TELKOM
NIKA
, Vol.13, No
.3, Septembe
r 2015, pp. 1
014
~10
2
0
ISSN: 1693-6
930,
accredited
A
by DIKTI, De
cree No: 58/DIK
T
I/Kep/2013
DOI
:
10.12928/TELKOMNIKA.v13i3.1792
1014
Re
cei
v
ed Ma
rch 2
6
, 2015;
Re
vised Ma
y 20, 2015; Accepted June 1
0
, 2015
A New Selection Method of Anthropometric Parameters
in Individualizing Head-Related Impulse Responses
Huge
ng*
1
, Wahidin Wah
a
b
2
, Dadang
Guna
w
a
n
3
1
Departme
n
t of Computer En
g
i
ne
erin
g, Un
iv
e
r
sitas Multime
d
i
a Nus
antara (
U
MN),
Jl. Scientia Bo
ulev
ard, Gadi
n
g
Serpo
ng,
T
anger
ang
158
10
, Indonesi
a
, Ph
./F
ax: +
6221-5
422
08
08/00
2,3
Department of Electrical En
gin
eer
i
ng, Un
iv
ersitas Indo
nes
ia (UI),
Kampus Bar
u
UI, Depok 16
4
24, Indo
nesi
a
, Ph./F
ax: +
6221
-727
007
8/77
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: huge
ng@
um
n.ac.id*
1
, w
a
h
i
din.
w
a
ha
b@
ui.
a
c.id
2
, guna
@
eng.u
i
.ac.id
3
A
b
st
r
a
ct
A trend
issue
i
n
mod
e
li
ng
he
ad-rel
a
ted
i
m
p
u
lse r
e
sp
onses
(HRIRs) is
ho
w
to indiv
i
du
ali
z
e
HRI
R
s
mo
de
ls that ar
e conv
eni
ent for a parti
c
u
l
a
r
listener. T
he
obj
ective of thi
s
research
is to show
a ro
bu
st
selecti
on meth
od of
ei
ght an
thropo
metric
p
a
ra
meters
out
of al
l 2
7
p
a
ra
meters
d
e
fine
d
in
CIPIC HR
TF
Datab
a
se. The
prop
osed s
e
l
e
ction
meth
od is
systematic
al
ly
and sc
ientific
all
y
accepta
b
l
e
, compar
ed to
‘
t
ri
a
l
and error
’
met
hod in
s
e
lecti
n
g
the para
m
ete
r
s.
T
he
se
lect
e
d
anthr
op
o
m
etr
i
c par
a
m
eters
of a giv
en
liste
ner
w
e
re app
li
ed i
n
establ
ishi
ng
multipl
e
l
i
ne
ar re
gressi
on
mo
de
l
s
in or
der to
in
divid
u
a
l
i
z
e
his
/ her HRIRs. W
e
mo
de
lle
d
the e
n
tire mi
ni
mu
m phas
e
HRIRs i
n
hor
i
z
o
n
t
a
l p
l
ane
of 35 s
u
b
j
ects usin
g pri
n
cipal
co
mp
on
e
n
ts
ana
lysis (PCA
). T
he in
divi
d
ual
mini
mu
m
phas
e H
R
IRs
can
be
esti
mate
d a
d
e
q
u
a
t
ely by
a li
ne
ar
combi
natio
n of ten orthon
or
ma
l basis functi
on
s.
Ke
y
w
ords
:
HRIR mode
li
n
g
, HRIR indiv
i
dua
li
z
a
ti
on, multip
l
e
regr
essi
on an
alysis, p
r
incip
a
l co
mpo
nents
ana
lysis
Copy
right
©
2015 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
Hea
d
-related
impul
se
re
sp
onse i
s
the
i
m
pulse
re
spo
n
se
of a
hum
an e
a
r th
at functio
n
s
as
an
acou
stic filter of h
u
m
an a
uditory
system fr
o
m
a
sou
nd so
urce
to
the entran
c
e of
ea
r ca
n
a
l.
Two
main
cues in
locali
zing
the
dire
ction
s
of
so
und
so
urce
s on
the
ho
rizontal
pla
n
e
are
Interaural Time Difference (ITD)
and
Interaur
al
Le
vel Differe
nce (IL
D
) [1]. ITD a
nd IL
D
are
almost u
ndi
stingui
shabl
e o
n
the media
n
plane. Howe
ver, locali
zati
on of so
und
dire
ction o
n
this
plane is po
ssible by sp
e
c
tral modifi
ca
tion, mainly
due to reflection and diffraction on pi
n
nae
folds [2]. Filte
r
ing m
ona
ura
l
sou
nd u
s
in
g
huma
n
bin
a
u
ral
HRI
Rs creates Vi
rtual
Auditory Spa
c
e
(VAS) in virt
ual reality. This
depends
on the human
ps
y
c
hoacous
t
ic
c
h
ar
ac
t
e
ris
t
ic
s
.
Human
tends to find a convin
cing
spatial soun
d
sufficiently
using two e
a
r
cha
nnel
s. Co
ntrol of ITD, ILD,
and sp
ectral modificatio
n
i
s
signifi
cant
i
n
provid
ing
i
n
formatio
n of
so
und
source directio
n to
a
lis
tener in order to
create
VAS.
All these three primary s
o
und cu
es
are
enc
r
ypt
ed in HRIR.
The
Fouri
e
r-pai
r o
f
HRI
R i
n
fre
quen
cy d
o
ma
in is
kn
o
w
n
as hea
d-relat
ed
tran
sfer functio
n
(HRT
F).
Many re
se
arche
s
h
ad
shown
that
HRT
F vari
es among
sub
j
ects
due to
inter-i
ndivid
ual
differen
c
e
s
in
anthrop
omet
ric pa
ram
e
ters and
cha
nge
s in so
und
so
urces’ di
re
cti
ons [2, 3].
