TELKOM
NIKA
, Vol.11, No
.1, March 2
0
1
3
, pp. 95~1
0
6
ISSN: 1693-6
930
accredited by D
G
HE (DIKTI
), Decree No: 51/Dikti/Kep/2010
95
Re
cei
v
ed O
c
t
ober 2
6
, 201
2; Revi
se
d Ja
nuar
y 6, 201
3
;
Accepte
d
Ja
nuary 28, 20
1
3
Spectral-based Features Ranking for Gamelan
Instruments Identification using Filter Techniques
A
r
i
s
T
j
ah
yan
to
1
, Y
o
y
on K Supra
p
to
2
,
Diah
P Wu
landari
3
1
Information S
ystems Departm
ent,
2,3
Electrical Engi
neer
in
g Dep
a
rtment
Institut
T
e
knolo
g
i Sep
u
lu
h No
pemb
e
r, Surab
a
y
a, Indo
nesi
a
e-mail: atj
a
h
y
a
n
to@gm
a
il.co
m, yo
yo
nsupr
a
p
to@gm
a
il.co
m, diah_
bas
uki
@
yah
oo.com
Abs
t
rak
Pada pa
per in
i,
ka
mi me
nj
el
askan up
aya dal
a
m
men
ent
ukan
r
ank
ing
fitur
berb
a
sis spektra
l
den
ga
n me
ma
nfaatkan tek
n
i
k
filter
yang d
i
gun
aka
n
untuk
identifik
as
i i
n
strume
n ga
mel
an Jaw
a
. Mod
e
l
ya
n
g
d
i
p
a
k
a
i
m
e
n
g
e
kstra
ksi se
ke
lom
p
o
k
fi
tu
r b
e
r
b
a
s
i
s
sp
e
k
tra
l
d
a
r
i
si
n
y
a
l
sua
r
a
g
a
m
e
l
a
n de
n
gan
m
e
n
g
g
u
n
a
k
an
Sh
o
r
t Tim
e
Fou
r
ie
r Tra
n
s
fo
rm
(STFT). R
a
n
k
in
g
da
ri fitu
r d
i
te
n
t
u
k
an
d
e
n
g
a
n
mem
a
n
f
a
a
t
kan
li
ma
alg
o
rit
m
a,
yaitu
Rel
i
efF
,
Chi-S
quar
ed,
Informati
o
n
Gain, Gai
n
R
a
ti
o, da
n Sy
mmetric Unc
e
rtain
t
y.
Sela
njutny
a k
a
mi
men
g
u
ji r
anki
ng fitur s
e
cara v
a
li
dasi
silan
g
d
eng
a
n
men
ggu
nak
an Su
pport V
e
ctor
Machi
ne (SVM
). Eksperimen
me
nu
njukk
an
bahw
a al
gor
it
ma Ga
in Rati
o
me
mb
erik
an
hasil ter
baik, y
a
itu
me
ng
hasi
l
ka
n akuras
i sebes
a
r
98.93%.
Ka
ta
k
unc
i:
su
pport vector
machi
ne, transkri
p
si ot
o
m
atis, Gain R
a
tio, ekstraksi fitur
A
b
st
r
a
ct
In this p
aper,
w
e
descri
be
an
ap
proac
h of s
pectr
al-
base
d
f
eatures
rank
in
g for Jav
a
n
e
se
ga
me
la
n
instru
me
nts id
entificati
o
n
usi
ng fi
lt
er tec
hni
ques. T
h
e
mo
del
extracte
d s
pectral-
base
d
f
eatures
set
of
th
e
sig
n
a
l
u
s
ing
Sh
o
r
t Tim
e
Fo
u
r
ie
r Tran
sfo
rm (STFT). Th
e
ra
n
k
o
f
the
fea
t
u
r
e
s
wa
s de
te
rm
in
e
d
u
s
in
g th
e five
alg
o
rith
ms; n
a
m
e
l
y R
e
li
efF
,
Chi-Sq
uar
ed, I
n
formatio
n
Ga
i
n
, Gain
Rati
o,
and
Sy
mmetric
Uncert
ainty. T
hen,
w
e
tested th
e
ranke
d
fe
ature
s
by cr
oss va
li
datio
n
usin
g S
upp
ort Vector
Machi
ne (SVM
). T
he ex
peri
m
en
t
show
ed that Gain R
a
tio al
gor
i
t
hm g
a
ve the
b
e
st result, it yielde
d accuracy
of 98.93%.
Ke
y
w
ords
: su
pport vector
machi
ne, auto
m
atic transcr
ipti
o
n
, Gain Ratio, features extract
i
on
1. Introduc
tion
Feature sele
ction is a pro
c
e
ss of
fi
n
d
in
g an optimal feature sub
s
et, removes i
rrel
e
vant
or red
und
ant
feature. Feature sele
ction is
one of
the importa
nt steps in machi
ne lea
r
ning
esp
e
ci
ally for re
cognitio
n
tasks. The
perform
an
ce of recog
n
i
t
ion algorith
m
s are u
s
u
a
lly
depe
ndent
o
n
the q
uality of the featu
r
e
set. If the
fe
ature set co
n
t
ains red
und
ant
or
irrelev
ant
feature
s
, the algorithm m
a
y produ
ce
a less a
c
cu
rate or a less reco
gnition
rate. The fea
t
ure
sele
ction p
r
o
b
lem ha
s b
e
en stu
d
ied b
y
the statisti
cs
and m
a
ch
ine lea
r
nin
g
comm
unitie
s
for
many yea
r
s [
1
-4]. Th
e fea
t
ure
sele
ction
algo
rithm
s
can b
e
catego
rize
d a
s
fi
lte
r
, w
r
ap
p
e
r
,
and
embed
ded m
e
thod
s ba
sed
on the criteri
on functio
n
s.
Filter metho
d
s u
s
e
s
stati
s
tical p
r
o
perti
es
for evalu
a
tin
g
feature
su
bset
s. The
a
d
vantage
s of
fi
lters m
e
th
ods are fa
st and
efficien
t to
pro
c
e
ss
high
dimen
s
ional
dataset
s, h
o
weve
r
fi
lte
r
s app
ro
ach do not consi
der the feat
ure
depe
nden
cie
s
. Wrapp
er
method
s u
s
e
a learning
al
gorithm fo
r e
v
aluating the
sele
cted fe
a
t
ure
sub
s
et
s. Em
bedd
ed m
e
thod
s a
r
e
si
milar to
wr
a
pper meth
od
s, but l
e
ss
comp
utationa
lly
expen
sive and co
nsid
eri
ng feature d
epen
den
cie
s
[5]. Feature extraction can be viewe
d
a
s
fi
ndi
ng a sub
s
et of ra
w dat
a while redu
ci
ng the dimen
s
ion
a
lity.
