Internati
o
nal
Journal of Ele
c
trical
and Computer
Engineering
(IJE
CE)
V
o
l.
6, N
o
. 5
,
O
c
tob
e
r
201
6, p
p
. 2
211
~221
8
I
S
SN
: 208
8-8
7
0
8
,
D
O
I
:
10.115
91
/ij
ece.v6
i
5.1
202
2
2
211
Jo
urn
a
l
h
o
me
pa
ge
: h
ttp
://iaesjo
u
r
na
l.com/
o
n
lin
e/ind
e
x.ph
p
/
IJECE
Statistical Based Audio Forens
ic on Identical Microphones
Fajri Kurni
a
wan
1
, Mohd.
Shafry Mohd.
Rahim
2
,
Mohammed S. Khalil
3
, Muham
m
ad
Khurram Khan
4
1,2
Facult
y
of Co
m
puting, Univer
siti T
e
knolog
i M
a
lay
s
i
a
, Skudai,
Malay
s
ia
1,3,4
Center
of Ex
cel
lenc
e in
Infor
m
ation Assurance, King
Saud Un
iv
ersity
, Riy
a
dh, Saudi Arab
ia
Article Info
A
B
STRAC
T
Article histo
r
y:
Received
J
u
n 20, 2016
Rev
i
sed
Au
g
12
, 20
16
Accepted Aug 27, 2016
Microphone forensics has become a challenging f
i
eld due to the p
r
olifer
ation
of recording dev
i
ces and exp
l
osion in video/aud
i
o recording
.
Video or audio
recording h
e
lps a crim
inal
inves
tigator
to ana
l
yz
e the s
cen
e and
to colle
ct
eviden
ces. In this regards, a
robus
t method is required to
assure the
originality
of s
o
me recordings. In this paper
,
we focus on digital
audio
forensics and stud
y
how to iden
tif
y
th
e microphone model. Definin
g
microphone model will allow th
e investigat
o
r
s to conclud
e
in
teg
r
ity
of some
recordings
.
W
e
perform
s
t
atis
ti
c
a
l an
al
ys
is
on
th
e re
cording
that
is
coll
ect
ed
from two
microphones of the same model.
Experimental results and analy
s
is
indicate
that
the signal of soun
d reco
rding of
identical microp
hone is not
exac
tl
y s
a
m
e
an
d the
diff
eren
ce
i
s
up to 1
%
-
3%.
Keyword:
Dig
ital aud
i
o
Ide
n
tical microphone m
odel
M
i
crop
h
one
f
o
rensi
c
s
Micro
pho
n
e
iden
tificatio
n
Statistical analysis
Copyright ©
201
6 Institut
e
o
f
Ad
vanced
Engin
eer
ing and S
c
i
e
nce.
All rights re
se
rve
d
.
Co
rresp
ond
i
ng
Autho
r
:
Mohd. Shaf
ry Mohd.
Ra
him
,
Facu
lty of Com
p
u
tin
g
,
Un
i
v
ersiti Tekn
o
l
o
g
i
Malaysia,
Sku
d
a
i, Jo
ho
r,
Malaysia 8
1
310
.
Em
a
il: sh
afry
@u
tm
.
m
y
1.
INTRODUCTION
M
i
crop
h
one
f
o
rensi
c
s i
s
a
rec
e
nt
resea
r
ch
i
n
t
e
rest
u
nde
r a
u
di
o
fo
re
nsi
c
sc
i
e
nce. T
h
e
ob
j
ect
i
v
e i
s
t
o
aut
h
e
n
t
i
cat
e whet
he
r a di
gi
t
a
l
audi
o
was
m
a
de on a
gi
ven
reco
r
d
er
or i
t
has be
en
t
a
m
p
ered. C
o
py
ri
g
h
t
in
fri
n
g
e
m
e
n
t
has b
e
en
an im
p
o
rtan
t issu
e in
th
is 21
st cen
t
u
ry. Id
en
tify the
m
i
cro
p
h
o
n
e
m
o
d
e
l
o
f
d
i
g
ital
au
d
i
o
record
i
n
g will
p
r
ov
id
e
v
a
lu
able ev
id
en
ce for th
e actu
a
l
ownersh
i
p
wh
en
cop
y
righ
t d
i
spu
t
e o
c
cu
rs. In
additio
n
,
f
o
r
g
er
y
o
n
d
i
gital au
d
i
o con
t
en
t is
u
n
a
vo
idab
le
n
o
w
a
d
a
ys. Using
a
sophisticated
m
u
lti
med
i
a sof
t
w
a
re m
a
k
e
su
ch
fo
rg
er
y tr
u
l
y eff
o
r
tle
ssness. C
r
iminal evide
n
ce fro
m arb
itrary d
i
g
ital record
i
n
g
m
u
st b
e
v
e
ri
fied
to
assu
re its in
teg
r
ity and
orig
i
n
ality. Hen
ce,
a ro
bu
st and
fast m
e
th
o
d
to
au
th
en
ticate such
d
i
g
ital con
t
en
t is
p
r
og
ressi
v
e
ly v
ital th
ese d
a
y
s
. In
add
itio
n, in
fo
rm
atio
n
abo
u
t
th
e
reco
rd
in
g
sou
r
ce can
effectiv
ely assist o
t
h
e
r
stu
d
y
lik
e
gu
nsh
o
t
ch
aracteri
zatio
n
[1
], tam
p
erin
g
de
tectio
n
[2
], sp
eak
e
r
reco
gn
ition
[3
],[4
] and
sp
eech
enha
ncem
ent [5].
In ge
ne
ral, mic
r
ophone fore
ns
ics is a
study on the ba
sis of t
h
e digital traces
that leaves on arbitra
r
y
recordi
n
g. Suc
h
traces
we
re
occurs
du
e to intrinsic c
h
a
r
ac
teristics of t
h
e
de
vice, whic
h
can be sha
p
e
d
from
audi
o sens
o
r
s,
com
pone
nt
t
echn
o
l
o
gy
o
r
so
m
e
defect
fro
m
t
h
e
m
a
nufact
ure i
t
s
el
f. A
n
au
di
o rec
o
rd
i
ng t
h
a
t
prete
nds t
o
have inc
o
nsistent traces the
n
it indicat
ed the di
gital content has
bee
n
ta
m
p
ered.
