on funct
i
on
Copy
right
©
2014 In
stitu
t
e o
f
Ad
van
ced
En
g
i
n
eerin
g and
Scien
ce. All
rig
h
t
s reser
ve
d
.
1. Introduc
tion
Speaker lo
calizatio
n ba
sed on mi
cro
phon
e
array
is an impo
rtant topic in
Huma
n-
Comp
uter Int
e
ra
ction Research. It has wid
e
ly
ap
plicatio
ns in
several fields, incl
udin
g
the
multimedia
systems, vide
o co
nfere
n
ci
ng sy
stem
s
and mo
bile robotics a
nd
so o
n
[1]. The
traditional
sp
eaker lo
cali
zation ba
sed
on microp
ho
ne array ma
inly estimate
d the sp
ea
ker
locatio
n
by calcul
ating the
current time
delay
of the spee
ch
sign
al re
ceived b
y
microp
hon
e
array. In the
situation
of free
soun
d fiel
d, this
m
e
tho
d
can g
e
t a
great
effect
of locating a
n
d
tracking.
Ho
wever, it will generate lots o
f
fake
so
und
resou
r
ce und
er a st
ron
g
b
a
ckgroun
d no
ise
or the situatio
n of long reve
rbe
r
ation time
. As a re
sult, it will affect the accuracy of
locali
zation.
In re
ce
nt ye
ars,
with
the
develo
p
men
t
of non
-lin
e
a
r filteri
ng te
chni
que
s, re
sea
r
che
r
modele
d
the
spe
a
ker m
o
vement tra
c
kin
g
usi
ng th
e
way of
state-spa
c
e
metho
d
, simul
a
ting
the
situation
of spe
a
ker’
s m
o
vement wit
h
pro
p
e
r
dynamical eq
u
a
tion, synth
e
si
zing p
r
e
s
ent
informatio
n
with the p
a
st
o
nes,
su
rmo
u
n
t
the effe
ct
of virtual soun
d sou
r
ce effecti
v
ely
in
compl
e
x
noise situ
atio
n, and im
pro
v
e the a
c
curacy an
d
rob
u
stne
ss of
spea
ker lo
cali
zation
syste
m
.
Vermaa
k [2] i
n
trodu
ce
d th
e parti
cle filte
r
algo
rith
m to
spe
a
ker lo
ca
lization
syste
m
, establi
s
he
d a
rea
s
on
able
spea
ker motio
n
model to suppress
spurious noi
se
source
s, and con
s
tru
c
ted t
he
likeliho
o
d
fun
c
tion
acco
rdi
ng to th
e tim
e
del
ay e
s
tim
a
tion. Ward [
3
] mad
e
imp
r
ovements o
n
the
basi
s
of th
e p
r
eviou
s
, sum
m
ari
z
ed th
e
sound
so
urce l
o
cate
metho
d
s
an
d u
s
e
d
o
u
tput ene
rgy
of
a ste
e
red
be
am-former to
co
nst
r
u
c
t th
e likeliho
od f
unctio
n
, a
c
hi
eving a
c
cura
tely trackin
g
of
spe
a
ker by u
s
ing th
e pa
rticle filter. Respectively
, on
the basi
s
of
improved
pa
rticle filter, F
u
-
Liang Yin [4, 5] construct
ed likelih
ood
function
b
a
se
d on time-d
elay estimation
and the out
put
energy of SRP-PHAT b
e
a
m
forme
r
for trackin
g
a
nd
l
o
catin
g
the spea
ker. Nai
-
g
ao Jing
[6] u
s
e
d
quantum evol
utionary met
hod
s to improve particl
e
filter and ap
pli
ed the impro
v
ed particl
e filter
to spe
a
ke
r lo
cali
zation, an
d achieved g
ood result
s.
The ab
ove m
e
thod
s ca
n a
c
curately lo
cate
the sp
ea
ke
r i
n
sh
ort
reverberatio
n time
and l
a
rg
e S
NR
environm
ent, but
still cannot
effe
ctively
locate the
sp
eaker in the strong noi
se e
n
vironm
ent.