A seri
es
of empiri
cal m
e
asu
r
em
ents
of i
ndividual
HRT
Fs fo
r a
spe
c
ific li
st
ener
are
requi
red in synthesizi
ng perfect VAS system. These meas
urements will
inevitably grow
prohi
bitive, taking i
n
to a
ccount the
req
u
i
reme
nts of
speci
a
lized a
n
d expen
sive
equipm
ent a
nd
the meas
urement time s
p
ent. Commerc
ial VAS s
y
s
t
ems
are recently s
y
nthes
ized usually in an
inexpen
sive
way by u
s
in
g no
n-in
divid
ualized o
r
g
eneri
c
HRTF
s that i
gno
re
inter-individ
ual
differen
c
e
s
. T
he works in [
3
, 4], howeve
r
, sh
owe
d
tha
t
non-in
dividu
alize
d
HRTF
s, i.e. unsuitab
l
e
HRT
Fs ap
pli
ed to
a li
ste
ner,
suffer from di
stortio
n
s
su
ch
a
s
in
-hea
d lo
cali
zation
whe
n
u
s
ing
head
phon
es,
poo
r verti
c
a
l
effects, in
a
c
curate
late
ralizatio
n, and
wea
k
fr
ont-back di
stin
ction.
Thus,
it is e
s
sential to
d
e
velop
an in
dividualiz
atio
n metho
d
to
estimate
p
r
oper HRIRs
for a
listene
r which
is abl
e to p
r
ovide ad
equa
te sou
nd
cue
s
with
out ne
cessitating a
measurement
of
the
individual HRI
Rs.
A vast body of research
e
s
is devote
d
to
the individuali
z
ation of
HRTF in fre
quen
cy
domain
or
HRIR in tim
e
domain. A n
u
mbe
r
of HRTF in
dividu
alizatio
n met
hod
s have
b
een
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
A New Sele
ct
ion Method of
Anthropom
etric Param
e
ters in Indivi
dual
izing
…
(Huge
ng)
1015
prop
osed, su
ch
a
s
HRTF clu
s
terin
g
an
d
sel
e
ctio
n of
a few
most
repre
s
e
n
tative one
s [5], HRTF
scaling in fre
quen
cy [6], a
structu
r
al m
odel of
com
p
osition an
d d
e
com
p
o
s
ition
of HRTF
s
[2],
HRT
F d
a
taba
se
matching
[7], the b
ound
ary el
ement
method
[8], HRIR
su
bje
c
tive custo
m
izatio
n
of pinna resp
onses [9] an
d of pinna, h
ead, and
to
rso respon
se
s [10] on med
i
an plan
e, an
d
HRT
F p
e
rso
nalization
ba
sed
on
multi
p
le reg
r
e
ssi
o
n
an
alysi
s
(MRA) on
ho
rizontal
pla
n
e
[11].
Shin and Pa
rk [9] prop
osed HRI
R cu
stomizati
on m
e
thod ba
sed
on subje
c
tive tuning of only
pinna respon
se
s (0.2 m
s
out of entire HRI
R)
on m
e
dian plan
e u
s
ing PCA of
the CIPIC HRTF
Datab
a
se. Th
ey attained t
he custo
m
ize
d
pinn
a
re
sp
onses by
letting
a subj
ect tune
the weig
hts
on th
ree
ba
si
s fun
c
tion
s.
Hwan
g a
nd Pa
rk [10] follo
wed the
si
milar metho
d
a
s
i
n
[9], but they f
ed
PCA with the
entire medi
a
n
HRI
Rs; i.e. each
HRI
R wa
s 1.5 ms l
ong (6
7 sa
m
p
les) sin
c
e t
h
e
arrival
of dire
ct pul
se. T
h
is HRIR in
clu
d
ed the
pinn
a,
head, a
nd to
rso
re
spo
n
ses. They tune
d t
h
e
weig
hts of th
ree d
o
mina
nt basi
s
fun
c
tio
n
s
subje
c
tivel
y
accordi
ng to the thre
e large
s
t sta
n
d
a
rd
deviation
s at
each el
evatio
n. Hu
et al. [1
1]
perso
nali
z
ed the
e
s
tima
ted log
-
ma
gni
tude respon
ses
of HRTFs
by MRA. Firstly, the log-m
agnitud
e
re
spon
se
s were
approximate
d
usi
ng PCA
a
s
linear
co
mbi
nation of
we
ighted b
a
si
s function
s.
T
he weight
s
of the ba
sis function
s
were
sub
s
e
que
ntly approxim
ate
d
usin
g anth
r
opo
metri
c
p
a
ram
e
ters by MRA. Our i
ndividuali
z
ati
o
n
method
wa
s
better tha
n
the metho
d
i
n
[11],
beca
u
se we used
the
minimu
m
pha
se HRIRs
(H
RIR
s
mp
) i
n
time dom
ain t
o
be
mo
delle
d in P
C
A, an
d ou
r
anthrop
ometri
c p
a
ra
meters
sele
ct
ion
method was different. Modelin
g of HRIRs
mp
by PCA wa
s ba
sed on the fact that mode
ling
minimum
ph
ase
HRIRs
p
r
ovided
be
st
results am
o
n
g
othe
r
pre
p
roce
ssing
s
of HRIRs in
ti
me
domain, a
s
shown by Hug
eng et al. in [12].
This
re
sea
r
ch wa
s
a co
mpre
hen
sive
re
sea
r
ch in
fulfilling and
validating th
e goal
to
develop pa
ra
metric mo
del
s of HRT
Fs t
hat can
b
e
tuned ba
se
d on few numb
e
r of listene
r’s
own
anthro
pom
etries.