Actually,
ta
m
p
erin
g aud
i
o
d
e
tectio
n is
m
o
stl
y
in
sp
ired
fro
m
p
r
ev
i
o
us wo
rk
s i
n
im
a
g
e tam
p
erin
g detectio
n
[6
].
Th
e
p
i
on
eers i
n
th
is field
is
Kraet
zer et al.
[7
].
In
itially, t
h
ey u
tilized
K-mean
s and
Nai
v
e Bayes as
classifier along
with ste
g
a
n
a
l
ysis f
eatures
t
o
i
d
entify t
h
e
microphone m
odel.
After tha
t
, Kraetzer et
al. [8]
fuse
d Deci
si
on
Tree and Li
ne
ar Lo
gi
st
i
c
R
e
gressi
o
n
m
odel
t
o
achi
e
ve be
t
t
e
r perf
orm
a
nce. Few y
ears
back
,
Kraetzer et al
[9] propose
d
huge c
o
ntext m
odel as
a
gui
da
nce for ot
her res
earche
r
t
o
sele
ct suitable clas
sifie
r
and feat
ure
pri
o
r t
o
i
d
e
n
t
i
f
y
the m
i
crop
ho
ne
m
odel
.
The fe
at
ures t
h
at
co
m
m
onl
y
used f
o
r di
gi
t
a
l
audi
o dat
a
i
s
mel-scaled cepstral coe
ffici
ents (M
FCCs) features
. As
instance, Brew
et al [10] presente
d s
p
eake
r
v
e
rification
tech
n
i
q
u
e
with M
F
CC as th
e m
a
in
feat
u
r
es.
Dh
analaks
h
m
i
et
al [11] re
po
rt
e
d
di
gi
t
a
l
au
di
o
can be
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE
Vo
l. 6
,
N
o
. 5
,
O
c
tob
e
r
20
16
:
221
1
–
22
18
2
212
cl
assi
fi
ed t
o
al
l
o
w bet
t
e
r c
ont
ent
m
a
nagem
e
nt
. The M
F
CC
feature are fused with
lin
ear p
r
ed
iction
cepstru
m
coefficients (LPCCs),
pe
rc
eptually base
d linea
r
pre
d
ictive coe
fficie
n
ts (PLPCs
). Raba
oui et
al [12]
introduced s
o
und
recogn
ition
m
e
thod for s
u
rveillance appl
ication. Th
ey c
o
m
b
ined three
features to ac
hive
d
recognition
rate 96.8%. T
h
e features a
r
e MFCC, wavele
t
-
b
a
sed a
nd t
e
m
poral
-
fre
q
u
ency
feat
ur
es.
Hani
l
c
i
an
d
Ki
n
n
u
n
e
n
[
13]
pr
o
pose
d
cel
l
-
ph
o
n
e rec
o
g
n
i
t
i
on f
r
o
m
recor
d
ed a
u
di
o t
h
at
cont
ai
n
s
speec
h si
g
n
al
by
ext
r
act
i
ng
t
w
o
feat
u
r
es c
onsi
s
t
o
f
M
F
C
C
an
d LFC
C
.
Eski
de
re a
n
d
Karat
u
tlu
[14
]
wo
rk
s
on
source id
en
tificatio
n of
micro
p
h
o
n
e
u
s
in
g
m
u
ltitap
e
r-MFCC features.
In a
d
di
t
i
on,
o
t
her
wo
rk
s d
one
by
B
u
c
h
hol
z et
al. [15]. They c
o
ns
ider
e
d
Fourier coefficient
h
i
stog
ram
as th
e feat
u
r
es and
th
ey
u
tilizin
g
fou
r
cla
ssi
fier in
clud
ing
Si
m
p
le Lo
g
i
stic, J48
Decision Tree,
KNN and
SVM to
d
e
termin
e th
e
m
o
d
e
l. Besid
e
s th
at,
Esp
y
-W
ilso
n
[16] stu
d
i
ed
th
e dev
i
ce id
en
tificatio
n
for
lan
d
lin
e teleph
on
e
and
m
i
c
r
oph
on
e. Th
ey co
nsid
er
ed two features nam
e
d
MFCC
and
linea
r-c
epstral
coefficients. T
h
e m
o
st recent work
was
pe
rform
e
d by Vu e
t
al. [17] in 2012.
Vu et al. [17] introduce
d
novel
approach called
One
-
class cl
assifica
tio
n
(OCC) alon
g wit
h
represen
tativ
e
instance
clas
sification
fram
e
work
(RICF)
for m
i
croph
on
e fo
ren
s
ics. Th
e RICF is in
tro
d
u
c
ed
to
red
u
ce t
h
e no
isy sig
n
a
l su
ch
th
at it i
m
p
r
o
v
e
s
OC
C
per
f
o
r
m
a
nce. Ha
ni
l
c
i
et
al
. [18]
ex
pl
o
r
ed m
i
crop
h
one
i
d
ent
i
f
i
cat
i
on
pr
o
b
l
e
m
of spe
ech reco
r
d
i
n
g f
r
om
a
m
obi
l
e
ph
one
w
h
ere
au
di
o
rec
o
r
d
e
d
f
r
o
m
14 m
odel
s
o
f
m
obi
l
e
p
h
o
n
es
are
cl
assi
fi
ed
usi
n
g
vect
o
r
qua
ntization a
n
d SVM-base
d classifie
r
. M
o
re
over, Es
ki
dere [19] re
port
ed
his
recent
work in m
i
crophone
i
d
ent
i
f
i
cat
i
on
on
16 m
i
crop
h
one m
odel
s
us
i
ng GM
M
-
ba
s
e
d m
odel
i
ng t
echni
que al
on
g wi
t
h
t
h
ree d
i
ffere
nt
featu
r
es called
LPCC, PLPC an
d MFCC.
As presen
ted
ab
ov
e, th
ere is on
ly li
mited
stu
d
i
es h
a
ve bee
n
do
ne i
n
m
i
croph
o
n
e fo
re
nsi
c
s. Eve
n
vast
atte
m
p
t has be
en m
a
de but c
u
rre
nt works
only conside
r
e
d
s
i
ngle
device
for each va
rious microphone
models
.