Many algo
rithms h
a
ve be
en develo
p
e
d
to perfo
rm
audio featu
r
e extractio
n
; comm
on
method
s
su
ch a
s
tem
p
o
r
al
ba
sed
an
d
spectral
ba
sed
usi
n
g
Fa
st F
ourie
r
Tra
n
sf
orm
(FF
T
), S
hort
Time Fou
r
ier
Tran
sfo
r
m (S
TFT), Di
scret
e
Wave
let Transfo
rm (DWT), and Conti
nuou
s Wavel
e
t
Tran
sfo
r
m (CWT). T
here
are va
riou
s f
eature
s
h
a
ve
been
pro
p
o
s
ed fo
r au
dio
sign
al, su
ch
as
zero crossin
g
rate, RMS e
nergy, envelo
pe,
an
d spe
c
trum rep
r
e
s
e
n
tation
[6]. We
u
s
ed
a
set
o
f
spe
c
tral
-b
ase
d
featu
r
e
s
whi
c
h ha
s been
p
r
ev
io
usly d
e
velop
ed fo
r g
a
m
e
lan i
n
st
rum
ents
identi
fi
c
a
tion [7].
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 11, No. 1, March 2
013 : 95 – 10
6
96
There are
several a
p
p
r
o
a
ch
es
ha
s
been
develo
ped to e
s
ti
mate the pi
tch an
d
instru
ment
s i
n
the auto
m
atic mu
si
c
transcri
p
tion. We can use aut
ocorrelation func
tion [8] for
identifying
hi
dden pe
riodi
cities
in a
time
-dom
ain sign
al.
The auto
c
orrelation
fu
n
c
tion sho
w
s
t
h
e
pea
ks pe
riodi
city in a
sig
n
a
l.
Suprapto
et al [9] [10]
introd
uced
a
method
to g
enerate m
u
si
c
transcription
for gamel
an
usin
g sp
ectra
l
density
mo
del to extract
the waveforms of gamel
an
instru
ment
s sound u
s
in
g Adaptiv
e Cross Correlatio
n (ACC).
Another tech
nique
is patt
e
rn
re
co
gniti
on a
ppr
oa
ch
that requi
re
s a
set of fe
ature
s
to
identify the musi
cal in
struments [1
1] [12]. T
he co
mmon featu
r
es that ne
ed
ed for reco
g
n
ition
pro
c
e
s
s such
as pitch, fre
quen
cy mo
d
u
lation,
sp
e
c
t
r
al e
n
velop
e
, sp
ectral
cen
t
roid, inten
s
it
y,
amplitude e
n
v
elope, ampli
t
ude modul
ation, ons
et asy
n
ch
rony an
d in harm
oni
city.
The g
oal of
this pa
pe
r i
s
to get the
mi
nimal sp
ectral
-b
ased feature
s
sub
s
et
that
extracted
fro
m
gam
elan
reco
rdin
g u
s
in
g STFT. T
he
sele
cted
feat
ure
s
sub
s
et t
hen valid
ated
by
cro
s
s-vali
dati
on te
chni
que
s u
s
in
g
sup
p
o
rt ve
ctor
ma
chin
e
(SVM).
The
r
e
are t
w
o
main
re
asons
for ad
dre
s
sin
g
this ta
sks
u
s
ing SVM. Fi
rst, a
c
curate
recognitio
n
of
gamela
n
in
st
rume
nt is itse
lf
an importa
nt for automati
c
transcriptio
n
. Second,
be
cause of the effectiv
eness
of SVM [13]
[14]
and re
cently became one
of
the
mo
st
p
opula
r
recog
n
ition o
r
cla
s
si
fi
cat
i
o
n
met
hod
s.
S
V
M
h
a
v
e
been
used i
n
a va
riety of appli
c
at
io
ns su
ch as text
classi
fi
cation [15], f
a
cial
expression
recognitio
n
[16], gene anal
ysis [17], [18] and many others.
In the pro
posed app
roach, Java
nese gamel
an instrume
nts identi
fi
cation is
accompli
sh
ed
throu
gh id
e
n
ti
fi
cation of
individual
b
l
ade
s or keys u
s
ing
an
SVM cla
ssi
fi
er.
Javan
e
se ga
melan i
s
an
ensemble
of percu
ssi
on
in
strum
ents th
at mostly me
tallopho
ne [1
9],
xylophone
s, and go
ng type instru
ment
s which pro
d
u
c
e tone
s whe
n
stru
ck with
horn o
r
wood
en
mallets.
A co
mplete set of
gam
elan
con
s
ist of
72
in
st
rume
nt [20], f
o
r
example:
kenda
ng,
saro
n
grou
ps, bo
na
ng gro
u
p
s
, kethuk-keno
ng
and gon
gs.
Grou
p of sa
ron co
nsi
s
t of demung,
saron,
and p
e
ki
ng.
Those in
stru
ments
play th
e co
re m
e
lod
y
or bal
ung
a
n
gen
dhin
g
. Gamela
n i
s
o
ne of
percu
ssi
on ty
pe mu
si
cal i
n
stru
ment
s
which
do
not
prod
uce h
a
rmonic soun
d
s
[21].
Ho
we
ver,
becau
se of the han
dmad
e prod
uctio
n
, gamelan
still produ
ce th
e freque
nci
e
s of non-i
n
te
ger
overtone [22]
. The freque
ncy ra
nge of
saron gr
oup
s [7] can b
e
seen at Ta
ble 1. Individual
gamela
n
pitch are
som
e
times difficult to identify due to their o
v
erlappi
ng in
freque
ncy, fo
r
example fund
amental fre
q
u
ency of sa
ron
‘1’ equal
s to that of demun
g ‘1H’.
The rest of th
is pap
er i
s
organi
zed a
s
fo
llows. Sectio
n 2 de
scribe
s the re
sea
r
ch
method
how to get the optimal spe
c
tral
-ba
s
e
d
feature
sub
s
ets. Section 3 prese
n
ts ou
r experim
ents an
d
discu
ss the result
s. Finally, Section 4
gives co
ncl
u
si
o
n
s of ou
r experime
n
ts.
Table 1. Saro
n grou
p freq
u
ency ra
nge
Key
s
Funda
m
e
n
t
al Fr
equency (
H
z)
De
m
ung Saron
Pe
k
i
ng
6L
231
463
925
1
267
533
1062
2
307
613
1225
3
349
698
1400
5
402
805
1599
6
463
925
1858
1H
533
1062
2158
2H
613
1225
2477
Figure 1.
Flo
w
chart of the
resea
r
ch method
2. Rese
arch
Metho
d
A gene
ral vi
ew of th
e
fl
o
w
chart
of the
pro
p
o
s
ed
sy
stem i
s
d
epi
cted in
Figu
re 1. Th
e
output of th
e propo
se
d
system i
s
th
e select
ed
f
eature
sub
s
e
t
for ide
n
tifying the
gam
e
l
an
instru
ment
s. The
fi
rst sta
ge in o
u
r
propo
sed
syst
em is
preproce
s
sing. Be
fore a
gam
e
l
an
recording
is
subj
ecte
d to
the p
r
op
osed
metho
d
s, it
is p
r
e
p
ro
ce
ssed in
some
way in
o
r
de
r to
make the foll
owin
g task ea
sier.