Therefore, th
is pa
per
pre
s
ent
s an
acous
ti
c source localizatio
n method
b
a
se
d o
n
iterative un
scented pa
rticl
e
filt
ering met
hod, which
consi
ders the
inhibition
roo
m
reverbe
r
ati
on
role
of mea
s
u
r
eme
n
t information a
nd
spea
ker motio
n
mod
e
l aimi
ng to the
ina
c
curate p
r
o
b
le
m of
spe
a
ker l
o
cation in th
e en
vironme
n
t of
long
reverbe
r
ation time a
n
d
sm
all SNR. The meth
o
d
use
s
th
e aud
io
si
gnal
coll
ected
by
th
e
microp
hon
e array as the observation i
n
formatio
n, and
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Acou
stic Sou
r
ce L
o
calization Base
d on
Iterati
ve
Unscented Parti
c
le
Filter (Ji
e
Ca
o)
3903
con
s
tru
c
t
s
the likelih
ood f
unctio
n
by ca
lculatin
g the steered be
a
m
forme
r
ene
rgy formed by
the
microph
one
array. On the
other
hand,
the introd
uc
ti
on of iteratio
n un
scent
ed particl
e
filter and
the analy
s
is
of comp
ari
ng
to other filteri
ng alg
o
ri
thm
s
verified the e
ffectiveness
of this alg
o
rit
h
m
in aco
u
sti
c
so
urce localization.
In the rest of this pa
per, we
explain our a
l
gorithm in Se
ction 2. Experimental re
sult
s and
analysi
s
are reporte
d in Se
ction 3. We
concl
ude thi
s
pape
r in Secti
on 4.
2. The Propo
sed Algori
t
h
m
2.1. Particle Filter
Particle filter is a filter
in
g metho
d
b
a
se
d on
Mo
nte Carlo
an
d re
cu
rsive
Bayesian
estimation.
In
re
ce
nt yea
r
s,
it ha
s
be
com
e
a
commo
n
tool
for
locati
ng
target
und
er
non
-line
a
r
or
non-Gau
s
sia
n
con
d
itions
[7,8,9]. The core ide
a
is
th
at using the
weig
hted su
m of a serie
s
of
rand
om
sam
p
les t
o
repre
s
ent th
e po
st
erio
r p
r
ob
ab
ili
ty density. Assumin
g
that
n
online
a
r
syst
em
dynamic m
o
d
e
l as follo
ws:
State equatio
n:
11
kk
k
x
fx
v
(1)
Measurement
equation:
kk
k
zh
x
n
(2)
Whe
r
e
k
x
is the system state,
k
z
is the ob
serve
d
state, the map
()
f
and
()
h
r
e
pr
ese
n
t
the system
state tran
sition model fun
c
tio
n
and mea
s
u
r
ement mod
e
l function,
1
k
v
and
k
u
are the
pro
c
e
ss n
o
ise and ob
se
rvation noi
se.
Let
,,
1
,
,
ii
kk
x
wi
N
rep
r
e
s
e
n
ts a set of rand
om weig
hted
sampl
e
(p
arti
cle
s
), wh
ere
i
k
x
is the
i
-th pa
rticle state in ti
me
k
, the corre
s
po
ndin
g
wei
ght value is
i
k
w
, there:
1
()
(
)
N
ii
kk
k
k
i
p
xY
w
x
x
(3)
()
()
()
()
()
1
-1
()
(
)
0:
1
1
:
(|
)
(
|
)
(|
,
)
ii
i
ii
kk
k
k
kk
ii
kk
k
p
Yx
p
x
x
ww
qx
x
Z
(4)
Whe
r
e
()
k
x
x
is
the unit impuls
e func
tion, that is
()
0
,
kk
x
xx
x
a
nd
()
1
x
xd
, particle se
ts sampl
ed
from pro
p
o
s
al distrib
u
tio
n
()
()
0:
1
1
:
(|
,
)
ii
kk
k
qx
x
Z
, whos
e
weig
hts satisf
y the normali
zation
con
d
ition
1
1
N
i
k
i
w
.
2.2. The Proposed
Algorithm
I
n
t
he ac
ou
st
ic
sou
r
c
e
locali
zat
i
on
sy
st
em,
p
a
rt
i
c
le f
ilt
er al
g
o
rit
h
m
sele
ct
s pri
o
r
probability density as pr
oposal di
stribution, th
is approach lost measur
ed value of the current
time, which le
ads to th
e wrong lo
cali
zati
on. Whil
e
un
scente
d
pa
rticl
e
filter (UP
F
)
algorith
m
use
s
the unsce
nte
d
Kalman filter to gen
erat
e pro
posal
di
stributio
n, wh
ich can be
well integrate
d
into
the late
st me
asu
r
em
ent inf
o
rmatio
n to i
m
prove
lo
cati
on
a
c
cu
ra
cy.