In ot
her w
o
rd
, m
o
st of them
are foc
u
s o
n
in
ter-class pr
ob
le
m
,
wh
ich
is h
o
w
to
classify so
m
e
reco
rd
ing
then
i
d
ent
i
f
y
t
h
e d
e
vi
ce
m
odel
am
ong di
ffe
re
n
t
m
i
crop
h
one
m
odel
s
. Unf
o
r
t
unat
e
l
y
,
m
i
crop
h
one f
o
rens
i
c
s on
id
en
tical m
o
d
e
l o
f
reco
rd
ing
d
e
v
i
ce
(in
t
ra-cl
a
ss prob
lem
)
is still lack
of atten
tio
n fro
m
the co
mm
u
n
ity.
In t
h
i
s
pa
pe
r, a
n
au
di
o rec
o
rd
i
ng f
r
om
t
w
o i
d
ent
i
cal microphones of the
sa
m
e
m
odels are exam
ined
u
s
ing
statistica
l
an
alysis. Th
is stu
d
y
wou
l
d ad
d
n
e
w kn
owledg
e for mi
croph
on
e fo
ren
s
ics co
mm
u
n
ity an
d
fu
rt
he
r st
im
ul
at
e im
provem
e
nt
on pe
rf
orm
a
nce of m
i
croph
o
n
e i
d
ent
i
f
i
cat
i
o
n.
Whe
r
ei
n, m
i
cro
p
h
o
n
e f
o
re
n
s
i
c
s
practitione
rs should
not consid
er only digital traces betw
een one m
odel t
o
anothe
r m
odel, but also m
u
st take
into acc
ount t
h
e digital traces
within
t
h
e sam
e
m
odel in
order to care
f
ully
assess s
u
s
p
icous
digital audi
o.
In
section
1
presen
ts t
h
e literatu
re
rev
i
ew
o
n
m
i
cro
pho
ne forensics and
th
e
research
m
o
tiv
atio
n.
Section
2
pres
ents the s
p
eci
fications
of t
h
e devices
a
nd the explanati
on
of how the audi
o sam
p
le was
collected. In Section 3, a brief
descri
ption
of the sta
tistical an
alysis
tech
n
i
qu
e th
at u
s
ed
in
th
is stu
d
y
.
Afterwa
r
d, the
result is
discussed a
n
d analy
zed for each e
nvi
ronm
ent. Fina
lly, a conclusion
of this study is
prese
n
t
e
d
i
n
Se
ct
i
on
5 as
wel
l
as som
e
su
gge
st
i
ons
f
o
r
f
u
t
u
r
e
w
o
r
k
s.
2.
DAT
A COLL
ECTION
The aim
of this study is to e
xpl
ore t
h
e digi
tal
traces within the sam
e
mi
crophone m
o
del. Hence
we
col
l
ect
ed di
gi
t
a
l
audi
o
reco
r
d
i
n
g usi
ng t
w
o i
d
e
n
t
i
cal
m
i
cro
p
h
o
n
es
of
di
ffe
re
nt
m
ode
l
s
. The m
i
croph
o
n
e
m
o
d
e
l th
at is
u
s
ed
in th
is stu
d
y
is Sh
ure SM-58
.
Th
er
e a
r
e two ide
n
tical microphone
s for that m
o
del. It i
s
exciting t
o
study di
gital traces from
m
o
re microphones
of i
d
entical m
odels. Ho
weve
r, at
this stage
we fi
nd at
l
east
t
w
o
m
i
crop
h
one
s are eno
u
gh t
o
e
xpl
ore t
h
e
di
ffe
re
nce betwee
n i
d
entical micr
o
p
hon
e m
o
d
e
ls if ex
ists.
Th
e sp
ecificatio
n of t
h
e m
i
cro
p
hon
e th
at
u
tilized
in th
is st
u
dy is p
r
esen
ted
i
n
Tab
l
e
1
.
Th
ese two
m
i
c
r
oph
on
es were u
tilized
to
co
llect au
d
i
o
signals at two
d
i
fferen
t lo
cation
s
in
clu
d
i
ng
qui
t
e
r
o
om
and com
put
er l
a
b
o
rat
o
ry
. T
h
e
q
u
i
t
e
ro
om
i
s
a sou
n
d
pr
oo
f
ro
om
wi
t
h
alm
o
st
free f
r
o
m
noi
se. O
n
t
h
e ot
he
r ha
nd
,
com
put
er l
a
bo
rat
o
ry
i
n
t
r
o
duc
es
m
a
ny
noi
ses from
t
h
e act
ive C
P
Us
, ai
r con
d
i
t
i
one
r,
wal
k
i
n
g
pers
ons and other noises.
We conside
r
e
d
these two
locations to study the fre
quency response
of each
micro
p
h
o
n
e
again
s
t clear and
n
o
i
sy con
d
ition
s
.
Tot
a
l
t
w
o
rec
o
rdi
ng se
ssi
o
n
s
are pe
rf
orm
e
d d
u
r
i
n
g da
ta
co
llectio
n
.
No
ted
,
it is on
ly two
session
s
because we pe
rform
the recordi
ng sim
u
ltaneously for all microphones in each
environment. In this regards
,
all
micr
o
pho
nes ar
e or
gan
i
zed
in
a row
(h
or
izon
tally)
usin
g
stan
d
a
rd
micr
o
p
h
o
n
e
st
an
ds. Mor
e
ov
er
, on
e
session is a
three m
i
nutes rec
o
rding, which
consists
of
bot
h silence
and
s
p
eech rec
o
rdi
n
gs. Silence
rec
o
rding
means the m
i
c
r
ophone wa
s passively record the envi
ronm
ent without a
n
y hum
an speec
h exists. Meanwhile,
speec
h recordi
n
g
wa
s pre
p
are
d
from
a
pers
on who
rea
d
s
predefi
n
ed se
ntences
for t
w
o minutes. T
h
e
person i
s
sittin
g
on
a ch
air and
t
h
e
d
i
stan
ce
b
e
tween
t
h
e sp
eaker an
d th
e m
i
croph
on
e
was fix
e
d
t
o
30
c
m
. Th
e
gene
rat
e
d
fi
l
e
of
au
di
o
sam
p
le i
s
desc
ri
be
d i
n
Ta
bl
e
2.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
St
at
i
s
t
i
c
al
B
a
s
e
d A
u
di
o F
o
re
nsi
c
on
I
d
ent
i
c
al
Mi
cro
p
h
o
n
e
s
(
F
aj
ri
K
u
r
n
i
a
w
an)
2
213
3.
STATISTICAL-BASE
D METHOD
Th
is section
briefly d
e
scrib
e
s th
e statistical
an
alys
is techniq
u
e
u
tilized
in
th
is st
u
d
y
.
