Indonesian J
ournal of Ele
c
trical Engin
eering and
Computer Sci
e
nce
Vol. 2, No. 1,
April 201
6, pp. 161 ~ 16
7
DOI: 10.115
9
1
/ijeecs.v2.i1.pp16
1-1
6
7
161
Re
cei
v
ed
De
cem
ber 2
2
, 2015; Re
vi
sed
F
ebruary 27,
2016; Accept
ed March 1
0
, 2016
Sparse Modeling with Applications to Speech
Processing: A Survey
AN Om
a
r
a*
1
,
AA H
e
fn
a
w
y
1
, Abdelhalim Zekr
y
2
1
Computers a
n
d
S
y
stems D
e
p
a
rtment, Electr
onics R
e
searc
h
Institute, Giza, Eg
ypt
2
Communic
a
tio
n
s and El
ectro
n
ics De
partme
n
t, F
a
cult
y
of Engi
neer
in
g, Ain
Shams Univ
er
sit
y
, Cair
o, Eg
ypt
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: ahmed
_omar
a@eri.sci.e
g
A
b
st
r
a
ct
Now
adays, the
r
e has b
een
a
grow
ing i
n
ter
e
st in
the stud
y of sparse ap
proxi
m
ati
on of
signa
ls.
Using
an ov
er
-complet
e dicti
onary co
nsisti
ng of prototyp
e sign
als or
a
t
oms, sig
nals
are descr
ibe
d
by
sparse li
near
combi
natio
ns of
thes
e
ato
m
s. Applic
atio
ns
that us
e sp
ar
se
repr
ese
n
tati
on are many and
inclu
de c
o
mpr
e
ssio
n
, source
separ
atio
n, e
nha
nce
m
e
n
t, a
nd re
gul
ari
z
at
i
on i
n
inv
e
rse
prob
le
ms, feat
ure
extraction, a
n
d
mor
e
. T
h
is arti
cle intro
duc
es
a literat
ure
rev
i
ew
of sparse c
odi
ng a
p
p
licati
ons in t
he fie
l
d
of
speec
h proc
es
sing.
Ke
y
w
ords
: Sp
arse mod
e
li
ng,
signa
l repres
e
n
tations, sp
eec
h process
i
ng
Copy
right
©
2016 In
stitu
t
e o
f
Ad
van
ced
En
g
i
n
eerin
g and
Scien
ce. All
rig
h
t
s reser
ve
d
.
1. Introduc
tion
Natural
and
artificial
sensors a
r
e t
he
only tool
s
we have
for sensi
ng th
e
world
and
gatheri
ng si
g
nals of phy
si
cal proc
esse
s. The
s
e sen
s
ors a
r
e u
s
u
a
lly not aware of the physical
pro
c
e
s
s und
e
r
lying the p
h
e
nomen
a they
“se
e
,” h
e
n
c
e
they often sa
mple the
sign
al with a
high
er
rate tha
n
the
effective dim
ensi
on of the
pro
c
e
s
s. To
rep
r
e
s
ent the
sam
p
led
dat
a efficiently,
we
have to re
du
ce its di
men
s
ion to b
e
e
ffective.
In other
words, t
he sig
nal h
a
s
to be lin
ea
rly
rep
r
e
s
ente
d
with a few p
a
r
amete
r
s. Su
ch repres
ent
ations often y
i
eld su
peri
o
r
sign
al pro
c
e
s
sing
algorithms.
Recent theory inform
s us that, with high
probability,
a relatively small num
ber
of
rand
om proje
c
tion
s of a sig
nal ca
n co
nt
a
i
n most of its relevant info
rmation.
One of the e
fficient sig
nal
rep
r
e
s
entati
ons
i
s
the
sp
arse de
com
p
osition. Thi
s
type of
sign
al
de
co
mpositio
n
h
a
s re
cently received ex
tensive
re
se
arch inte
re
st
acro
ss several
comm
unitie
s
inclu
d
ing sig
n
a
l
processin
g
,
informat
io
n
theory, an
d o
p
timization
[1
-3]. Also, th
e
s
e
rep
r
e
s
entatio
ns have fo
un
d su
cces
sful
application
s
in data interpretation,
so
urce sepa
rati
on,
sign
al de-noi
sing, codin
g
, cla
ssi
fi
cation,
reco
gnition,
and many mo
re [4].