The detailed
descri
p
tion re
fer
pape
r [10-12]
, the UPF’s p
r
inci
ple will n
o
t be introdu
ced in thi
s
section. But UPF also can
not
accurately lo
cate the
spe
a
ke
r
whe
n
the filterin
g e
s
timation
pre
c
isi
on i
s
lo
wer. Thi
s
p
a
p
e
r
improve
s
the
UPF alg
o
rith
m by joinin
g i
t
erative Ka
lm
an filter
(IKF) algo
rithm to
revise
the
sta
t
e
mean an
d varian
ce up
date
d
by UKF, and optimize
s
t
he pro
p
o
s
al distrib
u
tion. IKF [13] is a kind
of maximum
poste
rio
r
ap
p
r
oa
ch,
whi
c
h
aims to
find
more
inform
ation ab
out u
s
ing th
e
kno
w
n
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3902 – 39
10
3904
state estimati
on app
roxima
tion of
ˆ
x
and
P
.
For any natural Numb
er
i
d
e
termin
ed by prior o
r
conve
r
ge
nce crite
r
ia, the m
easure
m
ent u
pdate form
ula
is:
11
ˆ
,
ii
x
xP
P
.Sequen
ce
{}
i
x
and
{}
i
P
are defin
e
d
as:
00
ˆ
,
x
xP
P
1
ˆˆ
((
)
(
)
)
ii
i
i
i
x
xK
z
h
x
H
xx
(5)
1
()
ii
i
PI
K
H
P
(6)
Whe
r
e
1
()
ii
i
i
KP
H
H
P
H
R
,
'
()
ii
Hh
x
,
R
is t
he mea
s
u
r
e
m
ent noi
se
covarian
ce
matrix,
I
is
a u
n
it matrix. It can get
ne
w m
ean a
nd
va
ri
ance by iterative
update. Combine
d
the
iterative thou
ght of formul
a (5
)
and
(6
)
with
UP
F
can
get iterated
unsce
nt
ed Ka
lman filter (IUPF
)
algorith
m
. Assume th
at the system a
s
formul
a
(1
) an
d (2), the p
r
o
c
e
ss n
o
ise a
nd mea
s
u
r
e
m
ent
noise are all
zero-m
ean
G
aussia
n
whit
e noise wh
ich are
not rel
a
ted to ea
ch
other, covari
ance
are
k
Q
and
k
R
, The improve
d
part
of the algorithm is de
scrib
ed as follo
ws:
(1) Initializ
a
tion. At the time
0
k
, gene
rat
e
sa
mple
s f
r
om the
pri
o
r di
strib
u
tio
n
0
{,
1
,
2
,
,
}
i
x
iN
.
0
00
()
ii
x
Ex
(7)
00
0
0
00
[(
)(
)
]
ii
i
i
i
T
PE
x
x
x
x
(8)
(2) Imp
o
rtant
sampli
ng
At time
0,
1
1
ii
kk
x
x
,
1,
2
,
,
s
iN
, use the UKF algo
rithm to updat
e the particl
e.
Select particl
e
0,
1
11
ii
k
kk
xx
.
,1
11
1
1
()
,
1
,
2
,
,
ii
i
jk
x
x
kk
kk
x
xn
P
j
n
(9)
,1
11
11
()
,
1
,
2
,
,
2
ii
i
x
xx
x
jk
kk
kk
x
xn
P
j
n
n
n
(10)
0
m
x
W
n
(11)
2
00
(1
)
cm
WW
(12)
1
2(
)
mc
jj
x
WW
n
(13)
Whe
r
e
()
m
j
W
and
()
c
j
W
are the wei
ght
s co
efficient o
f
the
first-ord
e
r stati
s
tical p
r
ope
rtie
s
and the seco
nd-o
r
d
e
r stati
s
tical p
r
op
erti
es.