The
s
e anthro
pom
etries shoul
d p
r
ovide
crucia
l perceptu
a
l
psychoa
cou
s
tic
effects
on
sp
atial so
und. A
t
first, for PCA mode
lin
g, a be
st prep
ro
ce
ssi
ng a
nd
data type of
HRI
R
in time dom
a
i
n; and a b
e
s
t pre
p
rocessing
and d
a
ta type of HRTF in freq
ue
ncy dom
ain
were
found a
s
pu
b
lishe
d in [12]
. The be
st d
a
ta types
we
re minim
u
m
pha
se
HRIRs and ma
gnitu
de
HRT
Fs.
The individu
alizatio
n of HRI
Rs
mp
for sound
sou
r
ce
s on the hori
z
ontal pl
ane wa
s
explained
an
d investig
ate
d
in [13] that
use
d
t
he
sa
me individu
al
ization m
e
tho
d
as
one tha
t
is
explained
in
this pa
per. I
n
[14], individuali
z
ation of
magnitud
e
HRT
Fs
fo
r sound so
urce
s
on
hori
z
ontal
pla
ne
wa
s inve
stigated. Th
e i
ndividuali
z
ati
on m
e
thod
u
s
ed
was si
mil
a
r to
the
met
hod
use
d
h
e
re,
ex
cept th
at the
i
ndividuali
z
ati
on of
mag
n
itu
de
HRTFs wa
s d
one
in
freq
uen
cy do
mai
n
.
After individu
alizatio
n pro
c
ess in freq
ue
ncy dom
ain
wa
s finish
ed,
the individua
lized ma
gnitu
de
HRT
Fs
sho
u
l
d
be re
con
s
tructed b
a
ck to
time domain
to yield individuali
z
ed
HRI
Rs.
In this
researc
h
, the
entire
hori
z
o
n
tal HRIRs
mp
from
the ori
g
inal
HRI
Rs i
n
the
CIPIC
HRT
F
Datab
a
se
were
in
cluded
in
a
single
analysi
s
. Thu
s
, all h
o
rizontal
HRIRs
mp
s
h
ar
e
the
same
set of basi
s
functio
n
s, whi
c
h co
mpri
se t
he in
ter-in
d
ividual
variation a
s
well as the i
n
ter-
elevation vari
ation. The re
spon
se
s of first 1.5 ms of HRIRs
mp
, which contai
n the effects of pin
na,
head, an
d torso, we
re i
n
clu
ded in
PCA, as p
r
opo
sed by [
10]. This p
a
per p
r
e
s
ent
s an
individuali
z
ati
on metho
d
by developin
g
the st
ati
s
tical PCA m
odel b
e
twe
e
n
the sele
cted
anthro
pom
etric pa
ramete
rs and the
HRIRs
mp
in a different an
d novel way co
mpared to [9-11].
Section 2
de
scribe
s d
e
tail
s of the alg
o
rithm of
individuali
z
ation m
e
thod. Sectio
n 3 explain
s
t
he
sele
ction m
e
thod of in
dep
ende
nt and
d
epen
dent variable
s
of mul
t
iple reg
r
e
s
si
on mod
e
ls
a
nd
then el
abo
rat
e
s th
e p
e
rfo
r
man
c
e
of th
e propo
se
d
method th
ro
ugh
cal
c
ul
ation of th
e e
r
ror
betwe
en ea
ch measured
HRI
R and
correspon
ding e
s
timated HRI
R
.
2. Rese
arch
Metho
d
In this pa
per,
we u
nde
rline
d
the metho
d
for sel
e
ctin
g
eight anth
r
op
ometrie
s
o
u
t of all 27
anthro
pom
etries. Th
ese
selecte
d
a
n
thropomet
ri
c pa
ramete
rs,
to
gether with
minimum
ph
ase
HRI
Rs
were
establi
s
hed
into multiple reg
r
e
ssi
on
models in
orde
r to indi
vidualize a
given
listene
r’s HRI
R
s.
Figu
re
1
sho
w
s the
al
gorithm
st
ru
cture
of ou
r
HRIR i
ndividu
a
lization
meth
od.
The d
a
taba
se used
wa
s
measured a
n
d
provided
by CIPIC Interf
ace
Lab
orato
r
y of Unive
r
si
ty o
f
California at
Davis. F
r
om
Figure 1, we ca
n s
ee t
hat at first, the entire o
r
i
g
inal
HRI
Rs
on
hori
z
ontal
pl
ane of 3
5
su
bject
s
are p
r
oce
s
sed
by
cep
s
tral
anal
ysis to b
e
converted to
their
corre
s
p
ondin
g
HRI
Rs
mp
. T
he mea
n
of the entire HRIRs
mp
is the
n
cal
c
ulate
d
. In orde
r to achi
eve
a kind of data
from HRI
Rs
mp
that have zero me
an, the mean
is
su
btracte
d
from
each
HRI
R
mp
to
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No. 3, September 20
15 : 1014 – 10
20
1016
obtain
corre
s
pondi
ng mini
mum ph
ase
dire
ct impul
se re
spo
n
se (DIR
mp
). Th
ese DIRs
mp
are
a set
of minimum phase HRI
Rs
data with ze
ro empiri
cal
m
ean whi
c
h i
s
requi
re
d to obtain a basi
s
that
minimizes m
ean squa
re
error of the
appr
oximated
data. The
whol
e set of
DIRs
mp
a
r
e
then
inputted to
PCA, whi
c
h
th
en results i
n
basi
s
fu
nctio
n
s
or
prin
cip
a
l compo
nent
s (P
Cs) to
get
her
with their wei
ghts. The line
a
r co
mbinatio
n of weighted
PCs form
s e
s
timated DI
R
mp
. We applie
d
multiple line
a
r
re
gre
s
sion
(ML
R
)
as th
e method
to
individuali
z
e
the estim
a
ted DI
R
mp
. MLR
utilizes
weights of P
C
s (P
CWs) and anthropometri
c
parameters i
n
order to
provide
regression
coeffici
ents th
at later can b
e
applie
d to model DI
R
mp
of a new liste
ner.