We co
n
s
i
d
er
fiv
e
m
e
trics
to
an
alyze th
e si
g
n
a
l, i.e. stan
dard
d
e
v
i
ation,
mean, the cres
t-factor Q,
dynam
i
c
range
D and
au
to
co
rr
elatio
n ti
m
e
. Th
ese
metr
ics h
a
s b
een
w
i
d
e
ly know
n
an
d
p
r
ov
en
cap
ab
le to
ch
ar
acter
ize an au
d
i
o
si
gnal
[2
0]
.
In
t
h
i
s
ex
peri
m
e
nt
, t
h
e c
o
l
l
ect
ed
audi
o si
gnal
i
s
anal
y
zed
un
de
r
M
a
t
l
a
b en
vi
r
o
nm
ent
.
Table
1. Microphone Feat
ures
And Specifica
tions
Shur
e SM
-
5
8
(Mic
1
: 2 units)
SPECIF
I
CAT
I
ON
S
T
y
pe: Dy
na
m
i
c
Fr
equency
Response: 50-
15,
000 Hz
Polar Patte
rn: Card
ioid
Sensitivity: -54.5 d
B
V/Pa
Im
pedance: 150
Ω
(
300
Ω
actual)
Polarity:
Positive pressure on
diaphrag
m
p
r
oduces positive voltage
on pin 2 with r
e
spect to pin 3
Connector
T
y
pe: 3
-
pin XL
R
Net
W
e
ight: 298 gr
am
s
Dim
e
nsions: 162 m
m
L x 51
m
m
W
Tabl
e
2. T
h
e
A
udi
o Sam
p
l
e
D
e
scri
pt
i
o
n
Descr
i
ption Value
For
m
at
Wave
Audio Form
at
PCM
Codec I
D
1
Bit r
a
te
705.
6 Kbps
Channel(
s)
1
channel
Sam
p
ling r
a
te
44.
1 KHz
Bit depth
16 bits
File size
~16.
9 M
B
Overall bit
rate
m
o
de
Constant
Bit r
a
te
m
ode
Constant
Form
at s
e
ttings, E
ndianness
Little
Form
at s
e
ttings,
Sign
Signed
3.
1.
Mea
n
Value
Of A
Si
gn
al
The a
v
era
g
e
value of the si
gnal
or
known as the m
ean (indicated by
μ
)
is u
s
ed
as th
e
b
a
sis
to
measu
r
e t
h
e si
g
n
a
l
power. It
is also
kn
own
as th
e
d
i
r
ect
cu
rren
t
v
a
l
u
e
(DC
v
a
lu
e). Rep
e
titiv
e sign
al
su
ch
as
sine wa
ve ca
n
be des
c
ribe
d si
m
p
ly
using t
h
e
DC value
.
Unfortunately,
most of nat
u
ral
signals, e
.
g. speech,
noi
se
or m
u
si
c, n
o
rm
al
l
y
have ran
d
o
m
peak-t
o-
pea
k
am
pl
it
ude.
He
nce, st
anda
r
d
de
vi
at
i
on
o
f
si
g
n
al
can
b
e
u
tilized
to
d
e
scrib
e
su
ch
si
g
n
al. Th
e m
ean
can
b
e
calcu
lated
as fo
llow
∑
,
whe
r
e t
h
e si
gnal store
d
in
x
i
, with nu
mb
er of sam
p
les N sign
al.
3.
2.
Stand
a
rd De
viation
O
f
A
Si
gn
al
The si
gnal
fl
uc
t
u
at
i
on
fr
om
i
t
s m
ean and fl
u
c
t
u
at
i
on
p
o
we
r
can
be est
i
m
ated f
r
om
st
anda
rd
de
vi
at
i
o
n
of a signal, called
σ
(sig
ma). Basically,
σ
is q
u
ite si
milar with
av
erag
e d
e
v
i
atio
n. Th
e d
i
fferen
t is th
e
avera
g
ing i
n
σ
u
s
ing
t
h
e
p
o
wer,
no
t th
e am
p
litu
d
e
. Calcu
l
atio
n of
σ
is
depi
cted as
below:
1
1
Whe
r
e,
t
h
e si
g
n
al
de
fi
ne
d as
xi
. T
h
e
n
, M
u
i
s
m
ean of t
h
e si
gnal
a
n
d
N i
s
n
u
m
b
er o
f
sam
p
l
e
s.
3.
3.
The Cres
t-Fac
t
or
Q
Qu
an
tity o
f
im
p
u
l
si
v
e
n
o
i
se,
sh
ort ev
en
ts
or sho
c
ks ca
n
b
e
esti
m
a
ted
u
s
ing
crest fact
o
r
.
It is ab
le t
o
co
m
p
u
t
e
th
e prob
ab
ility
o
f
un
n
ecessary wav
e
with
resp
ec
t to
th
e m
ean
o
f
sign
al. Su
ch un
wan
t
ed sign
al can
be co
nsi
d
e
r
ed
as di
st
ort
i
o
n o
r
bi
t
err
o
r o
f
t
h
e si
gn
al
. Cre
s
t factor is
m
e
asuri
ng a wa
veform'
s
peaks with
respect
t
o
m
e
an val
u
e
.
Nat
u
r
a
l
soun
d n
o
r
m
a
l
l
y
has hi
gh
crest factor, whereas crest
fact
or equal to one
m
eans
the signal
has
no pea
k
s
.
Crest
-factor calc
u
la
ted
in b
a
se-1
0 lo
g
a
rith
m
i
c fo
rm
as b
e
lo
w:
2
0
Whe
r
e
V
p
i
s
p
eak am
pl
i
t
ude
of
t
h
e si
gnal
an
d
Vrm
s
i
s
ro
ot
m
ean sq
uare
.
3.
4.
Dynamic
Range
D
Dy
nam
i
c rang
e i
s
sim
p
l
y
t
h
e rat
i
o
b
e
t
w
ee
n
pea
k
an
d
bot
t
o
m
of a si
g
n
al
. I
n
au
di
o
si
g
n
a
l
,
dy
nam
i
c
ran
g
e i
s
c
o
m
m
onl
y
co
m
put
ed as
base-
1
0
l
oga
ri
t
h
m
i
c v
a
l
u
e. It
o
f
ten
u
s
ed
to
ex
press ratio
of th
e
lo
ud
est
pos
si
bl
e wa
ve
wi
t
h
res
p
ect
t
o
R
M
S of
n
o
i
se am
pl
i
t
ude.