Time Up
date:
,1
,1
()
ii
jk
jk
k
xf
x
(14)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Acou
stic Sou
r
ce L
o
calization Base
d on
Iterati
ve
Unscented Parti
c
le
Filter (Ji
e
Ca
o)
3905
2
()
|1
,1
0
x
n
im
i
kk
j
jk
k
j
xW
x
(15)
2
()
|1
|1
1
1,
1
,
1
0
x
n
T
ic
i
i
i
i
j
kk
k
k
k
kk
j
k
k
j
kk
j
P
Wx
x
x
x
Q
(16)
,1
,1
()
ii
jk
k
j
k
k
zh
x
(17)
2
()
|1
,1
0
x
n
im
i
kk
j
jk
k
j
zW
z
(18)
Measurement
update:
2
()
|1
|1
,1
,1
0
x
n
ci
i
i
i
xz
j
k
k
k
k
jk
k
j
k
k
j
PW
x
x
z
z
(19)
2
()
|1
|1
,1
,1
0
x
n
ci
i
i
i
zz
i
k
k
k
k
jk
k
j
k
k
j
PW
z
z
z
z
(20)
1
kx
z
z
z
KP
P
(21)
|1
|1
()
ii
i
kk
k
k
kk
kk
xx
K
z
z
(22)
1
ii
T
kz
z
k
kk
kk
PP
K
P
K
(23)
(3) Iterative update
Use IKF to
revise th
e
sta
t
e mea
n
a
n
d
varia
n
ce u
p
dated
by me
asu
r
em
ents,
get ne
w
mean
i
k
x
and variance
i
k
P
.
Particle sampling:
0:
1
1
:
ˆ
(,
)
(
,
)
ii
i
i
i
kk
k
k
k
k
x
qx
x
z
N
x
P
(24)
Whe
r
e
()
N
represents a Ga
ussi
an functio
n
.
Weig
hts calculating:
1
0:
1
1
:
ˆ
ˆ
()
(
)
ˆ
(,
)
ii
i
kk
k
k
i
k
ii
kk
k
pz
x
p
x
x
w
qx
x
z
(25)
Whe
r
e
1,
2
,
,
s
iN
, Nor
m
alize
d
wei
g
hts:
1
s
i
i
k
k
N
i
k
i
w
w
w
(26)
(4) Re
sampl
e
process.
(5) State e
s
timation:
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3902 – 39
10
3906
1
s
N
ii
kk
k
i
x
wx
(27)
IUPF fully in
tegrated into
the current
meas
u
r
em
e
n
t of the latest inform
ation, can
acc
u
rately locate the s
p
eaker.
3. Acous
tic
Source Loc
a
lization Base
d on IUPF
3.1.
Speaker Localiza
t
ion Function
SRP-PHAT
sound
source
localization
algorit
h
m
co
mbined
robu
stne
ss of
be
amform
method a
nd
sho
r
t analy
s
i
s
of characte
ristics
with
th
e insen
s
itivity to the enviro
n
ment of ph
a
s
e
transfo
rmatio
n metho
d
, re
duci
ng the
a
c
ou
stic
so
ur
ce localization system’
s
sensitivity to noise
and
reverberation, improving t
he
syst
em's
ro
bu
stn
e
ss an
d po
sitioning a
c
cu
racy [1
4]. The
algorith
m
sea
r
ch
es th
e ma
ximum output
energy of
the beam i
n
the
total spa
c
e, t
o
determine t
he
sou
nd
so
urce
location, u
s
e
a
sho
r
ter dat
a an
alysi
s
,
so this method is
very
suitable for real
-time
tracking the
movement of the spe
a
ker. The
sp
ecifi
c
calcul
ation
s
form as b
e
lo
w.
Assu
ming tha
t
the sound source si
gnal
()
s
t
rea
c
he
s the
microph
one a
rray thro
ugh t
he
multipath pro
pagatio
n. The
i-
th micropho
ne re
ceived
signal
()
i
mt
can be
expre
s
sed a
s
:
()
()
()
()
ii
i
mt
s
t
h
t
v
t
(28)
Whe
r
e
()
i
vt
is th
e noi
se,
()
i
ht
is th
e impul
se
re
spo
n
se bet
ween the
so
urce a
nd the
first microp
h
one, it is a f
unctio
n
of th
e microp
hon
e and the
so
und
sou
r
ce l
o
catio
n
. “*”i
s the
convol
ution o
perato
r
. Set the spa
c
e ve
ctor of the sound sou
r
ce is
,the Fourier transfo
rm of
spe
e
c
h
si
gn
al
()
i
mt
is
()
M
,the nu
mber
of micropho
ne
s is
M
,the propag
ation del
ay of
sou
nd
sign
al
to the
i
-th
microp
hone
in the
dire
ct path
is
i
.SRP-PHA
T sound
source
locali
zation
al
gorithm
is by
cal
c
ulatin
g th
e mi
cro
pho
ne
array
beam
output e
n
e
r
g
y
to po
sitioni
ng,
loc
a
liz
a
tion func
tion is
as
follows
:
*
()
*
11
()
(
)
()
()
(
)
SR
P
ij
MM
j
ii
t
ij
ii
MM
ye
d
MM
(29)
3.2. The Implementa
tion
of Propos
ed Algorithm
This p
ape
r
use the
audi
o sign
al ca
p
t
ured
by mi
croph
one a
r
ra
ys as o
b
servational
informatio
n to
locate the
sp
eaker. T
he I
U
PF al
gorith
m
frame
w
o
r
k nee
ds to
e
s
tablish the
mo
tion
model of the
spea
ke
r an
d the sp
ea
ker lo
cali
za
tio
n
function. T
h
is pa
pe
r use the Lan
ge
vin
pro
c
e
ss to e
s
tablish the
sp
eaker’
s
motio
n
model, the
detailed d
e
scription refers
pape
r [15]. The
impleme
n
tation step
s
of spe
a
ker
l
o
ca
lization ba
se
d on ite
r
atio
n un
scented
parti
cle filte
r
a
s
follows
,
Step 1:
Initial
i
zation: k=1, i
n
itialize particle sets
0
1
,,
1
,
,
i
x
iN
N
.