The process
of reconst
r
u
c
tion to the desire
d
HRIR
model is
sho
w
n by da
she
d
lines in
Figure 1. A
model of DI
R
mp
that results from ML
R analysi
s
, based on th
e anthro
pom
etric
para
m
eters o
f
the listener, is
adde
d with
the mean of HRI
Rs
mp
to yield a model o
f
HRIR
mp
. Initi
a
l
left- and
right
-ea
r
time d
e
l
a
y (On
s
et in
Figure 1
)
du
e
to distan
ce
from sound
so
urce to e
a
r
drum
are in
se
rted
back to the
HRIR
mp
model,
resulting in t
he de
sired HRIR mo
del.
More d
e
tails
about
the minimum
pha
se HRIR,
and ML
R mo
deling a
r
e ex
plaine
d in the
following
sub
s
e
c
tion
s.
HR
IR
HR
I
R
mp
Ce
p
s
tr
a
l
an
al
y
s
i
s
-
DI
R
mp
w
i
PC
A
8 A
n
t
h
r
o
pom
et
r
i
c
m
eas
ur
e
m
e
n
t
s
ML
R
+
On
s
e
t
R
e
gr
es
s
i
on
Co
e
f
f
i
c
i
e
n
t
s
Figure 1. Pro
posed HRIRs Individualizat
ion Method [1
5]
2.1. Minimu
m Phase HRI
R
Kulka
r
ni et. al
[16] sug
g
e
s
ted that the p
hase of HRIR can
be ap
proximated by
minimum
pha
se. A system function,
H(z), of an HRIR, h(n
)
,
is
said to be mi
nimum ph
ase
if all poles and
zeros of H(z)
lie inside the
unit circle |z| =1. The mini
mum pha
se
HRI
R, h
mp
(n),
can be obtai
ned
throug
h the calcul
ation of real
ce
pst
r
um
of its origin
al
HRI
R, whi
c
h
has a
r
bitrary
pha
se. It can
be
said that the h
mp
(n) is the
removed initial time delay ve
rsion of HRIR, but both
kinds of HRIR
have the sa
m
e
magnitud
e
spe
c
tru
m
in the freq
uen
cy domain. Th
e real
cep
s
tru
m
, v(n), of HRI
R,
h(n
)
, is cal
c
ul
ated as follo
w,
v(
n)
=
R
e
{
F
1
D
{ln|
F
D
{h(n
)
}|}},
(1)
Whe
r
e ln an
d
Re{} de
note
respe
c
tively natural log
a
rith
m and the re
al part of a co
mplex variabl
e,
F
D
{} and
F
1
D
{} a
r
e the discret
e Fouri
e
r tra
n
s
form a
nd its
inverse re
sp
e
c
tively. This real ce
pstrum
is then weight
ed by the followin
g
wind
o
w
functio
n
, w(n), whi
c
h is gi
ven by:
0 if
n < 0,
w(n) =
1
if n =
0,
(2)
2 if
n >
0.
For a
ration
al
H(z), the
wi
ndo
w fun
c
tio
n
is a
co
mpl
e
x conj
ugate
inversi
on of t
he zero
s
outsid
e
the
u
n
it circle,
so t
hat a mi
nimu
m pha
se
HRI
R
i
s
p
r
ovided
. Hen
c
e th
e d
e
sired mi
nim
u
m
phas
e
HRIR,
h
mp
(n), is yiel
ded by:
h
mp
(n) =
Re{
e
x
p
(
F
D
{w(n).v(n)})}.
(3)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
A New Sele
ct
ion Method of
Anthropom
etric Param
e
ters in Indivi
dual
izing
…
(Huge
ng)
1017
2.2. Multiple Linear Regre
ssion Modeli
ng
Suppo
se th
at the relation
b
e
twee
n the
weights ve
cto
r
of the PC
i
(i=1,2,…,10)
i
n
azimuth
θ
of all subje
c
ts,
w
i
(
θ
) (3
5
x
1), and the corre
s
p
ondin
g
anthro
pom
etric pa
ram
e
ters,
X
(35x9
)
, as
follows
,
w
i
(
θ
) =
X . ß
i
(
θ
) +
E
i
(
θ
)
,
(
4
)
Whe
r
e
X
is t
he matrix co
mposed of a
35x1 colu
m
n
vector
with
all 1’s an
d 8 anthropom
etric
para
m
eters o
f
all subje
c
ts being a
naly
z
ed,
ß
i
(
θ
) i
s
t
he reg
r
e
s
sio
n
co
ef
f
i
cient
s
colum
n
v
e
ct
or
(9x1), and
E
i
(
θ
) is the e
s
ti
mation errors column vect
or
(35x1
)
. Th
e regressio
n
coeffici
ents a
r
e
found
by le
ast-squ
a
re e
s
timation. T
h
is i
s
p
e
rfo
r
med by
solving the
opti
m
ization
p
r
o
b
lem
min{E
i,n
(
θ
)}, whe
r
e E
i,n
(
θ
) is the n-th
depen
dent
variable’
s estimation e
rro
r. PCWs
and
anthro
pom
etric pa
ram
e
ters are
re
spe
c
tively
the model
’s de
pen
dent
and in
dep
end
ent variabl
es.
From Eq
uati
on (4
), the regre
s
sion
co
e
fficients
due
to i-th PCW in azim
uth
θ
,
ß
i
(
θ
), can
be
written a
s
,
ß
i
(
θ
) =
(
X
T
.
X
)
-1
.
X
T
w
i
(
θ
)
.