In
sp
arse re
pre
s
entatio
n,
the signal
ca
n
be
co
nstructed
by el
ementa
r
y wa
veforms
chosen in a family called a dictionary [5]. T
he diction
a
r
y element
s a
r
e called ato
m
s that may
be
orthog
onal
or non-orth
ogo
nal [6].The ov
er-com
pleted
diction
a
rie
s
who
s
e ve
ctors a
r
e la
rge
r
t
han
bases
are
n
eede
d to bui
ld sp
arse re
pre
s
entatio
ns of compl
e
x sign
als [7]. But choo
sin
g
is
dif
fi
cult and requires m
o
re
compl
e
x algo
rithms.
This arti
cle
aims
at pre
s
entin
g an
ove
r
view of re
se
arch effort
s o
n
sparse
decompos
i
tions
of
speec
h
s
i
gnals
.
So, t
he
s
t
ruc
t
ure
of the artic
l
e is
as
follows
. In Sec
t
ion II, we
review
the b
a
si
c
d
e
finitions of
the sp
arse c
odi
ng.
And
we ill
u
s
trate th
e m
e
thod
s of
sp
arse
optimization probl
em. In Section III, we show the
aspect of over-complet
e di
ctionaries and its
approa
che
s
.
In Sectio
n IV
, we
illust
rat
e
the
impo
rt
ance of
sparse
co
ding
in
different
sp
e
e
ch
pro
c
e
ssi
ng a
pplication
s
.
2. Sparse M
odeling
It was
fi
rst introdu
ce
d in [8], [9] as a method to
fi
nd sparse line
a
r combinatio
ns
of basi
s
function
s to encode nat
ural im
age
s. Sparse repr
esentation of signal
s is a g
r
owi
ng
fi
eld of
resea
r
ch whi
c
h aim
s
at
fi
nding a set of prototype si
gnal
s calle
d atoms
∈
w
h
ic
h fo
r
m
s
a
diction
a
ry
∈
that can
be u
s
ed to represe
n
t a
pa
rticul
a
r
set of given
sign
als
∈
by
some
sp
arse
linear
com
b
in
ation of the atoms in t
he di
ctiona
ry. Mathematica
lly, for a given
set
of
sign
als rep
r
e
s
ente
d
by
Χ
, we nee
d to
fi
n
d
a suitable
diction
a
ry
suc
h
that
whe
r
e
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 25
02-4
752
IJEECS
Vol.
2, No. 1, April 2016 : 161 –
167
162
∈
is
a sp
arse vector whi
c
h contai
ns
th
e coe
ffi
cient
s f
o
r the li
nea
r
combi
nation
and
∈
Χ
.
2.1. Non-Co
nv
ex Sparse Optimizatio
n
Problem
The pro
b
lem
of sparse rep
r
esentation
can
thus be fo
rmulate
d
as
a non-co
nvex or (
ℓ
)
optimizatio
n probl
em of
fi
nd
in
g
,
whic
h satis
fi
es
,
∶
A
r
g
m
i
n
,
‖
‖
Subject
t
o
‖
‖
(1)
whe
r
e
is some
prede
fi
ne
d thre
sh
old whi
c
h
control
s
the
spa
r
sene
ss of the
rep
r
e
s
entatio
n and
‖
‖
denotes the
ℓ
pseu
do norm whi
c
h co
unts th
e numbe
r of
non-ze
ro
element
s of the vector
. This pro
b
lem
ca
n alternately
be formul
ated
as
,
∶
A
r
g
m
i
n
,
‖
‖
Subject
t
o
‖
‖
(2)
whe
r
e
is the
tolerabl
e limit of erro
r in
re
con
s
tru
c
tion.
Thoug
h the
solution
s to Eq
.1 and
Eq.2 nee
d
n
o
t be th
e
sa
me math
ema
t
ically, they
are
simil
a
r in
esse
nce to
what th
e
spa
r
se
rep
r
e
s
entatio
n probl
em ai
ms at achi
e
v
ing. This
problem is thu
s
involves a
choi
ce of the
diction
a
ry an
d a sp
arse li
near
co
mbin
ation of t
he
atoms in th
e
dictiona
ry to
rep
r
e
s
ent e
a
ch
desi
r
ed sign
a
l
.
2.2. Non-Co
nv
ex Sparse Optimizatio
n
Problem
Usi
ng the
ℓ
norm in th
e
sp
arse a
pproximation p
r
obl
em ma
ke
s it
a NP
-Hard
with a
redu
ction to
NP-compl
ete
sub
s
et
sele
ction p
r
obl
e
m
s
in com
b
i
natorial opti
m
ization.