Dy
nam
i
c range o
f
h
u
m
a
n speec
h o
n
ave
r
age i
s
aro
u
n
d
4
0
dB
[
20]
.
Eq
uat
i
o
n
bel
o
w c
o
m
put
ed t
h
e
dy
nam
i
c
ran
g
e
of
si
g
n
al
:
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE
Vo
l. 6
,
N
o
. 5
,
O
c
tob
e
r
20
16
:
221
1
–
22
18
2
214
2
0
Whe
r
e, Vpea
k
and Vbottom
a
r
e
peak an
d
bo
tto
m
o
f
sign
als, resp
ectiv
ely.
3.
5.
Autoc
o
rrelation Time
Au
t
o
correlation
is a fu
n
c
tion
to
m
easu
r
e si
m
ilarit
y
b
e
t
w
een
its o
r
i
g
in
al sig
n
a
ls v
e
rsu
s
tim
e-lag
ap
p
lied
on
same th
e sig
n
a
l. Th
is fun
c
tion h
a
s cap
ab
ility
to
search
repetitiv
e p
a
ttern
th
at su
pp
ressed
wit
h
anot
her si
g
n
al
.
Aut
o
c
o
r
r
el
at
i
o
n com
m
onl
y
appl
i
e
d
on st
a
t
i
s
t
i
cal
si
gnal
s
. Eq
uat
i
on b
e
l
o
w defi
ned t
h
e
di
scret
e
au
to
co
rrelatio
n R of
d
i
screte sig
n
a
l:
∈
In g
e
n
e
ral, th
e
au
to
co
rrelatio
n related
t
o
a
d
e
lay ti
m
e
t is d
e
termin
ed
as
fo
llo
w:
1)
Calculate signa
l value at
a time t, de
noted
as
S
1
2)
Calculate signa
l value at
a time t +
τ
,
de
not
e
d
as S
2
3)
Co
m
p
u
t
ed
m
u
ltip
ly o
f
t
w
o sign
als,
4)
Perf
o
r
m
steps 1-
3
fo
r all
desired
tim
e
s t
5)
Finally, calcul
a
te the a
v
era
g
e
̅
∑
.
4.
RESULT AND DIS
C
USSI
ON
The collected
audi
o contents
are an
alyzed using statistica
l
analysis tech
nique as prese
n
ted above
.
Firstly, the audio si
gnal is s
e
parate
d between silen
ce
re
cording a
n
d s
p
eech rec
o
rding. T
h
e se
pa
ration i
s
sim
p
ly
based
o
n
tim
e fram
e
. The
first
60 se
conds
are
cons
idere
d
as
silen
c
e
r
e
co
rd
ing
an
d a
f
ter
6
0
s
e
co
nd
s a
r
e
tag
g
e
d
as sp
eech
r
e
co
rd
ing
.
In
th
is exp
e
r
i
men
t
, th
e co
m
p
ar
ison
is tak
i
ng
b
e
tw
een
id
en
tical
m
i
cr
o
p
hone, sa
me
recordi
n
g type
and at
sam
e
envi
ronm
ent. For e
x
am
pl
e, silence rec
o
rdi
n
g at
qu
ite ro
om
from
m
i
crop
h
o
n
e
Sh
ure
SM
-
5
8
i
s
com
p
ared
wi
t
h
si
l
e
nce
rec
o
r
d
i
n
g at
qui
t
e
ro
om
of S
h
ure
S
M
-5
8
of
an
ot
h
e
r m
i
croph
o
n
e.
Fo
ur
gra
p
hi
cal
pl
ot
s i
n
cl
u
d
es
si
gnal
i
n
t
i
m
e-
dom
ai
n, am
pl
itude s
p
ect
r
u
m
,
hi
st
o
g
ram
,
aut
o
co
rrel
a
t
i
o
n
are de
picted and c
o
m
p
ared in a table
m
a
nner to visualize
th
e d
i
fferen
ce if ex
ists. In
add
itio
n
,
fiv
e
statistical
values
are c
a
lculated then com
p
ar
ed am
ong identical m
i
cr
ophone
s.
4.
1.
Microphone
F
o
rensics In Quite Room Re
cordings
In t
h
e
qui
t
e
r
o
om
, we expect
t
h
e
m
i
crop
ho
ne sh
o
u
l
d
gi
ve
hi
gh
er si
m
i
l
a
ri
t
y
am
ong t
h
em
. As i
t
i
s
kn
o
w
n t
h
at
no
noi
se
was p
r
esent
an
d t
h
e
m
i
crop
ho
ne s
h
oul
d n
o
t
capt
u
re any
n
o
i
s
y
si
gnal
d
u
ri
ng si
l
e
nce
reco
rdi
n
g
.
Ho
weve
r, t
h
e e
x
peri
m
e
nt
sho
w
an
o
p
p
o
si
t
e
r
e
sul
t
as e
xpec
t
ed. Fi
gu
re
1
sho
w
s
h
o
w
si
gnal
o
f
silence recordi
ng
of Mic
1a
in
th
e ti
m
e
-d
o
m
ai
n
is b
i
gg
er th
an
Mic
1b
. Mo
r
e
o
v
e
r
,
w
e
fo
und also
Mic
1b
perfo
rm
s
anom
al
ous bas
e
d
o
n
obse
r
vat
i
on on
a
u
t
o
c
o
r
r
el
at
i
on pl
ot
.
T
h
e
aut
o
co
rrel
a
t
i
on pl
ot
ex
pl
ai
ned
t
h
at
t
h
e
ca
pt
u
r
ed
silen
ce u
s
ing
Mic
1b
h
a
s sev
e
ral p
a
ttern
s. Mean
wh
ile, Mic
1a
sho
w
s o
n
l
y
si
ngl
e pat
t
e
r
n
exi
s
t
s
on t
h
e s
i
gnal
,
wh
ich
is m
a
k
e
sen
s
e
fo
r silence sign
al in
qu
ite roo
m
.
Not
i
c
e, b
o
t
h
m
i
crop
ho
ne a
r
e
reco
r
d
i
n
g si
m
u
l
t
a
neousl
y
.