Step 2:
Particles sam
p
le.
1,
2
,
k
, obtain the sa
mpling pa
rticl
e
s at time
k
by usin
g
IUPF algorith
m
and sp
ea
ker motion mo
del and lo
cali
zation fun
c
tio
n
.
0:
1
1
:
ˆ
(,
)
(
,
)
ii
i
i
i
kk
k
kk
k
k
kk
x
qx
x
z
N
x
P
(30)
Step 3:
Weig
hts up
date.
Obtain p
a
rticl
e
weig
hts
i
k
w
at time
k
acco
rdi
n
g
to formula
(2
5)
and (2
6).
Step 4:
Re
sa
mple process.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Acou
stic Sou
r
ce L
o
calization Base
d on
Iterati
ve
Unscented Parti
c
le
Filter (Ji
e
Ca
o)
3907
Step 5:
State output. Finall
y
, particle
set
s
an
d correspondi
ng weig
hts can b
e
o
b
tained.
State estimation is
1
s
N
ii
kk
k
i
x
wx
.
Step 6:
Det
e
rmin
e wh
ether the
sp
ee
ch
signal
ge
ts an en
d. If true sto
p
s
runnin
g
,
otherwise go
to step 2.
4. Experimental Re
sults
and An
aly
s
is
4.1. Experimental Param
e
ter
s
Settin
g
s
The p
e
rfo
r
ma
nce
of the
propo
sed
a
c
ou
stic
so
urce l
o
cali
zation
alg
o
rithm
are
ev
aluated
in a
simul
a
te
d (
57
3
mm
m
) recta
ngul
ar
room,
as shown in
Fi
gu
re 1. In
two
dire
ction
s
of
X
and Y, the
two g
r
ou
ps
of linea
r arrays whic
h
containin
g
two microp
hon
es a
r
e l
o
cated
respe
c
tively, and the
dist
ance bet
wee
n
the mi
cro
p
hone
s in
ea
ch group i
s
1
m
, the sp
ea
ker
moves unifo
rmly along the x axis inclined 45° di
re
ctio
n in the room
, and keep
s spea
king du
rin
g
the moveme
nt, and the
starting
poi
nt of the mov
e
ment i
s
(1,1). Th
e reve
rbe
r
ation i
m
p
u
lse
respon
se fu
n
c
tion in the
room is g
ene
rated
by the IMAGE model
, using G
a
u
s
sian
white n
o
i
se,
the sp
ee
ch
signal i
s
obtai
ned by mi
cro
phon
e arra
y
sampl
ed
with
the sam
p
lin
g rate of fs=16
KHZ. The hei
ght of the spe
a
ke
r is
set to a con
s
tant va
lue.
Duri
ng the
experim
ent,
Th
e paramete
r
s in the spea
ker movem
ent
model a
r
e
1
10
s
x
，
1
1m
s
x
v
，
32
m
s
T
, the initial state of the
spe
a
ker is
0
1,
0
,
1,
0
a
,
with the c
o
varianc
e
of
2
0
00
00
01
0
0
P
00
00
00
0
1
,the tra
n
sf
orm l
ength
of
FFT i
s
L=51
2, win
d
o
w
fu
nction
is ham
ming
wind
ow. Sou
nd sou
r
ce mo
ving uniforml
y
along the x-axis inclin
ed
45° directio
n with the spe
e
d of
0.