The
p
r
ep
ro
cessing co
nsi
s
ts
of noi
se redu
ction,
l
o
w-p
a
ss
fi
lteri
ng, and
sam
p
ling
rate
conve
r
si
on. T
he
se
cond
st
ep i
s
to
creat
e time-f
requ
e
n
cy rep
r
e
s
ent
ation o
r
spe
c
trogram from
a
gamela
n
re
co
rding. Th
e 2D matri
c
e
s
spectrog
ra
m o
f
the given gamelan
recording is calcul
ated
Pre
Pro
cessi
ng
Gamelan
So
un
d
ST
FT
O
n
set
D
e
t
ect
io
n
&
S
e
g
m
e
n
ta
ti
on
Feat
ur
es
Ext
r
act
i
o
n
F
eatu
r
es
Ra
n
k
i
n
g
Cross
V
a
lida
t
io
n
Feat
ur
e
s
Sub
sets
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Spectral-b
ased Featu
r
e
s
Ran
k
in
g for Ga
m
e
lan Inst
rum
ents…
(Aris Tjah
yanto
)
97
by the sh
ort-time Fou
r
ie
r tran
sfo
r
m
(STFT)
usi
n
g Ham
m
ing
wind
ow
wit
h
win
d
o
w
si
ze
approximatel
y 2048 sam
p
l
i
ngs an
d hop
size 6%.
Before extra
c
ting the fea
t
ures
set, se
gment
ation i
n
the time-freque
ncy do
main wa
s
perfo
rmed. T
he process o
f
segme
n
tation for t
he ti
me-fre
que
ncy
rep
r
e
s
entati
on re
quires
note
onset inform
ation. Note
onset can b
e
detecte
d u
s
ing
sud
den
chan
ge
s of acou
stic e
n
ergy
approa
che
s
[24]. In the case of stron
g
gamela
n
not
e, this abrupt
energy ch
an
ging will be very
sha
r
p. We
ca
n find the on
set location u
s
ing the
p
e
a
k
dete
c
tion fu
nction [25]. The feature
s
set
then cal
c
ul
ated ba
sed o
n
the seg
m
ent
ed sp
ect
r
og
rams Th
e feat
ure
s
set sho
u
ld co
ntain u
s
eful
informatio
n for id
entifying
and
differen
t
iating gam
el
an in
strum
e
n
t
s. In this
pa
per,
we
use
d
34
feature
s
fo
r
gamela
n
in
st
rume
nts i
den
ti
fi
catio
n
ta
sks. The
featu
r
es [2
6] have
bee
n calcul
ated
and ad
dition
al feature
s
h
a
ve been
extracted i
n
cl
u
d
ing the
statistical p
r
o
perties like m
e
an,
varian
ce of the spe
c
tral e
n
v
elope.
W
e
co
mp
ar
ed
th
e
fi
ve feature ra
nki
n
g
algorithm
s of the
fi
lters appro
a
ch. They are
Information Gain,
Gain Ratio,
Chi
-
squared,
Symmetrical Un
certainty and
Relief. Ran
k
in
g
algorith
m
s p
r
odu
ce a ran
k
ed list, acco
rding to the
e
v
aluation of criterio
n functi
on. For the
sake
of perform
an
ce compa
r
iso
n
, we also
co
nsid
er
the cross validation
accura
cy. We cal
c
ulated t
h
e
cro
s
s v
a
lidat
i
on ac
cu
ra
cy
in t
e
rms of
S
V
M
clas
si
fi
er
.
2.1. Time-fre
quencie
s An
aly
s
is
The go
al of
automatic
ga
melan tran
scripti
on i
s
to
extract the
seque
nce of
gamela
n
notes f
r
om
g
a
melan
re
co
rding. Ga
mela
n note
s
a
r
e
any syste
m
t
hat re
pre
s
e
n
ts the
pitch
of a
gamela
n
so
u
nd. This pa
p
e
r is pa
rt of the proj
ec
t ai
ms to develo
p
a system that extracts
note
events from g
a
melan
sou
n
d
s spe
c
trog
ra
m.
Spectrogram is
a spe
c
tro
-
t
e
mpo
r
al
represe
n
tation of
the so
und.
Spectrogram
provide
s
a time-freq
u
e
n
cy port
r
ait of gamelan so
und
s.
The STFT has b
e
e
n
the commo
nly used met
hod
for g
ene
ratin
g
time-f
req
u
e
n
cy
rep
r
e
s
en
tations
or sp
ectro
g
rams o
f
musi
cal
si
g
nal. Th
e
re
su
lt of
STFT can be
plotted on a 2D or 3D
sp
ectro
g
ram
(a
s sh
own in Fi
gure 2
)
as a
function of time
and fre
que
ncy, and magni
tude is
rep
r
e
s
ente
d
a
s
the height of a
3D surfa
c
e
spe
c
tro
g
ram
or
intensity in
2
D
spe
c
trogra
m
. Ho
weve
r, STFT
su
ffe
rs from th
e
common
sho
r
tcomin
g th
at the
length
of the
win
d
o
w
det
ermin
e
s the t
i
me an
d fr
eq
uen
cy re
sol
u
tion of th
e
sp
ectro
g
rams [27]
[21].
The
size of t
he wi
ndo
w
u
s
ed
for STF
T
is
relate
d to
the time
re
solution a
nd freque
ncy
resolution. If
we
apply a
sho
r
t wi
ndo
w, we
will
h
a
ve goo
d time re
sol
u
tio
n
. Ho
weve
r, if we
impleme
n
t a long wi
ndo
w, we will get hi
gh frequ
en
cy resolution b
u
t low time re
solution. For pi
tch
analysi
s
su
ch a
s
autom
atic g
a
mel
a
n
note
tran
scription, th
e f
r
equ
en
cy re
solutio
n
of t
h
e
spe
c
tro
g
rams is mo
re im
p
o
rtant tha
n
th
e time re
sol
u
tion [21]. The
n
STFT
with
long
wind
ow
is
good e
nou
gh
for automati
c
gamela
n
note
transcri
p
tion.
Evaluation Warning : The document was created with Spire.PDF for Python.
TEL
K
98
2.2.
S
give
seg
m
the
m
seg
m
dete
r
enve
l
the
m
gam
e
ener
g
rec
o
r
to id
e
clo
s
e
cor
r
e
from
prom
valid
by a
n
K
OM
NIKA
V
S
egmentati
o
Segment
significant
e
m
entation
m
e
m
agnitud
e
o
m
entation
tas
r
mi
ned at
ti
m
l
op
e. An on
s
m
oment whe
n
It is
pos
s
e
lan inst
rum
g
y produ
ce
s
r
ding h
a
s a
s
e
ntify the b
o
e
proximity s
u
A
fter get
e
sp
ondin
g
to
different fr
e
i
nent on
es
a
onset if thei
r
n
onset dete
c
V
ol. 11, No.