(5)
From Equ
a
tion (5
), in order to en
ha
nce the
per
f
o
rma
n
ce of the multiple li
near
reg
r
e
s
sion
model
s, it is
need
ed to
se
lect b
o
th d
e
p
ende
nt
an
d i
ndep
ende
nt v
a
riabl
es caref
u
lly. Correl
ation
analysi
s
is
u
s
ed to
sele
ct
the indep
en
dent varia
b
le
s in obtai
nin
g
more accu
rate an
d sim
p
ler
MLR mo
del, as explai
ned i
n
[13].
3. Experiments’ Re
sults
and Disc
uss
i
on
The CIPIC
HRTF Databa
se use
d
co
ntai
ns not only th
e mea
s
ured
HRI
Rs, but al
so some
anthro
pom
etric pa
ramete
rs for 4
5
subj
ects, in
clu
d
in
g the KEMAR man
neq
uin
with both
small
and la
rge pi
n
nae. The d
e
tail definition
s
of the a
ll 27 anthro
pom
etric pa
ramete
rs are given i
n
[1,
17]. Estimation of the listener’
s
o
w
n HRIRs via
his or her o
w
n a
n
throp
o
metri
c
paramete
r
s will
directly affect
the feasi
b
ility and
compl
e
xity of the
syst
em. It is
cl
early not advisable to introduce
all parameters into the model. Some useful info
rmati
on will be conceal
ed by t
he unnecessary
para
m
eters, whi
c
h re
sult
s in a worse regre
s
sion m
odel. Besid
e
s
, many para
m
eters are v
e
ry
difficult to be measured co
rre
ctly. Howe
ver, some
pa
ramete
rs of
8
subje
c
ts are not
available in
the datab
ase. Acco
rdi
ng to
our a
n
thropo
metric
param
eters sel
e
ctio
n,
8
s
e
le
c
t
ed
p
a
r
ame
t
e
r
s
ar
e
inclu
ded only
in 35 su
bje
c
ts.
The pe
rforma
nce
s
of the e
s
timated
HRI
Rs
on the h
o
r
izo
n
tal pla
n
e
were evalua
ted by
the com
p
a
r
ison of mea
n
-square e
rro
r o
f
the di
fferen
c
e
s
bet
wee
n
the estimate
d
HRI
Rs
and t
he
measured
HRIRs to
the
mean
-squa
re
erro
r of th
e
mea
s
u
r
ed
HR
IRs in
pe
rcentage,
whi
c
h is
defined by:
e
j
(
θ
) = 10
0 % x
|| h
j
(
θ
) -
ĥ
j
(
θ
)
||
2
/
|| h
j
(
θ
)
||
2
(
6
)
Whe
r
e
h
j
(
θ
)
is the j-th m
easure
d
HRI
R
mp
with azi
m
uth
θ
in h
o
rizontal pl
a
ne,
ĥ
j
(
θ
) is t
he
corre
s
p
ondin
g
e
s
timated
HRI
R
mp
of
h
j
(
θ
). If the e
r
ror is l
a
rg
er, the
perfo
rma
n
ce
of the
estima
te
d
HRI
R
mp
is worse,
where better locali
zation
results
will
be achieved
with sm
all e
j
(
θ
). The
avera
g
e
errors a
r
e dif
f
erent am
ong
subje
c
t
s
in the dat
ab
ase. The go
od p
e
rform
a
n
c
e o
f
the estimat
e
d
left-ear HRI
R
mp
of a subject is not alway
s
followed by
small e
rro
r of the right-ear
one
s.
We
have
accomplished
th
e mo
deling
o
f
HRI
Rs
mp
fr
om 3
5
su
b
j
ec
ts
, fo
r so
ur
c
e
s
on
th
e
hori
z
ontal
pl
ane
usi
ng P
C
A
with 10
basi
s
fu
nctio
n
s.
He
re th
e
avera
ge
error, a
s
define
d
by
Equation 6, a
ttained from t
h
is PCA mo
d
e
ling a
c
ro
ss
35 su
bje
c
ts a
nd so
urce
s o
n
the hori
z
o
n
t
al
plane
is 8.11
%. In the fol
l
owin
g pa
ra
g
r
aph
s,
we
wi
ll discu
s
s a
system
atic a
nd
sci
entifica
lly
accepta
b
le selectio
n pro
c
ess of 8 anth
r
opo
metri
c
p
a
ram
e
ters fro
m
all 27 para
m
eters define
d
in
[16], for individuali
z
ation of
HRIRs.
As the first st
ep, we individ
ualized PCA
model
s of HRIRs
mp
on the
hori
z
ontal pl
a
ne with
all 27 anthro
pometri
c pa
rameters u
s
in
g the ML
R. By looking carefully in the CIPIC HRTF
Datab
a
se, these pa
ram
e
ters
were only
complete
d for 35 su
bje
c
ts. The avera
ge error obtai
ned
in this
co
nditi
on is 11.8
5
%. A seri
es of
experim
ents
were the
n
pe
rforme
d u
s
in
g ea
ch
of 26,
25,
24, and 2
3
pa
ramete
rs.
We
sele
cted
k p
a
ram
e
ters ou
t of 27 param
eters
usi
ng combinatio
ns
of k
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No. 3, September 20
15 : 1014 – 10
20
1018
out of 27 parameters, wh
e
r
e k < 2
7
. The result
ed av
erag
e errors
are 12.4
6
%, 13.04%, 13.6
0
%,
and 14.1
0
%, respe
c
tively.
The average
error of
usi
n
g
23 param
ete
r
s could b
e
consi
dered no
t
drop
ping
sign
ificantly comp
ared to the u
s
e of
27 para
m
eters (14.1
0
% comp
are
d
to 11.85%).