1
/
ms
. To verify the perfo
rma
n
ce of IUPF al
gorithm
on th
e sp
ea
ker l
o
cali
zation, thi
s
pa
per
comp
are it with the PF an
d the UPF
which
us
ed in
pape
r [10], a
nd intro
d
u
c
e
the root m
e
a
n
squ
a
re
e
rro
r
(RMSE) a
s
a
stan
da
rd
of
pre
c
isi
on
me
asu
r
em
ent. T
he
root m
e
a
n
squa
re
error
(RMSE) i
s
de
fined as:
21
/
2
1
1
((
)
)
T
i
kk
t
RM
S
E
x
x
T
The sm
aller t
he RMSE, the highe
r the p
o
sition a
c
curacy.
Figure 1. The
Placeme
n
t of Micro
pho
ne
Array
4.2. The Experimental Re
sults and
An
aly
s
is
Und
e
r the
co
ndition of different SNR a
nd reve
rbe
r
at
ion time (
60
T
)
,
w
e
co
mpa
r
ed
th
e
perfo
rman
ce
of PF, the
UPF and
the
prop
osed
me
thod in
this
pape
r. Fig
u
re 2
sh
ows t
h
e
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3902 – 39
10
3908
locali
zation
e
ffect of thre
e
algo
rithm
s
i
n
X di
re
ction
wh
en
15
,
SNR
d
B
60
100
Tm
s
. Figure 3
sho
w
s t
he l
o
cali
zat
i
on
ef
f
e
ct
of
t
h
ree a
l
gorithm
s
in
Y
dire
ction when
15
,
SNR
d
B
60
100
Tm
s
.
Figure 4 shows the locali
zatio
n
effect
of three al
gorit
hms in X
directio
n whe
n
60
5,
2
0
0
SNR
d
B
T
m
s
. Figure
5
shows th
e lo
cali
zation
effect of
thre
e
algo
rithm
s
i
n
Y
dire
ction whe
n
60
5,
2
0
0
SNR
d
B
T
m
s
. The h
o
ri
zontal axis re
pre
s
ent
s tim
e
, the vertica
l
axis
rep
r
e
s
ent
s th
e po
sition in
each directio
n. Experimen
t has 5
0
time
s si
mulation
s re
spe
c
tively in
two differe
nt SNR a
nd reverbe
r
atio
n time (
60
T
), getting the averag
e values
of the RMSE a
s
sho
w
n in Ta
b
l
e 1 and Tabl
e 2.
F
i
gure 2.
60
15
,
1
00
SNR
d
B
T
m
s
, t
he Locati
on
Results of T
h
ree Algor
ithms in
X Directi
on
F
i
gure 3.
60
15
,
1
00
SNR
d
B
T
m
s
, t
he Locati
on
Results of T
h
ree Algor
ithms in
Y Direction
Figure 4.
60
5
,
200
SN
R
d
B
T
m
s
, t
he
Locatio
n
Results of T
h
ree Algor
ithms in
X Directi
on
Figure 5.
60
5
,
200
SN
R
d
B
T
m
s
, th
e Location
Re
sults of Th
ree Algo
rithm
s
in Y Dire
cti
o
n
Figure 2 and
Figure 3 sho
w
the lo
catio
n
re
sult
s of t
he PF, the UPF used i
n
p
aper [1
0]
and the p
r
op
ose
d
algo
rith
m unde
r the
con
d
ition of
h
i
gher
sig
nal n
o
ise
ratio (S
NR) an
d shorter
reverberation
time. Obvio
u
sly, the l
o
ca
tion resu
lts
o
f
PF an
d
UPF are
coa
r
se
and
imp
r
e
c
i
s
e.
The p
r
op
ose
d
algo
rithm
can a
c
hieve
p
r
eci
s
e l
o
calization of the
spea
ker
co
mp
ared to
PF a
n
d
UPF. Results indicate that
the p
r
opo
se
d metho
d
is
sup
e
rio
r
to P
F
and
UPF
a
l
gorithm
on t
he
locat
i
o
n
ac
cu
racy
in hig
h
S
NR.
Figure
4 an
d Figu
re 5
show th
e location re
sult
s of three al
g
o
rithm
s
a
s
the SNR
decrea
s
e
s
an
d reve
rb
erati
on time i
n
cre
a
se
s. Th
e lo
cation a
c
curacy of three
alg
o
rithm
s
redu
ced
in different d
egre
e
, the estimation error of PF
and UPF algorithm
increa
sed o
b
v
iously whil
e the
0
2
4
6
8
10
1
1.