1
Figure 2.
o
n
atio
n is a
n
e
ffec
t
for ins
e
th
od
ba
se
d
o
f energy e
x
k can be
vie
m
e-freq
uen
c
y
s
et ca
n b
e
d
n
a ne
w not
e
i
b
le to
distin
en
ts
(s
uch
a
ha
rd o
n
set
s
s
et of featur
e
o
u
nda
rie
s
b
e
u
ch a
s
not
e
s
ting the
sp
e
energy cha
e
quen
cy ch
a
a
re c
o
n
s
id
er
e
r
amplitu
d
e
a
c
tion al
gorit
h
1
, Marc
h 20
3D surfa
c
e
s
impo
rtant p
r
tru
m
ent
s re
c
on
on
se
t in
f
x
cee
d
s a lo
c
wed a
s
a
pr
o
y
domai
n by
l
e
fi
ne
d as
th
e
begi
ns [
27]
gu
ish d
i
ffe
r
e
a
s dem
un
g,
s
that a
r
e
s
e
s t
hat
so
m
e
e
tween the
n
s
candi
date
f
o
Figure 3. T
h
e
ct
rog
r
ams
f
ng
es a
r
e d
e
a
nnel a
c
co
r
e
d a
s
note
o
n
a
bove a gl
o
b
h
m will be us
e
13 : 95 – 1
0
s
pe
ct
rog
r
a
m
r
ocess in a
u
c
o
gnition pe
f
ormation. T
h
c
al or glo
b
a
o
ce
ss f
o
r
fi
n
d
l
ooki
ng for t
h
e insta
n
t w
h
.
e
nt not
e
on
s
e
saro
n, pe
k
ho
wn as an
e
times mak
e
n
otes, e
s
pe
c
o
r saron ’3’
a
h
resh
olde
d
s
f
rom gamel
a
e
tected. The
r
din
g
to ga
m
n
sets
. The
w
b
al thre
sh
old
e
d to segm
e
0
6
m
of ga
melan
u
tomatic ga
m
rforma
nce.
I
h
e simple
a
p
l
thresh
old
d
ing the n
o
t
e
h
e on
set tim
e
h
en the
play
e
e
t
s
in a gam
e
k
ing, an
d b
o
abru
pt ene
e
the note b
o
ially if the n
o
a
nd bon
a
ng
’
s
pec
tr
o
g
r
a
m
a
n re
co
rd
ing
algorith
m l
o
m
elan in
str
u
w
ea
k on
set c
a
value (see
F
e
nt the spe
c
t
r
recor
d
ing
m
ela
n
no
tes
I
n
this pa
pe
p
proa
ch
is to
(a
s sh
own
e
s bo
und
ari
e
e
of the not
e
e
r st
rik
e
the
e
lan recordi
n
o
nang
) the
s
rgy inc
r
eme
o
unda
rie
s
di
f
o
tes
with
lo
c
’
3’
(as
sho
w
n
, ons
e
t
c
a
n
o
ok
e
d
fo
r
e
n
e
u
ments. Th
e
a
ndid
a
tes a
r
F
i
gure 4).
T
h
r
og
r
a
m.
ISSN: 169
3
transc
riptio
r
w
e
us
ed
s
fi
nd p
o
ints
w
in Figure
3
)
e
s. Bounda
ri
e
e
s i
n
the am
p
gamela
n
bl
a
n
g. For
not
e
s
s
u
dden cha
n
nts. The
ga
f
fus
e
. So it i
s
c
ations th
at
n
in
Figure 3
didate
s
or
e
e
rgy c
h
ang
e
e
most
sali
e
r
e c
o
ns
id
er
e
d
h
e onsets
de
t
3
-6
930
n that
s
imp
l
e
w
he
r
e
)
. The
e
s ar
e
p
litude
a
de
o
r
s
from
n
ge
of
mel
a
n
s
hard
are in
).
e
vents
e
s [
2
8]
e
nt
or
d
as
a
t
ec
te
d
Evaluation Warning : The document was created with Spire.PDF for Python.
TEL
K
2.3.
F
ef
fi
ci
e
fi
rs
t
s
featu
extra
featu
com
b
(ZC
R
They
sou
n
d
featu
(i)
(ii)
(iii
)
K
OM
NIKA
F
eatur
e Ext
r
The m
a
i
n
e
nt, s
o
the
p
s
egm
ent
ed
t
r
e
s
such a
s
c
t
e
d
and
c
a
re
s sh
ou
ld
b
Stan Z.
L
b
i
nation
s
for
R
), brig
htnes
s
also u
s
ed c
e
d
. Bas
e
d o
n
re
s give a c
o
There ar
e
Temporal f
e
A
uto
c
Zero
Loca
(tem
p
sk
e
w
A
mpl
"grai
n
Ceps
tral fe
Mel-
f
enve
)
Spec
tral fe
a
A
su
spe
c
t
asy
m
A
udi
o
Spectral-b
a
r
ac
tion
n
i
d
e
a
o
f
f
e
a
t
p
ro
ce
ss requ
t
he
spe
c
t
r
o
g
s
spe
c
t
r
al
c
e
a
lculated.
B
e
b
e incl
ude
d f
o
L
i et al [29] e
x
au
d
i
o c
l
as
s
s
, an
d ba
n
d
e
pstral coef
fi
n
th
e
CC,
m
o
o
mpl
e
ment t
o
e
four featur
e
e
ature
s
c
o
rrel
a
tion c
o
cr
os
sing
rat
e
l temp
oral
w
p
o
r
al ce
ntroi
w
ne
ss
Tk
)
itude M
o
d
u
n
i
n
e
ss" o
r
"r
o
ature
s
f
re
que
ncy c
lo
pe
o
v
er
th
e
a
tures
b
s
et of f
eat
u
t
ral
c
ent
roi
d
m
met
r
y (
Sa
)
,
o
Spec
trum
F
I
S
a
sed Featu
r
e
Figure
t
ur
e
extr
a
c
ti
o
ires a
sm
all
g
ram
ba
se
d
e
ntroid, sp
e
c
e
fore
pe
rfo
r
m
o
r re
cog
n
itio
n
Figure 5. O
n
x
plai
ned a
b
o
i
fi
cation. T
h
e
d
wi
dth for
c
a
fi
cient
s (C
C)
o
st of
the or
o
the perce
p
e
s
e
ts
[12]
s
u
o
efficient
s (
A
e
s:
u
s
ing
sh
o
w
aveform
m
d
Tc
, temp
o
u
lation feat
u
o
ug
hne
ss"
ep
stral coe
f
e
firs
t few
c
o
u
re
s obtai
n
e
d
(
Sc
), sp
e
,
and sp
ectr
a
F
lat
n
e
ss (
A
S
S
SN: 1693-6
9
e
s Ran
k
in
g
f
4. Onsets
c
a
o
n i
s
to
perf
o
and sim
p
l
e
d
on
the
on
s
c
tral
fl
ux, m
e
m
in
g fe
a
t
ur
e
n
p
r
oc
ess
.
n
set base
d
s
o
ut pe
rce
p
tu
a
e
y us
ed s
h
o
a
pturin
g the
for captu
r
i
n
g
iginal sound
tual ch
ara
c
t
e
u
itable f
o
r a
u
A
C
): signal
o
rt wind
ows
m
oments
, in
c
o
ral width
T
w
u
re
s (
AM
f
f
i
cient
s (
M
o
effici
ents.
e
d from the
e
ct
ral widt
h
a
l sk
ew
ne
ss
SF
) and Sp
e
c
9
30
f
or Gam
elan
a
ndid
a
te
o
rm
re
co
gniti
o
d
ata sp
ace.
ets info
rma
t
e
an an
d va
r
extrac
tion,
s
egmentatio
n
a
l featu
r
es,
m
o
rt time
ene
r
perceptu
a
l
c
g
the sh
ape
o
si
gnal ca
n
e
rist
i
cs.
u
d
i
o
s
i
gna
l p
r
sp
ectral di
s
t
(
ZCR
) and
c
lud
i
n
g
the
w
, temporal
), meant
M
FFC
), ten
firs
t four
s
t
a
(
Sw
), sp
e
(
Sk
)
c
tral Crest
F
a
Inst
rum
ents
o
n pro
c
e
s
s
m
To obtain th
t
i
on (see Fi
g
r
i
a
nce of th
e
it
is i
m
porta
n
m
el-cep
stral
r
gy
(STE),
z
e
c
hara
c
te
risti
c
o
f the frequ
e
be recon
s
tr
u
r
oc
es
s
i
n
g
:
t
ri
bu
tion in t
h
long windo
w
firs
t four
s
t
as
ymmetr
y
to
descri
b
d to
re
pre
s
a
tis
t
ic
al mo
m
e
ct
ral s
k
e
w
a
ct
or
s (
SC
F
)
…(
Ari
s
T
j
a
h
y
m
o
r
e effecti
v
e feature
s
s
e
g
u
r
e 5).