Observing
th
e anth
r
op
om
etries data
b
a
s
e, the
r
e a
r
e
many pa
ra
meters, nam
ely x
4
, x
5
,
x
13
, d
8
,
θ
1
, an
d
θ
2
that coul
d not be sim
p
ly measu
r
ed
becau
se of the difficulty in determi
ning t
h
e
referen
c
e poi
nts and the
para
m
eters a
r
e so
sma
ll (only several millimeters) that one nee
d
s
a
very preci
s
e
i
n
stru
ment
s.
These
6 p
a
ra
meters
al
so
have wea
k
correlation
s
wi
th
maximum
ITD
(ITD
max
). The
co
rrelation
coeffici
ents b
e
twee
n ITD
ma
x
and each
of these p
a
rameters,
, are
0.161, 0.09
8
,
0.222, 0.3
97, 0.
243, a
nd 0.28
4 re
spe
c
tively. The expe
rime
nts of u
s
ing
21
para
m
eters b
y
excluding x
4
, x
5
, x
13
, d
8
,
θ
1
, and
θ
2
, resulted in ave
r
age erro
r of 15.29%.
We ma
de a
n
observation
on the regression
co
effi
cie
n
ts, ß, from
MLR. Th
e re
gre
ssi
on
coeffici
ents,
ß, have ve
ry
small
ne
ar
ze
ro val
u
e
s
a
n
d
very
sm
all
variation
s
fro
m
ML
R
between
PCWs
;
w
i
(
θ
), i =
1,2,...,10
where
θ
= azi
m
uth angl
es,
and ea
ch of x
11
, x
14
, x
15
, x
16
, and x
17
. Thus
,
these p
a
ra
m
e
ters
we
re al
so not in
clud
ed in t
he ML
R model. At this poi
nt, we achi
eved ave
r
age
error of 17.97
%.
The exp
e
rim
ents
of ML
R modeli
ng
were
co
ntinue
d furthe
r
by i
gnori
ng th
e
neck
and
tors
o items
(i.e. x
7
, x
8
, x
9
, and x
10
)
.
T
h
is
is
du
e
to o
u
r
o
b
s
e
r
va
tio
n
on
th
e
r
e
s
p
on
se
s
o
f
n
e
ck
and
torso
that
we
re
rep
r
e
s
ente
d
by the l
a
st
few
sa
mpl
e
s
in the
HRI
R,
whi
c
h h
a
v
e
v
e
ry
s
m
all n
e
a
r
zero value
s
. Thus, thei
r contributio
ns t
o
overa
ll HRI
R
co
uld be n
egle
c
ted. Ne
verthele
ss, to
rso
top width, x
9
, is repre
s
e
n
ted by neck width, x
6
, a
nd sh
oulde
r
width, x
12
, fr
om the cho
s
en
anthro
pom
etrie, while ne
ck d
epth, x
8
, is represe
n
ted by hea
d depth, x
3
, for the de
pth
cha
r
a
c
teri
stic. Individualization of
HRI
Rs
mp
u
s
in
g
MLR
with
12
paramete
r
s,
igno
ring
x
7
, x
8
, x
9
,
and x
10
, resul
t
ed in averag
e error of 20.
13%.
In the las
t
s
t
ep, we reserved 8
anth
r
op
o
m
etric
param
eters, i.e. x
1
,
x
3
,
x
6
,
x
12
, d
1
, d
3
, d
5
,
and d
6
, whi
c
h were
expl
ained
befo
r
e
in [13].
Correlation
an
al
ysis i
s
appli
ed to
determine
para
m
eters t
hat have
stro
ng correlatio
n with IT
D
max
. Applying th
e HRIRs
mp
of 35 subje
c
ts
for
sou
r
ces on
the h
o
ri
zontal
plan
e, indivi
duali
z
at
ion
u
s
ing
ML
R b
e
t
ween
the
ch
ose
n
p
a
ra
me
ters
and P
C
Ws,
w
i
(
θ
), provid
es the
avera
ge e
rro
r of 2
2
.22 %. Thi
s
re
sult is clo
s
e to
wh
en
we
employed
da
ta from 37
subje
c
ts, that is 22.50%
[1
3]. Here
we
came to the
opinio
n
that our
sele
ction met
hod
of anthropomet
ric pa
ramete
rs
bef
or
e
as i
n
[13
]
is fully conf
irmed
with th
e
sele
ction m
e
thod explai
ne
d in this p
a
p
e
r. T
abl
e 1 summari
ze
s t
he average e
rro
rs
ca
used
by
MLR mo
del
s usin
g variou
s numbe
rs of a
n
throp
o
metri
e
s explai
ned
above.
Table 1. Average Errors of MLR Mo
del
s Us
i
ng Vari
ou
s Num
b
e
r
s of
Anthropom
etries
No. of Anthr
opo
metries
Average Err
o
r (
%
)
27 11.85
23 14.10
21 15.29
16 17.97
12 20.13
8 22.22
Figure 2. Left- and Ri
ght-E
ar Errors of Subje
c
t
003 an
d Subject 16
3
in the Frontal
Hori
zontal
Plane
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
A New Sele
ct
ion Method of
Anthropom
etric Param
e
ters in Indivi
dual
izing
…
(Huge
ng)
1019
In the following
s
u
bs
ec
tions
,
the res
u
lt
s
of
HRIRs
mp
individuali
z
at
ion a
r
e
sh
own u
s
ing
data from
su
b
j
ect 00
3 and
subj
ect 1
63, to be
comp
are
d
to the re
sult
s attaine
d
in [11]. Here eig
h
t
anthro
pom
etric pa
ram
e
ters, x
1
,
x
3
,
x
6
,
x
12
, d
1
, d
3
, d
5
, a
nd d
6
were fe
d into ML
R m
odelin
g in o
r
d
e
r
to cal
c
ulate
regre
s
sion
co
efficients. T
h
e re
gre
s
sion
coeffici
ents
then we
re ap
plied
in
e
s
timating
the PCWs of DIR
mp
at each dire
ction on
the horizonta
l
plane of ea
ch subj
ect.