Refer to
statistical an
alysis in
Ta
ble 3, the
big differe
n
ce between Mic
1a
and M
i
c
1b
is th
e d
y
n
a
m
i
c
range val
u
e. M
i
c
1b
pro
duce
hi
ghe
r dy
nam
i
c range
rat
h
e
r
t
h
a
n
M
i
c
1a
. This means Mic
1b
g
e
n
e
rate m
o
re no
ise in
silence rec
o
rdi
n
g. T
h
e
reas
on proba
b
ly
there
is m
a
nufacture de
fect on Mi
c
1b
.
The
n
,
the peak
crest fact
or Q of
Mic
1b
al
so hi
gher t
h
a
n
M
i
c
1a
. Fi
nal
l
y
, bot
h
i
d
ent
i
cal
m
i
cr
op
h
one
pr
o
duc
e sim
i
l
a
r
m
a
xim
u
m
aut
o
cor
r
e
l
a
t
i
o
n
value t
h
at
near
to
55
second
s.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
St
at
i
s
t
i
c
al
B
a
s
e
d A
u
di
o F
o
re
nsi
c
on
I
d
ent
i
c
al
Mi
cro
p
h
o
n
e
s
(
F
aj
ri
K
u
r
n
i
a
w
an)
2
215
S
a
.Mic
1a
S
a
.Mic
1b
Fig
u
re
1
.
Fro
m
to
p to
bo
tto
m
:
Plo
t
for si
g
n
a
l
in
tim
e-d
o
m
ai
n
,
am
p
litu
d
e
spectru
m
,
p
r
o
b
a
bilit
y d
i
stribu
tion
and aut
o
correl
a
tion
of the sil
e
nce recordi
ng
in quite room
Tab
l
e
3
.
Statistical An
alysis of Silen
c
e Record
i
n
g in
Qu
ite Ro
o
m
Fo
r Sh
ure SM-58
(
Mic
1a
and
Mic
1b
)
Metrics
S
a
.Mic
1a
S
a
.Mic
1b
|
S
a
.M
i
c
1a
-
S
a
.M
i
c
1b
|
Sig
m
a 0.
1520
4
0.
1243
3
0.
0277
10
M
u
-
0
.
01482
2
-
0
.
01041
7
0.
0044
05
Peak (
c
rest)
factor
Q (
d
B)
16.
319
8
18.
078
4
1.
7586
00
Dy
nam
i
c range D
(
d
B)
31.
595
7
34.
647
9
3.
0522
00
Autocor
r
e
lation tim
e
(
s
ec.)
54.
921
2
54.
920
2
0.
0010
00
Aver
age Differ
e
nce
0.
9687
83
Ex
perim
e
nt result fo
r speec
h
reco
rdi
ng s
h
o
w
s m
o
re st
ab
le with
h
i
gh
si
milarit
y
co
m
p
are to
silen
ce
reco
rdi
ng as
expl
ai
ne
d
pre
v
i
o
usl
y
. Fi
g
u
r
e
2 desc
ri
be
d
t
h
e sim
i
l
a
r
si
gnal
,
am
pl
i
t
ude
, hi
st
o
g
r
a
m
and
aut
o
c
o
r
r
el
at
i
o
n
p
r
o
d
u
ces
by
bot
h m
i
crop
ho
nes.
The
st
at
i
s
tical analysis as prese
n
ted i
n
table 4 de
noted t
h
e
avera
g
e differe
n
ce of
the
t
w
o
identical im
age
s
is m
u
ch sm
aller com
p
ar
e t
o
the silence
rec
o
rding i
n
Ta
bl
e 3.
S
a
.Mic
1a
S
a
.Mic
1b
Fig
u
re
2
.
Fro
m
to
p to
bo
tto
m
:
Sign
al in
tim
e
-
do
m
a
in
, am
p
l
i
t
u
d
e
sp
ectru
m
,
p
r
ob
ab
ility d
i
stribu
tio
n and
autoc
o
rrelation of
t
h
e
speec
h
r
ecor
d
i
n
g i
n
qui
t
e
ro
om
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE
Vo
l. 6
,
N
o
. 5
,
O
c
tob
e
r
20
16
:
221
1
–
22
18
2
216
Table
4.
Statistical Analysis of Speec
h R
ecord
ing
in Qu
ite
Ro
o
m
Fo
r Sh
ure SM-58
(
Mic
1a
and
Mic
1b
)
Metrics
S
a
.Mic
1a
S
a
.Mic
1b
|
S
a
.M
i
c
1a
-
S
a
.M
i
c
1b
|
Sig
m
a 0.
1020
4
0.
1066
7
0.
0046
300
0
M
u
-
0
.
00040
956
-
0
.
00042
485
0.
0000
152
9
Peak (
c
rest)
factor
Q (
d
B)
19.
824
2
19.
438
7
0.
3855
000
0
Dy
nam
i
c range D
(
d
B)
62.
791
3
62.
483
6
0.
3077
000
0
Autocor
r
e
lation tim
e
(
s
ec.)
0.
0912
93
0.
0706
8
0.
0206
130
0
Aver
age Differ
e
nce
0.
1436
916
6
4.
2.
Microphone F
o
rensics
In
Computer
Lab Recordings
In la
b
reco
rdi
n
gs, M
i
c
1b
again s
h
ows
strange a
u
toc
o
rrelation
on
silence
recording as
prese
n
ted i
n
Fi
gu
re
3 an
d t
h
e di
ffe
rence
on a
u
t
o
c
o
rrel
a
t
i
on val
u
e i
s
h
i
ghe
r t
h
a
n
i
n
qui
t
e
r
o
om
(Tabl
e
3
)
. M
e
a
n
whi
l
e
,
anot
her
m
e
t
r
i
c
s (si
g
m
a
, m
u
, peak
crest
-fact
or
,
dy
nam
i
c range
)
does
not
show m
u
ch di
ffe
rence
com
p
are t
o
Mic
1a
, as shown
in Tab
l
e 5.
S
a
.Mic
1a
S
a
.Mic
1b
Fig
u
re
3
.
Fro
m
to
p to
bo
tto
m
:
Sign
al in
tim
e
-
do
m
a
in
, am
p
l
i
t
u
d
e
sp
ectru
m
,
p
r
ob
ab
ility d
i
stribu
tio
n and
au
to
co
rrelatio
n of th
e silen
ce
recordi
n
g in c
o
m
puter lab
Tab
l
e
5
.
Statistical An
alysis of Silen
c
e R
ecordi
n
g in C
o
m
puter La
b
.