T
h
e
e
segm
ent
w
nt
to de
cid
e
features an
d
e
r
o
c
r
o
ssi
ng
c
s of the s
o
e
ncy sp
ectru
m
u
cted again,
h
e time dom
a
w
s (
lZCR
)
t
atistical
mo
Ta
, and t
e
m
b
e t
he "tre
m
s
ent
the
s
p
m
ents
, nam
e
w
ne
ss o
r
s
p
)
y
anto
)
99
v
e and
e
t, we
e
n th
e
w
ill
be
e
wha
t
d
their
ra
tes
o
un
ds.
m
of a
t
hese
a
in
m
ents
m
po
ral
m
ol
o"
,
p
ect
r
al
e
ly the
p
ectral
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 11, No. 1, March 2
013 : 95 – 10
6
100
Spectral slo
p
e
(
Ss
)
,
s
p
ec
tr
al d
e
c
r
e
as
e
(
Sd
), spe
c
tral variation
(
Sv
), spectral
rolloff or frequenc
y
c
u
toff (
Fc
), and spect
r
a
l
flatness
(
So
).
Freq
uen
cy de
rivative of the con
s
tant-Q coefficient
s (
Si
)
Octave Band
Signal Intensi
t
ies (
OBSI
)
(iv) Perceptual feature
s
Relative spe
c
ific loudn
ess (
Ld
), sha
r
pn
ess (
Sh
), and spre
ad
(
Sp
).
In this p
ape
r,
we p
r
ovid
e 3
4
sp
ect
r
al fea
t
ures,
su
ch
a
s
: funda
ment
al freq
uen
cy, spe
c
tral
centroid (
Sc
), two spec
tral
rolloff (
Fc
), sp
ectral flux (
SF
), sp
ectral
ske
w
n
e
ss (
Sa
), spe
c
tr
al
kurt
o
s
i
s
(
Sk
), spectral
slop
e
(
Ss
),
and sp
e
c
tral ban
dwi
d
th
(
Sw
). The
s
e
feature
s
a
r
e
then
combi
ned a
s
a feature
set of a gamelan
soun
d.
The feature
set is
norm
a
lized b
y
dividing ea
ch
feature
com
p
onent by a
re
al num
ber
so
the re
sult
i
s
betwe
en
-1 a
nd 1. Th
e no
rmalize
d
featu
r
e
set is con
s
ide
r
ed a
s
the final rep
r
e
s
entat
ion of the ga
melan soun
d.
Spectral ske
w
ne
ss
(
Sa
) i
s
a
mea
s
u
r
e
of
the a
s
ymmet
r
y of the
sp
e
c
trum
aroun
d
th
e
mean value. If
0
<
Sa
indicate
s m
o
re en
ergy o
n
the right sid
e
. If
0
>
Sa
indicate
s more ene
rgy
on the left sid
e
. Spectral kurtosi
s
K
is a
measure of the pea
ke
dne
ss
or flatne
ss of the sha
pe
o
f
power spe
c
trum distrib
u
tion. Positive kurto
s
i
s
3
>
K
indicate
s a pea
ked di
strib
u
tion, the
stand
ard
normal dist
ributi
on ha
s a
kurtosis
0
=
K
, and n
egative ku
rto
s
is
3
<
K
indicates a
flatter distrib
u
t
ion [30]. Those featu
r
e
s
(spe
ctra
l
sk
ewness
, s
p
ec
t
r
al moment, s
p
ec
tral
k
u
rtos
is
,
and spe
c
tral
entropy
) we
re
implemente
d
using
statisti
cal fun
c
tion.
Spectral cent
roid (
Sc
) is a measure of the cente
r
of
gravity of
the spe
c
trum.
The
spe
c
tral
ce
ntroid is
co
mput
ed by multipl
y
ing t
he valu
e of ea
ch fre
quen
cy by its magnitud
e
, then
the sum of al
l these divide
d by the
sum
of
all the
m
agnitud
e
s. T
he
spe
c
tral
centroid
(
Sc
) [
31]
[29] [32] can be define
d
as Eq. (1),
)
(
)
(
)
(
=
1
=
1
=
i
f
M
i
f
M
i
f
Sc
N
i
N
i
(
1
)
whe
r
e
)
(
i
f
M
is the
magnitud
e
fo
r the f
r
equ
en
cy
f
at bin
i
,
N
is the nu
mbe
r
of frequ
en
cy
bins.
Schei
rer a
nd
Slaney define
d
the spe
c
tral
rolloff point (
Fc
) as the 95th
percentile of the
power
sp
ect
r
um di
strib
u
tio
n
[33]. Spe
c
tral roll
o
ff is th
e freq
uen
cy
whe
n
9
5
% of
the si
gnal
en
ergy
is co
ntaine
d. Spectral rollof
f
(
Fc
) is define
d
as Eq. (2
),
)
(
0.95
)
(
1
=
1
=
i
f
M
i
f
M
N
i
Fc
i
(
2
)
2.4. Featur
e Rankin
g
The goal
s of feature sele
ction are imp
r
o
v
ing
comp
uta
t
ional efficien
cy but preserving or
even in
crea
si
ng
re
cog
n
ition rate. It be
comes imp
o
rt
ant to th
e
su
ccess
of the
tasks that a
p
p
ly
machi
ne l
earning
app
roa
c
h e
s
pe
cially
wh
en th
e
data h
a
ve
many irrelev
ant or redun
dant
feature
s
. In g
eneral, the fe
ature
s
sele
ction
alg
o
rithm
s
can be cate
gori
z
ed as wrappe
r
ap
pro
a
c
h
and filter app
roach [34] [1].
The five filter-ba
s
ed fe
ature ran
k
in
g techniqu
es b
e
in
g com
p
a
r
ed
are d
e
scribe
d
belo
w
.
Those te
ch
ni
que
s a
r
e
Info
rmation
Gai
n
(
IG
), Gai
n
Rati
o (
GR
), ReliefF
(
R
F
),
Chi-Square
d
(
CS
) and Symmetric Uncertai
nty (
SU
), and available in the
We
ka data mi
ning tool [44]:
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Spectral-b
ased Featu
r
e
s
Ran
k
in
g for Ga
m
e
lan Inst
rum
ents…
(Aris Tjah
yanto
)
101
(i)
ReliefF (RF)
is a
n
exten
s
i
on of th
e
Rel
i
ef algo
rithm
develop
ed by
Kira
and
Re
ndell [3
6].