5
2
2.
5
t/
s
x
/
m
Tr
u
e
x
PF
UP
F
IU
P
F
0
2
4
6
8
10
1
1.
5
2
2.
5
t/
s
y
/
m
Tr
u
e
x
PF
UP
F
IU
P
F
0
2
4
6
8
10
0
0.
5
1
1.
5
2
2.
5
t/s
x
/m
Tr
u
e
x
PF
UP
F
IU
P
F
0
2
4
6
8
10
-0
.
5
0
0.
5
1
1.
5
2
2.
5
3
t/s
y/
m
Tr
u
e
x
PF
UP
F
IU
P
F
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Acou
stic Sou
r
ce L
o
calization Base
d on
Iterati
ve
Unscented Parti
c
le
Filter (Ji
e
Ca
o)
3909
proposed
algorithm
can st
ill keep good locali
zati
on
accuracy. T
h
e experiment
s
show that t
he
stability of the propo
se
d me
thod is better
than PF and
UPF algo
rith
ms.
Table 1.
60
15
,
1
00
SNR
d
B
T
m
s
, C
ompari
s
o
n
of RMSE Value
Filtering algorith
m
RMSE
X Y
PF 0.0627
0.0578
UPF
0.0415
0.0451
Our algo
rithm
0.0333
0.0318
Table 2.
60
5
,
200
SN
R
d
B
T
m
s
,
Co
mpari
s
o
n
of RMSE Value
Filtering algorith
m
RMSE
X Y
PF 0.4429
0.6294
UPF
0.2570
0.4906
Our algo
rithm
0.1655
0.2119
By comp
ari
n
g the
mea
n
squ
a
re
e
rro
r
of Tabl
e1
an
d Ta
ble
2, th
e IUPF
algo
ri
thm ha
s
minimum RM
SE, the localization a
c
cura
cy is 50%
-6
0
%
higher tha
n
that of the
PF, 10%-30
%
highe
r tha
n
that of the
UP
F. It sho
w
s th
at the
p
r
op
osed alg
o
rithm
i
n
this
pap
er i
s
al
ways bett
e
r
than the algo
rithm in pape
r [10]
and the standard filter method.
In summ
ary, unde
r the
co
ndition of lo
wer SN
R an
d stronge
r reverberatio
n, PF and
UPF
algorith
m
ca
n appe
ar la
rge localizatio
n errors.
Ob
viously, our
algorith
m
gre
a
tly improves the
locat
i
o
n
ac
cu
racy
.
5. Conclusio
n
Analyzing th
e traditional
locali
zation
met
hod ba
se
d on particl
e
filter algorithm, an
effective loca
lization
algo
ri
thm is
pro
p
o
s
ed
in thi
s
p
aper which
combi
n
e
s
UKF and IKF
to
gene
rate th
e
pro
p
o
s
al
distribution fu
nction of
the
P
F
. Un
der the
frame
w
o
r
k
of the imp
r
o
v
ed
algorith
m
, thi
s
p
ape
r
co
nstructed
the
li
kelih
ood
fun
c
tion u
s
ing
S
R
P-PHA
T
. Si
mulation
re
sults
sho
w
that th
e localization
accu
ra
cy of t
he propo
sed alg
o
rithm
has
obviou
s
imp
r
oveme
n
t
comp
ari
ng to PF algorithm
and UPF al
go
rithm.
Ackn
o
w
l
e
dg
ements
This wo
rk was su
ppo
rte
d
by
the
National Natu
ral
Scie
nce Found
ation of
Chi
na
(612
630
31).
The Fina
nce Dep
a
rtme
nt Found
ati
on of Gan
s
u Provin
ce, Chi
na (0
9
14ZTB14
8).
Referen
ces
[1]
Jin NG, Yi
n F
L
, Ch
en Z
.
W
e
ig
hted s
ubs
p
a
ce fittin
g
so
u
nd s
ource
loc
a
lizati
o
n
meth
od
base
d
o
n
particl
e filterin
g
.
Journal of El
e
c
tronics & Informati
on T
e
ch
no
logy
. 20
08; 30(
9): 2134-
21
37.
[2]
J Vermaak, A
Blake.
No
nli
n
e
a
r filtering for speak
er trackin
g
in noisy a
nd
reverb
erant e- nviro
n
ments
.
Proc. IEEE Int. Conf. Acoust. Speec
h, Sign
al
Processin
g
(ICASSP-01), S
a
lt Lake C
i
t
y
, U
T
. 2001.