Fo
r Sh
ur
e
SM-
58 (
Mic
1a
and
Mic
1b
)
Metrics
S
a
.Mic
1a
S
a
.Mic
1b
|
S
a
.M
i
c
1a
-
S
a
.M
i
c
1b
|
Sig
m
a 0.
2141
8
0.
2098
8
0.
0043
00
M
u
-
0
.
00881
91
-
0
.
00854
53
0.
0002
738
Peak (
c
rest)
factor
Q (
d
B)
13.
377
2
13.
553
4
0.
1762
00
Dy
nam
i
c range D
(
d
B)
36.
123
6
36.
390
9
0.
2673
00
Autocor
r
e
lation tim
e
(
s
ec.)
31.
228
5
45.
265
7
14.
037
200
Aver
age Differ
e
nce
2.
8970
547
6
Lo
oki
ng i
n
t
o
t
h
e speec
h rec
o
rdi
n
g
,
we f
o
u
n
d
b
o
t
h
ide
n
tical
microphones
produce alm
o
st the sa
me
values
.
Figure 4 depicts plots for
s
p
eech
rec
o
rding in com
pute
r
laborat
o
ry.
More
over, from
Table 6, it can
be
concl
ude
d that
in t
h
is s
p
eech rec
o
rding
both ide
n
tical
m
i
crophones a
r
e
produci
n
g quite similar signa
ls.
In
ad
d
ition
,
co
m
p
aring
th
is
resu
lt with
p
r
ev
ious resu
lt fo
r sp
eech
record
i
n
g
in
qu
ite ro
o
m
(Figu
r
e 2 and
Tab
l
e
4
)
, bo
th
resu
lts sho
w
s
alm
o
st
si
m
ilar p
a
ttern
. In
ano
t
h
e
r se
nse, spee
ches
re
cording in t
h
e
quite room
and the
co
m
p
u
t
er laborato
r
y
h
a
v
e
similar statistical v
a
lu
es.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
St
at
i
s
t
i
c
al
B
a
s
e
d A
u
di
o F
o
re
nsi
c
on
I
d
ent
i
c
al
Mi
cro
p
h
o
n
e
s
(
F
aj
ri
K
u
r
n
i
a
w
an)
2
217
S
a
.Mic
1a
S
a
.Mic
1b
Fig
u
re
4
.
Fro
m
to
p to
bo
tto
m
:
Sign
al in
tim
e
-
do
m
a
in
, am
p
l
i
t
u
d
e
sp
ectru
m
,
p
r
ob
ab
ility d
i
stribu
tio
n and
au
to
co
rrelatio
n of th
e sp
eec
h
rec. in c
o
m
p
. la
b
Table
6.
Statistical Analysis of Speec
h Reco
rd
ing
in Co
m
p
uter
Lab. For
Shu
r
e SM-5
8 (
Mi
c
1a
and
Mic
1b
)
Metrics
S
a
.Mic
1a
S
a
.Mic
1b
|
S
a
.M
i
c
1a
-
S
a
.M
i
c
1b
|
Sig
m
a 0.
0624
3
0.
0660
88
0.
0036
580
0
M
u
-
0
.
00012
704
-
0
.
00012
876
0.
0000
017
2
Peak (
c
rest)
factor
Q (
d
B)
24.
092
2
23.
597
6
0.
4946
000
0
Dy
nam
i
c range D
(
d
B)
72.
93
72.
809
6
0.
1204
000
0
Autocor
r
e
lation tim
e
(
s
ec.)
0.
0713
15
0.
0712
7
0.
0000
450
0
Aver
age Differ
e
nce
0.
1237
409
4
5.
CO
NCL
USI
O
N
In t
h
i
s
pa
per,
we ha
ve
prese
n
t
e
d a
no
vel
t
echni
que t
o
i
d
e
n
t
i
f
y
t
h
e m
i
cro
p
h
o
n
e m
odel
.
M
i
crop
h
o
n
e
classificatio
n
an
d
v
e
rification are cru
c
ial task
s to
en
sure
orig
in
ality o
f
some in
fo
rm
atio
n
,
wh
ich
are grow to
b
e
m
o
re and
mo
re im
p
o
r
tan
t
th
ese
d
a
ys in
un
lawfu
l
inv
e
sti
g
atio
n.
As co
nclu
sion
,
o
u
r
wo
rk
s
p
r
ov
e th
at
d
i
g
ital
traces
not
only present in different m
i
crophone
m
odels but also
it found in identical m
odel. The
plot
of
signal
i
n
t
i
m
e
-dom
ain, am
pl
i
t
ude, hi
st
o
g
ram
and aut
o
co
rrel
a
tion are presente
d to anal
yze the differe
n
ce betwee
n
gene
rated audi
o signals. Moreover,
fi
ve st
atistical values are c
o
m
puted from
each signal a
n
d us
ed as
com
p
arison tools. The e
xpe
rimental result prove
n
that
digital traces on ide
n
tical
mi
crophone
s are different up
t
o
1%
-
3%
.
Hence
,
fo
ren
s
i
c
ex
pert
s
h
oul
d c
onsi
d
er
t
h
i
s
di
f
f
ere
n
ce
pri
o
r t
o
a
n
al
y
ze t
h
e i
n
t
e
gri
t
y
o
f
au
di
o
cont
e
n
t
.
In add
itio
n, th
i
s
wo
rk
can
b
e
u
s
ed
as a b
a
se
to
im
p
r
o
v
e
th
e tech
n
i
q
u
e
for
micro
p
h
o
n
e
i
d
en
tificatio
n
i
n
l
a
rger s
p
ace
of m
i
crop
ho
n
e
m
odel
s
. Furt
her w
o
rks ca
n expl
ore m
o
re n
u
m
b
er of i
d
e
n
t
i
cal
m
i
crop
ho
n
e
s and
m
odels. Vari
ous statistical analysis techni
que that ca
pa
ble to c
h
aracteriz
e the a
udi
o signal can be
considere
d
to
stu
d
y
m
o
re th
e d
i
fferen
ce b
e
tween
id
en
ti
cal
micro
pho
nes. Co
mm
o
n
featu
r
es as rep
o
rted
in
th
e literatu
re,
e.g. MFCCs, L
P
CCs, etc., that used for m
i
cr
ophone ide
n
t
i
f
i
cat
i
on ha
ve t
o
exam
i
n
e agai
n by
i
n
cl
udi
n
g
a
udi
o
recordi
n
g of i
d
entical microphones
in t
h
e te
sting
dataset.