The main ide
a
of Relief algorithm is to
evaluat
e the worth of a feature or attri
b
utes ba
sed
on ho
w well their value
s
can be u
s
ed to distin
g
u
ish
among the in
stan
ce
s. Reli
ef algorith
m
can
not ha
ndl
e incompl
e
te
data an
d o
n
l
y
limited to two-cla
s
s p
r
o
b
lems.
The
ReliefF is the
extended ve
rsion
of Relief.
ReliefF
can
handl
e incom
p
lete data a
n
d
not limited t
o
two cl
ass
probl
em
s. Howeve
r, if we apply th
e
algorith
m
for a hig
h
ly noi
sy data th
at have ma
ny
irrel
e
vant feature
s
and/o
r
mislab
eling, t
he perfo
rma
n
c
e of Reli
efF can g
e
t worse [37].
(ii)
C
h
i-
Sq
ua
r
e
d (
C
S)
can be
use
d
to evaluate the worth
of
a feature by calculating
the value
of
the Chi
-
Sq
uare
d
with re
spe
c
t
to
th
e class.
Th
e n
u
ll
hypothe
si
s i
s
the a
s
sum
p
tion that th
e
two featu
r
e
s
are
un
relate
d
,
and it i
s
te
st
ed by
Chi
-
Sq
uare
d
fo
rmul
a from
Pla
c
ke
tt [38]. If we
got a large va
lue of CS, then we can det
ermin
e
that the feature is a
n
importa
nt feature.
(iii)
Information gain
(IG)
can
also be
u
s
ed
for d
e
termi
n
i
ng the fe
ature ra
nk.
The
main id
ea
of
IG
is to
sele
ct
feature
s
b
a
sed o
n
e
n
trop
y.
Entropy is a me
asure
of ho
w mixe
d up
or
uncertainty o
r
the di
so
rd
er
degree
of a
system. [39] [4
0].
IG
mea
s
u
r
e
s
the n
u
mb
er
of bits
of informatio
n
gaine
d abo
u
t
the cla
ss
predictio
n whe
n
usi
ng a
given feature to
sup
port the
predi
ction. Informatio
n gain
[40] of the fe
ature o
r
attrib
ute
A
is
defined as
Eq. (3),
)
|
(
)
(
=
)
(
A
C
E
C
E
A
IG
(
3
)
whe
r
e
)
(
C
E
is the entropy of cl
asse
s
C
and
)
|
(
A
C
E
is the conditio
nal entropy
of
C
given
A
when the value
of the attribute
A
is known.
(iv
)
The
Gai
n
Ratio
(G
R)
is an
extend
ed ve
rsi
on
of
Information
G
a
in.
GR
is more efficient
a
nd
effective than
Informatio
n
Gain
and
ca
n be
u
s
ed
to
evaluate
the
co
rrelation
o
f
attribute
s
with re
spe
c
t to the cla
ss
concept of an incom
p
lete
da
ta set in [41] [42] [35]. Th
e gain ratio
of
A
is defin
e
d
as th
e info
rmation
gain
[40] of
A
divided by its intri
n
si
c value
)
(
A
IV
usin
g Eq. (4),
)
(
)
(
=
)
(
A
IV
A
IG
A
GR
(
4
)
whe
r
e
N
A
log
N
A
A
IV
i
i
k
i
2
1
=
=
)
(
whic
h
i
A
is the num
ber of in
stan
ce
s wh
ere
attribute
A
takes the
value
of
i
A
,
k
is the
n
u
mbe
r
of di
st
inct value
s
of
attribute
A
,
and
N
is the total numbe
r of instan
ce
s in the data
s
et.
(v
)
Symmetr
ic Unce
rtain
t
y (SU)
is
a correlation m
e
asu
r
e b
e
twe
e
n
the feature
s
an
d the
cla
ss,
and it is obtai
ned by [44] [1] Eq. (5),
)
(
)
(
)
|
(
)
(
2
=
C
E
A
E
A
C
E
C
E
SU
(
5
)
where E
(
A) and E(C) are t
he ent
ropi
es
based
on the probability associ
ated
with feature
A
and cl
ass val
ue
C
.
2.5. Cross V
a
lidation
As di
scu
sse
d
in the
p
r
evio
us
se
ction,
we ne
ed to
m
a
ke
a
co
mpa
r
iso
n
of
pe
rfo
r
man
c
e
betwe
en different ran
k
ing
appro
a
che
s
using
cro
ss validation method. Cross validation is a
statistical me
thod of evalu
a
ting an
d co
mpari
ng
le
arning al
gorith
m
s by dividi
ng data i
n
to
two
portion
of da
ta for trai
nin
g
and vali
dat
ing o
r
testin
g
the mod
e
l. The g
oal i
s
to co
mpa
r
e t
he
perfo
rman
ce
of
different ranki
ng
a
p
p
r
o
a
ch
es and
fi
nd out th
e b
e
st a
pproa
ch
for the
gam
elan
instru
ment
s reco
gnition. Cross validatio
n can
al
so b
e
use
d
to un
derstand the
generalizatio
n
power of a cl
assi
fi
er.
Evaluation Warning : The document was created with Spire.PDF for Python.
TEL
K
102
insta
n
are u
In th
sev
e
r
popu
into
a
the
m
ker
n
e
on t
h
para
m
s
e
ts
:
instr
u
tone
s
two
c
(OA
O
agai
n
are
u
traini
pair
o
this
a
3. E
x
audi
o
frequ
=
M
of m
e
the d
desc
e
K
OM
NIKA
V
W
e
us
ed
n
ce
s in t
he
o
se
d for t
r
ain
is
r
e
s
ear
c
h
,
r
a
l
ke
r
nel fu
n
l
a
r tool f
o
r
d
a
high di
me
n
m
aximal mar
g
To train
a
e
l, we
n
e
ed
d
h
e pro
c
edu
r
m
eters. The
l
og2
(C)
u
ment
s a
r
e
c
s
, an
d 1
0
b
o
c
o
mmon a
p
O
) [13]. I
n
O
n
st
all cla
s
s
e
u
sed in
con
s
ng data an
d
o
f
cl
as
se
s b
y
a
p
p
r
o
ac
h is
M
x
periments
a
The d
a
ta
o
,
colle
ct
ed
f
e
n
cy
sa
mpli
31
=
cla
s
s
e
s
e
tal with thei
r
A
fter the
ata into
trai
n
e
nding
o
r
de
r
V
ol. 11, No.
1
1
0
-f
old cro
s
o
ri
ginal da
ta
s
in
g an
d th
e
l
,
cr
os
s
v
a
li
d
n
ctions: line
a
d
ata cla
s
si
fi
c
n
si
onal sp
ac
e
g
in [45].
a
n S
V
M, w
e
d
e
t
er
mine t
h
r
e Grid
-S
ea
grid
-sea
rch
{-3, -2, ...,
1
c
la
ss
i
f
i
e
d
in
t
o
na
ng ton
e
s.
proa
ch fo
r
m
O
A
A
app
ro
a
c
s (
M-1
). Th
e
s
truc
ting an
S
d
thei
r out
pu
t
y
t
r
ainin
g
i
t
t
1)/
2
(
M
M
a
nd Disc
us
s
ba
se u
s
ed i
n
f
rom
Elekt
r
o
ng 44
100
H
. We
prod
u
c
r
own ha
mm
Figur
e
feature d
a
t
a
n
ing data se
t
r
u
s
in
g th
e
f
1
, Marc
h 20
s
s-vali
dation
s
et into 1
0
a
p
l
ast i
s
u
s
ed
f
d
ation we
re
a
r, p
o
ly
no
mi
a
c
ation o
r
re
c
o
e
a
nd
fi
nd a
e
m
u
st
sele
c
h
e value
of
rch [44].