[3]
DB
Ward, RC Williams
on.
P
a
rticle filter
b
e
a
m
for
m
i
ng f
o
r a
c
oustic s
ource
local
i
z
a
ti
on
in
a rev
e
rber
ant
envir
on
me
nt.
Proc.IEEE Int. Conf. Acousti
c, Speach,
Sig
nal Proc
essin
g
(ICASSP-02), Orlando, FL.
200
3.
[4]
Hou DW
, Yi
n
F
L
. An IMM particle filt
erin
g
method for s
peak
er trackin
g
.
ACTA Elect
r
onic
a
Sinica
.
201
0; 38(4): 83
5-84
1.
[5]
Jin
NG, Yin
F
L
, Che
n
Z
.
Micr
oph
on
e arr
a
y
speak
er
tracki
ng usin
g la
ye
r
ed
s
a
mpl
i
n
g
p
a
rticle
filteri
ng.
ACT
A
Electron
ica Sin
i
ca
. 20
0
8
; 36(1): 19
4-1
98.
[6]
Jin NG, Yin
F
L
. Quantum
Evoluti
onar
y
P
a
rticle
F
ilt
erin
g Alg
o
rithm
w
i
th App
licati
o
n
to Spe
a
ke
r
T
r
acking.
Sign
al Process
i
ng
.
200
8; 24(6): 98
2-98
7.
[7] Gordon
N
，
S
a
lmo
nd D
，
Sm
i
t
h
A
.
Nove
l appr
oac
h to nonl
ine
a
r/no
n
-
Gaussia
n
Ba
yesia
n
state
estimation.
Proceed
ing
of Institute Electric
En
gin
eeri
ng. 19
9
3
; 140(2): 1
07-
113.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3902 – 39
10
3910
[8]
Gao Y, Yua
n
GW
, Z
hou H, et al. Object
D
e
scriptor Com
b
ini
ng Col
o
r
a
nd
D
e
tectio
n.
TELKOMNIKA
Indon
esi
an Jou
r
nal of Electric
al Eng
i
ne
eri
n
g
.
2013; 1
1
(6): 3
072-
307
9.
[9]
Indah A, Sira
dj
udd
in M, Rah
m
at W
,
T
Basarudd
in.
Particl
e
F
ilter
w
i
t
h
Gaussia
n
W
e
ighti
ng for Huma
n
T
r
acking.
T
E
LKOMNIKA Indones
ian J
ourn
a
l of Electrica
l
Engi
neer
in
g
. 2012; 10(
4): 801
-806.
[10]
Z
H
ANG MH, LIU XX. Man
e
u
v
erab
le targ
et tracking i
n
d
y
n
a
mic scen
e
ba
sed on
unsc
e
n
t
ed particl
e
filter.
Journa
l o
f
Optoelectroni
cs
Laser
. 201
0; 6(21): 924-
92
9.
[11]
PAN Q, YANG F
,
YE L. Surve
y
of
a kind of
nonl
ine
a
r filter
s-UKF
.
Control
and Dec
i
si
on
. 200
5; 20(5)
:
481-
489.
[12]
QU YW,
ZHA
N
G EH, YAN
G JY.
Im
proved unsce
nted p
a
rticle filter
.Control T
h
e
o
ry & Applic
ations
.
201
0; 9(27): 11
52-1
158.
[13]
Shmaliy
YS. An
Iterative Kalman-Lik
e Algor
it
hm Ign
o
rin
g
N
o
ise
a
nd In
itial
Co
n
d
itio
ns.
IEEE
T
r
ansactio
n
s o
n
Sign
al Proc
e
ssing.
20
11; 59
(6): 2465 –
247
3.
[14]
Dante
A B
l
aut
h ,Vice
n
te P
M
i
notto, C
l
a
udi
o
R Ju
ng,
B
o
w
o
n L
ee, T
on K
a
l
k
er. Voic
e
acti
vit
y
detecti
o
n
and sp
eak
er lo
calizati
on us
in
g aud
iovis
u
a
l
cues.
Pattern recognition letters
. 2012; 33(
4): 373-
380.
[15]
Hou
DW
, Yin
F
L
,T
ao Z
.
Quasi-Monte
Carl
o
filterin
g for s
p
eaker track
i
n
g
.
ACTA Automatica Sinic
a
.
200
9; 35 (7): 1
016-
102
1.
Evaluation Warning : The document was created with Spire.PDF for Python.