Figure 2
sho
w
s the l
e
ft- a
nd ri
ght-ear e
rro
rs of
azim
uth an
gle
s
in
the frontal
h
o
rizontal
plane of subj
ect 003 a
nd
subj
ect 16
3. The left- an
d
right-ear
erro
rs of subje
c
t 003 in the fro
n
tal
hori
z
ontal pl
a
ne are gen
erally good, nu
mberi
ng
bel
o
w
20% exce
pt at 2 azim
uth angle
s
. T
h
e
averag
e e
r
ror obtaine
d fro
m
the d
a
ta of
HRIRs of
left
ear of
subje
c
t 00
3 is 13.6
1
%, and th
at
of
right ea
r is
10.56%. As
can b
e
se
en
from Figu
re
2(a), the e
r
rors a
r
e com
m
only larg
er for
c
o
n
t
ra
la
ter
a
l
s
o
ur
ce
s th
a
n
fo
r
ips
ila
tera
l s
o
u
r
c
e
s
.
T
h
e
er
ro
rs
of s
u
b
j
ec
t 163
in
th
e fr
on
ta
l
hori
z
ontal
pla
ne
can
be
sai
d
worse
than
those
of
subj
ect 0
03. T
h
e
averag
e e
r
ror obtain
ed f
r
o
m
the data of HRIRs of left e
a
r of subj
ect 163 is
22.2
8
%
, while of right ear is 29.
22%. The so
urce
s
of extreme
contralate
ral
f
o
r ri
ght ea
r, i.e. -80
o
a
nd -65
o
, re
sult in
errors of 3
7
.8
2% and 5
0
.7
9%,
respe
c
tively. The u
n
sy
ste
m
atic b
ehavi
o
r of
wei
ghts of PCs in
the PCA
a
c
ro
ss subje
c
ts
and
across di
re
ctions
cau
s
e
s
di
fficu
lty for MLR to es
timate them.
The
estimat
ed mi
nimum
pha
se
HRIR of
su
bje
c
t 00
3
ca
n
well
ap
pro
x
imate
corre
s
p
ondin
g
measure
d
minimum ph
ase HRIR p
a
r
ticula
rly at the first 20 sample
s. Figu
re 3
sho
w
s the e
s
timated and
measured
HRIRs
mp
of bot
h left and
rig
h
t ear i
n
the
extreme lo
cat
i
ons
in the frontal
hori
z
ontal
pla
ne. The top,
middl
e, an
d b
o
ttom panel
correspon
ds t
o
azim
uth an
gle
-80°, 0
°
, and
80° re
sp
ectiv
e
ly.
Figure 3. Measu
r
ed a
nd E
s
timated Mini
mum P
hase HRI
Rs of Sub
j
ect 003 in th
e Frontal
Hori
zo
ntal Plane
4. Conclusio
n
Our p
r
o
p
o
s
e
d
sele
ction
m
e
thod of anth
r
opo
metri
c
p
a
ram
e
ters ex
plaine
d he
re,
whi
c
h is
simple,
sy
st
e
m
at
ically
and
scie
n
t
i
f
i
cally
ac
cept
a
b
le
, is fully confirmed with the
sele
ction met
hod
as in [13]. T
he re
sulte
d
eight anth
r
op
ometrie
s
were incorp
orate
d
in a si
mpl
e
and effici
e
n
t
individuali
z
ati
on metho
d
of the model of
minimum ph
ase
HRI
Rs
b
a
se
d on mult
iple re
gre
s
sio
n
analysi
s
. Thi
s
pro
p
o
s
ed i
ndividuali
z
ati
on method
showed better perform
an
ce
in the objective
simulatio
n
experim
ents tha
n
the perfo
rm
ance in [11] whi
c
h ha
s be
en discu
s
sed
in [13].
Ackn
o
w
l
e
dg
ements
The a
u
tho
r
s
gratefully tha
n
k to
CIPIC
Interf
ace La
b
o
rato
ry of
Ca
lifornia
Unive
r
sity at
Davis, USA for providing t
he CIPIC HRTF Datab
a
se.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No. 3, September 20
15 : 1014 – 10
20
1020
Referen
ces
[1]
Algaz
i VR, Duda RO,
T
hompson DP, Avend
ano C.
T
he CIPIC HRT
F
Databas
e
. Procee
din
g
s of IEE
E
Workshop o
n
Appl
icatio
ns of
Signa
l Proces
sing
to Au
dio
a
nd Acoustics (
W
ASPAA). New
P
a
ltz. 200
1:
99-1
02.
[2]
Algaz
i VR, Duda RO, Morrison RP, T
hompson DM.
Structural Co
mp
os
ition
a
nd Dec
o
mp
ositi
on o
f
HRTF.
Procee
din
g
s of IEEE Workshop o
n
Appl
icatio
ns
of
Signa
l Proces
sing to Au
di
o and Ac
oustics
(WASPAA). New
Pa
ltz. 2001:
103-
106.
[3]
W
enzel EM, A
rruda M, K
i
stle
r DJ, W
i
ghtma
n F
L
.
Loc
aliz
at
ion Usin
g
N
o
n
i
ndivi
du
aliz
ed Hea
d
-Re
l
ate
d
T
r
ansfer F
unctions.
T
he Jour
n
a
l of the Acous
tical
Soci
ety of Amer
ica (JASA
).
1993; 94: 11
1-12
3.
[4]
Blau
ert J. Sp
atial
He
arin
g:
T
he Ps
y
c
hop
h
y
s
i
cs of
Hu
man S
oun
d
L
o
caliz
atio
n
. Revised Edition.
Cambri
dg
e: T
h
e MIT
Press. 1
997.