G
c
a
n
be
don
e
1
2}
and
log
2
o 31 cl
asse
s
Cla
s
sif
i
ca
t
i
o
m
ulticl
ass
S
c
h, an
SVM
e
n the num
b
e
S
VM for a
c
t
s
)
,
(
i
i
y
x
.
t
o differenti
a
2
.
s
ion
n
o
u
r experi
m
Budoyo ITS
z. The trai
ni
c
ed th
e so
u
n
er at cent
er,
e
6.
Diff
eren
t
a
set w
a
s
c
a
t
s and
te
sti
n
f
ive te
chni
q
u
13 : 95 – 1
0
pr
oc
edu
r
e
.
p
proximatel
y
f
or testin
g. T
impleme
n
t
e
a
l, radial
b
a
s
o
gnition. Th
e
hyperpl
ane
t the
proper
and
C
tha
t
G
rid s
e
a
r
ch
e
by
sel
e
ct
i
n
g
2(
)
{-6,
s
, they are
7
o
n of multicl
a
S
VM: one-a
g
is create
d
f
e
r of SVMs
c
la
ss.
The S
V
For O
A
O a
p
ting the two
m
ents is co
m
gamela
n
s
e
ng data
set
n
d
s
dat
a sa
m
uppe
r, an
d
l
t
st
ru
ck a
r
e
a
a
lculate
d
an
d
g data
set
s
.
u
es
. T
h
e
r
a
n
0
6
Th
e
pr
oce
d
u
y
equal
sets
his whol
e
pr
o
e
d u
s
ing Li
b
s
is
func
tion
(
e
main idea
that sepa
ra
value of
th
e
t
can
be
obt
a
can be co
n
g
th
e va
lu
e
o
-5, ..., 10}
.
7
demun
g
t
o
a
ss
c
a
n be
a
g
ainst-all (O
A
f
or ea
ch cla
s
c
reated in O
A
V
M for clas
s
p
p
r
oa
ch
, an
cla
s
se
s.
Th
m
p
o
sed of
a
p
e
t. All audio
a
con
s
ist
s
of
m
ple
s
by
ran
l
owe
r
area
s
a
for data col
l
d
extract
ed,
The trai
nin
g
n
kin
g
of
f
e
a
t
u
re randoml
y
of size N/1
0
o
ce
ss t
h
en
r
b
SVM [44].
(
RBF
)
, and
s
of SVM is
t
t
e
betwe
en
t
e
ker
n
el par
a
a
ine
d
us
in
g
n
du
cted to
o
f
and
C
In this
res
e
a
o
n
e
s,
7
sar
o
a
chiev
e
d by
A
A) and th
e
s
s by
differ
e
AA
is
M
. All
t
s
k
is create
SVM is
c
o
n
e n
u
m
be
r o
f
p
p
r
oximat
el
y
a
re 16-bit,
m
2790
=
N
a
domly hittin
g
[46] as sho
w
l
e
c
ting
then we
ra
n
g
feature
s
d
a
t
ure
s
obtain
e
ISSN: 169
3
y
partitione
d
. Then 9
pa
r
r
ep
eated 1
0
LibSVM pr
o
s
ig
m
o
i
d
. SV
M
o proje
c
t th
e
t
he
tw
o
c
l
as
a
meters
. Fo
gri
d
sea
r
ch
b
cho
o
s
e
t
h
e
from the foll
a
rc
h, the
ga
o
n t
o
ne
s, 7
p
SVMs
.
The
e
one-agai
n
s
e
ntiating the
t
he
N
trainin
g
d usin
g the
n
st
ru
ct
ed f
o
r
f
SV
M
s
c
r
e
a
y
2790
seg
m
m
ono-ch
ann
e
a
udio samp
l
g
the keys
o
w
n at Figure
6
n
domly p
a
rti
t
a
ta w
e
re ran
e
d for the tr
3
-6
930
the N
r
titions
t
imes.
o
vides
M
is a
e
da
ta
s wit
h
r RBF
b
ase
d
e
best
owi
n
g
me
lan
p
ek
in
g
re are
st
-o
ne
cla
s
s
g
dat
a
set
of
eve
r
y
a
ted in
m
ented
e
l, and
es
fo
r
r ba
r
s
6
.
t
ioned
ked
in
ai
ning
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Spectral-b
ased Featu
r
e
s
Ran
k
in
g for Ga
m
e
lan Inst
rum
ents…
(Aris Tjah
yanto
)
103
data is p
r
e
s
e
n
ted in Ta
ble
3. The first 7
feature
s
are
con
s
i
s
tently ranked
a
s
the
top. The first
4
feature
s
p
r
ed
icted by the f
i
ve techniq
u
e
s
gives
th
e same re
sult
s, althoug
h feature
s
5, 6 an
d
8
are reversed i
n
some
ran
k
i
ngs.
Table 2. Spe
c
tral
-ba
s
e
d
feature
s
No
Feat
ures
Num
b
er o
f
feat
ures
1 Fundament
al
Fre
quenc
y
1
2 Spectral
Centroi
d
1
3-4
Spectral
Rolloff
2
5 Spectral
Flux
1
6 Spectral
Skew
ne
ss
1
7 Spectral
Moment
1
8 Spectral
Kurtosis
1
9 Spectral
Entrop
y
1
10 Spectral
Slope
1
11 Spectral
Band
w
i
dth
1
12 Mean
1
13 Standard
D
e
viation
1
14 Mode
1
15 Median
1
16 Variance
1
17-25
Percentile
9
26-34
Quantile
9
For ea
ch
ran
k
ing m
e
thod,
investigatio
n of
recognitio
n
accuracy on
the testing d
a
t
a as a
function of th
e feature
s
ha
s bee
n don
e
in asce
n
d
ing
orde
r an
d de
scendi
ng ord
e
r. Re
co
gniti
on
rat
e
o
r
a
ccu
r
a
cy
w
a
s t
a
ke
n f
r
om p
r
edi
c
t
ion ac
cu
ra
cy
perf
o
rmed
b
y
S
V
M
s.
A
c
c
u
ra
cy
re
sult
s as a
function
n
number of feat
ure
s
in asce
nding o
r
de
r ar
e p
r
e
s
ente
d
in Figure 7, for descend
in
g
orde
r
are
p
r
e
s
ente
d
in
Fig
u
re
8.
We
me
asu
r
ed
the
p
e
rform
a
n
c
e
for
su
bsets
co
nsi
s
ting
of th
e
n
ran
k
ed featu
r
es. Wh
ere
n
varie
s
betwee
n
1 and 34, started fr
om the least impo
rtant feature
s
for ascen
d
ing
orde
r and fro
m
the most im
porta
nt feature
s
for de
scendin
g
ord
e
r.