REFERE
NC
ES
[1]
I. L. Freir
e
and J. A. Apolinario
, “Guns
hot detection in nois
y
envir
onments,” in
Proceed
ing of the
7th International
Telecommunica
tions Sym
posium, Manaus,
Brazil
, September
,
201
0.
[2]
L. Cuccovillo,
et al.
, “Audio tampering detection vi
a m
i
croph
one cl
as
s
i
fica
tio
n,”
Multimedia
Signal
Processing
(
MMSP)
, 2013 IEEE 15th
Int. W
o
rkshop on
, Sep
t
. 30
2013-Oct. 2
,
pp. 177-182, 20
13.
[3]
L. Burge
t
,
et al.
, “Analy
sis of feature ex
traction and channel co
m
p
ens
a
tion in
a GM
M
s
p
eaker rec
ognition s
y
s
t
em
,
”
IEEE Transactio
ns on Audio
,
Sp
eech, and Langua
ge Processing
, v
o
l. 15
, pp
. 1979-
1986, 2007
.
[4]
D. A. Re
y
nolds
,
“
H
TIM
I
T and
LLHDB: s
p
eech
corpora for the
s
t
ud
y
of hands
et trans
duc
er eff
ects
,
”
IEEE Int.
Conf. on
Acoustics, Speech
, and
Signal
Processin
g
, vol. 2
,
pp
. 153
5-1538, 1997
.
[5]
Y. Jiang,
et al.
, “
A
Dual Microphone Speech E
nhancem
ent Alg
o
rithm
for CloseTa
l
k S
y
stem
,”
TE
LKOMNIKA
Indonesian Jour
nal of El
ectrical Engineering
, vol/issue: 12(6)
, pp
. 4475-4484, 201
4.
[6]
H. J
i
n, “
R
es
earch of Blind F
o
rens
ics
Algorithm
on Digital Im
age Tam
p
ering,
”
TELKOMNIKA Ind
onesian Journa
l
of Electrical En
gineering
, vol/is
sue: 12(7), pp
. 5
399-5407, 2014
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE
Vo
l. 6
,
N
o
. 5
,
O
c
tob
e
r
20
16
:
221
1
–
22
18
2
218
[7]
C. Krae
tzer
,
et al.
, “Digital Audio Forensics: A First Practi
cal
Evalu
a
tion
on Microphone and Env
i
ronment
Cl
a
ssi
fi
ca
t
i
o
n,
” in
Proceedings o
f
the 9th
workshop on Mu
ltimedia
and securi
ty, D
a
llas, Texas
, pp
.
63–74, 2007
.
[8]
C. Kraetz
e
r,
et al.
, “Unweighted
fusion in microphone forensics usi
ng a decision tree and lin
ear lo
gistic regr
ession
m
odels,” in
11th
workshop on M
u
ltimedia
and security.
Pr
inceton
,
New
Jersey, USA: ACM
Press
, p
p
. 49–56
, 2009
.
[9]
C. Kraetzer,
et al.
, “A contex
t model for microphone for
e
nsic
s and
its app
lication
in ev
aluations,”
M
e
dia
Watermarking, S
ecurity, and For
e
nsics III
, vol. 7
880, 2011
.
[10]
A.
Brew,
et al.
, “An evaluation of one-class
clas
s
i
fic
a
tion t
echniqu
es
for speaker ve
rifi
cat
ion,”
Art
ifi
ci
a
l
Intell
igen
ce
Re
vi
ew
, vo
l/issue: 27
(4), pp
. 295–307
, 2008
.
[11]
P
.
Dhanal
aks
h
m
i
,
et al.
, “Classification of
audio
signals using ANN and
GMM,
”
Applied Soft Computing
,
vol/issue: 11(1), pp.
716–723
,
20
11.
[12]
A. Rabaou
i,
et a
l
.
, “Using one
-c
la
ss svms
and wavele
ts
for aud
i
o
s
u
rveill
ance
,”
IEEE Transactions
on Information
Forensics and S
ecurity
, vol/issue: 3(4)
, pp
. 763–
775, 2008
.
[13]
C. Hanilçi and
T. Kinnunen
,
“Sour
ce Cell-
P
hone Recognition from Reco
rded Speech U
s
ing Non-Speech
Segments,”
Digital Signa
l
Processing
, vol. 35, p
p
. 75-85
, 2014
.
[14]
Ö. Eskidere an
d A. Karatutlu
, “Source m
i
crophone
ident
i
fi
cat
ion using multit
aper MFCC features
,”
9th
International Co
nference on
Electrical and Electr
onics
Eng
i
neerin
g (
E
LECO)
,
Bursa,
pp. 227-231,
2015.
[15]
R. Buchho
lz,
et al.
, “Microphon
e Classification
Us
ing Fourier C
o
efficients,”
LNCS
., vo
l. 5806, p
p
. 235–246
, 200
9.
[16]
D.
G.
Romero and C.
Y.
E.
Wilson,
“Automatic acqu
i
sition
device iden
tif
i
c
ation from
s
p
ee
ch record
ings
,”
in
Proceed
ing of I
EEE In
ternation
a
l Conferen
ce o
n
Acous
ti
cs Spe
ech and Signa
l
Processing, Dall
as, Texas, US
, p
p
.
1806–1809, 201
0.
[17]
H.
Q.
Vu,
et al.
, “Identif
y
i
ng Microphone from Nois
y
Recordin
gs
by
Using Representative Instance One Class-
Classification
Approach,”
Journal of Networks
,
vol/issue:
7(6), p
p
. 908-917
, 201
2.
[18]
C.
Ha
ni
lc
i,
et al.
, “
R
ecogni
tion o
f
brand and m
o
d
e
ls of c
e
ll-phon
e
s
from
recorded
speech signa
ls,”
IEEE T
r
ans. on
Information For
e
nsics and S
ecur
ity
, vol. 7
,
pp
. 62
5-634, 2012
.
[19]
O. Eskidere
, “
S
ource m
i
cropho
ne ident
i
fic
a
tion
from
speech
recordings based on a Gaussian m
i
xture m
odel,”
Turkish Journal
of Electric
al En
g.
&
Co
mp. Sciences
, vo
l. 22, pp. 754-767, 2014.
[20]
J. Earg
le, “Hand
book of Recording Engin
eerin
g
,
” Springer
,
pp
. 4
,
2005
, ISBN 0-
387-28470-2.
Evaluation Warning : The document was created with Spire.PDF for Python.