[5]
Shima
da S, H
a
yash
i M, Ha
yashi S. A Clus
t
eri
ng Meth
od
for Soun
d Loc
alizati
on T
r
ansfer F
unctions
.
Journ
a
l of Aud
i
o Engi
ne
erin
g Society
.
199
4; 42(7/8): 57
7-5
84.
[6]
Middl
ebr
ooks
JC. Virtual Loc
alizati
on Impro
v
ed b
y
Scal
ing
Non-Ind
i
vid
ual
ized E
x
tern
al-
E
ar F
r
ansfe
r
F
unctions in
F
r
equ
enc
y.
The
Journal of the
Acoustica
l Society of America
(JASA).
1999;
106(
3): 149
3-
151
0.
[7]
Zotkin DN, H
w
a
n
g
J, Dura
is
w
a
mi R, D
a
vis LS.
HRT
F
Persona
li
z
a
ti
on Usi
ng Ant
h
rop
o
m
entric
Measur
e
m
ents
.
IEEE Workshop
on A
p
p
lic
ations
of Sig
n
a
l Proc
essin
g
to Aud
i
o
an
d Acoustics
(WASPAA). New
Pa
ltz. 2003:
157-
160.
[8]
Katz BF
G. Measurem
ent a
n
d
Ca
lcul
atio
n
of Indi
vi
du
al Hea
d
-Re
l
ate
d
T
r
ansfer
F
unctions Usin
g
a
Boun
dar
y
Ele
m
ent Mo
del
In
cludi
ng
the
Me
asurem
ent
and
Effect of Ski
n
an
d
Hair
Imp
eda
nce. P
h
D
Dissertati
on. Penns
yl
van
i
a: Gradu
ate Progr
a
m
in Acoustics
- Penns
ylv
ani
a
State Universit
y
; 19
98.
[9]
Shin
KH, Pa
rk Y. Enha
nc
ed V
e
rtical
P
e
rc
epti
on T
h
roug
h H
e
a
d
-R
elate
d
Impu
ls
e Res
p
o
n
se
Customiz
ation
Based
on P
i
nn
a Res
p
o
n
s
e
T
uning i
n
T
he Median
Plan
e.
IEICE Transactions
F
unda
menta
l
s
. 2008; E91-A(
1): 345-3
56.
[10]
H
w
an
g S, Park Y.
HRIR Customi
z
at
io
n in
the Media
n
Pl
an
e via Prin
cipal C
o
mp
on
ents Analys
is
.
Procee
din
g
s of
AES 31st Int.
Conf. on Ne
w
Directio
n
s in H
i
gh Res
o
luti
on
Audi
o. Lond
on
. 2007.
[11]
Hu H, Zho
u
J
,
Ma H, Wu
Z.
Head-R
e
lat
ed T
r
ansfer
F
unctio
n
Pers
o
nali
z
a
t
io
n Bas
ed o
n
Mu
ltipl
e
Regr
ession Analysis
. Proc
ee
din
g
s of IEEE
Int. Conf. on
Comp
utatio
na
l Intell
ige
n
ce
and S
e
cur
i
t
y
.
Guangz
ho
u. 2006; 2: 18
29-1
832.
[12]
Hug
eng, W
a
h
ab W
,
Gun
a
w
a
n
D. Effective
Pre
p
roc
e
s
s
ing in
M
o
d
e
l
i
ng He
ad-R
e
la
ted
Impu
lse
Resp
onses B
a
sed o
n
Princ
i
p
a
l Com
p
o
nent
s Anal
ysis.
Si
gna
l Process
i
n
g
: An Intl. Jou
r
nal (SPIJ).
201
0; 4(4): 201
-212.
[13]
Hug
eng, W
a
h
ab W
,
Guna
w
an D.
Enh
anc
ed Ind
i
vid
ual
i
z
ation
of Hea
d
-
R
elat
ed Impu
l
s
e Resp
ons
e
Mode
l i
n
H
o
ri
zontal
Pla
ne
Ba
sed
on M
u
ltip
le
Regr
essi
on A
nalysis
. Pr
oce
edi
ngs
of IEEE
2
nd
Int. Conf.
on Com
puter E
ngi
neer
in
g and
Applic
ations
(I
CCEA). Bali Isl
and. 20
10; 2: 2
26-2
30.
[14]
Hug
eng, W
a
h
ab W
,
Gun
a
w
an
D. Improve
d
Meth
od f
o
r I
ndivi
du
aliz
atio
n of
Hea
d
-Re
l
ated T
r
ansfer
F
unctions on Horizo
ntal
P
l
a
ne
Usi
ng Red
u
c
ed
N
u
mber
of Anthrop
o
metri
c
Measur
emen
ts.
Journal of
T
e
leco
mmunic
a
tions
. 20
10; 2
(
2): 31-41.
[15] Hug
eng.
Ind
i
v
i
du
aliz
ation
of Parametric Mode
ls of Head-
Rel
a
ted
T
r
ansfer F
unctions. PhD
Dissertati
on. D
epok: De
pt. Elect. Eng. - Universit
y
of Ind
o
n
e
sia; 20
11.
[16]
Kulkar
ni A, Isabell
e
SK, Col
b
urn HS.
On the
Mini
mu
m-Ph
a
s
e Approx
i
m
ati
on of He
ad-R
e
l
a
ted T
r
ansfer
F
unctions.
Proceedings
of IEEE Worksh
op on Applications
of Signal Processing t
o
Audio and
Acoustics (WASPAA).
Ne
w
P
a
ltz.
199
5: 84-8
7
.
[17]
CIPIC HR
TF Datab
a
se Fil
e
s. Re
l
ease
1.2. Septemb
e
r
23, h
ttp://
interface.ci
pic.
ucdav
is.ed
u
/CI
L_html/CIL
_H
RT
F
_database
.
htm. 2004.
Evaluation Warning : The document was created with Spire.PDF for Python.