The SVM pe
rform ve
ry well wh
en all f
eature
s
o
r
subsets of the
origin
al feat
ure
s
a
r
e
u
s
ed
. T
h
e pe
a
k
ac
cu
r
a
cy w
a
s
re
ac
he
d
on
th
e
19
u
n
t
il 22
be
s
t
fe
a
t
ur
es
in
r
a
n
k
in
g
by a
l
l
techni
que
s
at accu
ra
cy of
more
or eq
ua
l to 98.
8
7
%,
and i
n
crea
sin
g
the
su
bsets did
not imp
r
o
v
e
the a
c
cura
cy. Then
the
re
st of the
feat
ure
s
ca
n
be
deleted
due
to no
n-signifi
cant influen
ce
for
the perfo
rma
n
ce. Inte
re
stingly, the GR techni
que
show th
e pe
a
k
at a
c
curacy of 98.93%
(as
sho
w
n at Tab
l
e 5), the high
est accu
ra
cy
achi
evable u
s
ing the five techni
que
s.
Table 3. The
first twenty of feature ran
k
s on the spe
c
tral-b
ased ga
melan featu
r
e
s
; see
descri
p
tion in
the text; all entries d
enote feature
nu
mb
ers sho
w
n
in Table
2
Methods
Feat
ure Ra
nk
1 2 3 4
5
6
7
8
9
10
11
12 13
14 15
16 17 18
19 20
CS
1 3 2 4
8
5
6
11
10 28
34
29 30
15 31
32 12 33
7
26
IG
1 3 2 4
5
8
6
11
34 28
32
29 33
31 15
30
7
10
27 12
SU
1 3 2 4
5
8
6
11
34 32
31
29 28
30 15
33 10
7
12 27
GR
1 3 2 4
8
5
6
11
31 32
18
14 17
30 15
29 10 34
33 28
RF
1 3 2 4
6
8
5
13
27 32
26
33 28
30 15
29 31 12
10 11
Table 4. Accura
cy for gam
elan data
s
et
as a fun
c
tion
of the worst
n
ranked feature
s
(asce
ndin
g
order); for
n
=24
..34
Methods
A
c
c
u
rac
y
(
%
) for the
w
o
rst
n
ranked fea
t
ure
s
us
ed for
classif
i
ca
tion
34
33
32
31
30
29
28
27
26
25
24
CS
98.87
98.53
97.97
97.74
95.99
95.88
94.69
92.66
81.36
68.47
68.36
IG
98.87
98.53
97.97
97.74
95.99
95.37
94.69
92.66
81.36
79.66
79.60
SU
98.87
98.53
97.97
97.74
95.99
95.37
94.69
92.66
81.36
79.66
79.49
GR
98.87
98.53
97.97
97.74
95.99
95.88
94.69
92.66
81.36
81.41
81.41
RF
98.87
98.53
97.97
97.74
95.99
95.08
95.20
92.66
92.43
92.43
92.43
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 11, No. 1, March 2
013 : 95 – 10
6
104
Figure 7
sh
o
w
s the d
e
g
r
a
dation in
the
recogniti
o
n
ra
te or
accu
ra
cy when
the
n
u
mbe
r
of
feature
s
sub
s
ets is redu
ce
d. A compa
r
i
s
on of t
he five method
s shows t
hat th
e accu
ra
cy over
90% a
c
hi
eved with
RF
su
bset
s a
r
e
bett
e
r th
an
anoth
e
r
re
sults (se
e
Ta
ble
4). Al
l the
tech
niqu
e
s
sho
w
the
sa
me beh
avior
without a
n
y significa
nt di
fferen
c
e
s
. Th
e
accuracy i
s
almost
sam
e
until
the
sub
s
ets are
re
du
ced
to
26 or
le
ss
featur
es, th
en the
a
c
curacy ten
d
s to
de
crease
wi
th
redu
cin
g
the sub
s
et
s (see
Table 4 an
d Figure 7).
Figure 7.
Accura
cy for ga
melan data
s
e
t
as a
function of th
e worst
n
r
ank
e
d
fe
a
t
ur
es
(asce
ndin
g
order)
Figure 8.
Accura
cy for gam
elan data
s
et
as a
function of th
e best
n
ranked feature
s
(de
s
cendi
ng orde
r)
Table 5. Accura
cy for gam
elan data
s
et
as a fun
c
tion
of the best
n
ra
nke
d
feature
s
(de
s
cendi
ng orde
r);
for
n
=19...23, 30…
34
Methods
A
c
c
u
rac
y
(
%
)
f
o
r the b
est
n
rank
ed fea
t
ures
use
d
for cl
assifi
cati
o
n
34
33
32
31
30
...
23
22
21
20
19
CS
98.87
98.87
98.87
98.87
98.87
98.87
98.87
98.81
98.81
98.81
IG
98.87
98.87
98.87
98.87
98.87
98.87
98.87
98.87
98.81
98.53
SU
98.87
98.87
98.87
98.87
98.87
98.87
98.87
98.81
98.81
98.81
GR
98.87
98.87
98.93
98.93
98.93
98.93
98.93
98.93
98.59
98.59
RF
98.87
98.87
98.87
98.81
98.87
98.87
98.87
98.87
98.87
98.87
For d
e
sce
ndi
ng orde
r, the
accuracy i
s
q
u
ite st
abl
e un
til the sub
s
et
s redu
ced
to
7 or l
e
ss
feature
s
. The
seven
features a
r
e fu
nda
mental fre
que
ncy, sp
ect
r
al
roll off 40%,
spectral
centro
id,
spe
c
tral
roll
off 90%, spe
c
tral flux, sp
ectral
ku
rtosi
s
and
sp
ectral skewne
ss.
The first b
e
s
t
feature give accuracy of
5
3
.96%
, the seco
nd b
e
st fe
ature
s
give
6
6
.95%, the th
ird b
e
st featu
r
es
give 72.03%, and the seve
n best featu
r
e
s
give
accu
ra
cy of 96.55% (a
s sho
w
n
at Figure
8).
4. Conclusio
n
In this pa
per,
we h
a
ve prese
n
ted in
d
e
tails o
u
r a
p
p
roa
c
h to
pe
rform fe
ature
ran
k
ing
usin
g five filter-ba
s
ed
ra
n
k
ing
metho
d
s. Although
th
ey all p
e
rfo
r
m in
a
similar way, a
c
cu
ra
cy of
the SVM classifier h
a
s b
e
e
n
signifi
cantly
influenc
ed b
y
the feature
ran
k
ing. It sh
ows that Gai
n
Ratio (G
R)
te
chni
que gave
better re
sult than
the ot
he
r fou
r
te
chni
que
s. The
hi
ghe
st accu
ra
cy
98.93% for G
R
wa
s re
ache
d usin
g the 2
1
best featu
r
e
s
.
Five filter-b
a
s
ed
ra
nki
n
g
method
s h
a
ve
bee
n e
v
aluated. Th
e first
seven
feature
s
predi
cted by
the five te
chni
que
s gives the sam
e
results. T
he first sev
en feature
s
are:
fundame
n
tal freque
ncy, spectral roll of
f 40%, spectr
al centroid, spectral roll of
f 90%, spectral
flux, spectral
kurto
s
i
s
an
d spe
c
tral
skewn
ess.
Tho
s
e featu
r
e
s
give accuracy of 96.55% for
gamela
n
inst
rument identifi
c
ation.
Evaluation Warning : The document was created with Spire.PDF for Python.