A conve
x
relaxation of the probl
em can instea
d be
obtained by takin
g
the
ℓ
norm instead of the
ℓ
norm,
whe
r
e
‖
‖
∑
,
. The
ℓ
norm in
du
ce
s sp
arsity under
ce
rtain
conditio
n
s [
10]. The
solution of the convex opti
m
ization problem will be in
the form of
,
∶
A
r
g
m
i
n
,
‖
‖
Subject
t
o
‖
‖
(3)
Or
,
∶
A
r
g
m
i
n
,
‖
‖
Subject
t
o
‖
‖
(4)
Efforts devot
ed to this p
r
o
b
lem have
re
sulted
in th
e
cre
a
tion of a
numbe
r of al
gorithm
s
inclu
d
ing b
a
si
s pu
rsuit (BP) [11], matchin
g
pursu
it (MP
)
[12], orthog
onal mat
c
hin
g
pursuit (OM
P
)
[22], s
u
bs
pace pursuit (SP
)
[13], [14],
regres
s
i
on shrink
age
and sele
c
t
ion (LASSO) [15], focal
unde
r-determ
i
ned syste
m
solver
(F
O
CUSS) [16],
and gradi
e
n
t pursuit (GP) [17]. Sparse
decompo
sitio
n
s
of a
si
gnal
, however,
rel
y
gre
a
tly on t
he d
e
g
r
ee
of
fi
tting b
e
twe
e
n
the
data
an
d
the diction
a
ry
, which le
ad
s to the second
pr
oble
m
, i.e., the issu
e of diction
a
ry de
sign.
3. O
v
er-Com
plete Dic
t
ion
a
ries
An over-co
m
plete di
ctiona
ry, one in
which
th
e num
ber of ato
m
s is greate
r
than the
dimen
s
ion
of
the sig
nal,
ca
n be
obtaine
d by eithe
r
a
n
analytical o
r
a le
arning
-b
ase
d
ap
proach.
The analytical approa
ch
generates
the diction
a
ry based o
n
a prede
fi
n
ed mathem
a
t
ical
transfo
rm, su
ch a
s
di
scret
e
Fou
r
ier tran
sform
(D
FT), disc
rete c
o
s
i
ne
trans
f
orm (DCT), wavelets
[18], curvelet
s [19], contou
rlets [20], an
d bandel
ets
[21]. Such dictionarie
s are relatively easy to
obtain and
more
suitabl
e for gene
ric signal
s. In
learni
ng
-ba
s
e
d
approa
che
s
, howeve
r
, the
diction
a
rie
s
a
r
e ad
apted from a set of training
data
[8], [9], and [22]-[27]. Although thi
s
ma
y
involve
high
er comp
utational compl
e
xity,
lear
ned
diction
a
rie
s
have the
potential to
offer
improve
d
pe
rforma
nce a
s
compa
r
ed
with
pre
d
e
fi
ned
di
ctiona
rie
s
, sin
c
e the
atom
s
are
derive
d
to
captu
r
e the salient inform
a
t
ion
dire
ctly from the sign
al
s.
Evaluation Warning : The document was created with Spire.PDF for Python.
IJEECS
ISSN:
2502-4
752
Sparse Mod
e
ling with Appli
c
ation
s
to Speech Pro
c
e
s
sing: A Surve
y
(AN Om
ara
)
163
Optimizin
g
the di
ctiona
ry
is a
chall
engin
g
probl
em, and th
e
nume
r
ical strategy
comm
only e
m
ployed
co
n
s
ist
s
in ite
r
at
ive algo
rith
m
s
that
start f
r
om
an initia
l diction
a
ry a
nd
alternate b
e
twee
n the followin
g
step
s [28]:
Sparse co
din
g
: given a
fi
xed diction
a
ry
, the matrix
∅
of sparse ap
proximation co
e
ffi
cient
s
is cal
c
ul
ated
usin
g any suit
able al
g
o
rith
m for spa
r
se
approximatio
n.
Dictio
nary up
date:
given a
fi
xed a
p
p
r
oxi
m
ation mat
r
ix
∅
, the di
ctiona
ry
is
upd
ated
in orde
r
to minimize t
he re
sidu
al cost functio
n
‖
∅
‖
.
More speci
fi
cally, several
method
s hav
e bee
n propo
s
ed to fo
rmal
ize the
notio
n of the
suitability of a dictionary for sparse approximati
on. These include the mut
ual
coherence [29], the
cumul
a
tive coheren
ce [30
], the exact recove
ry coef
fi
ci
ent (E
RC) [30], the sp
ark [31], and
the
rest
ricte
d
iso
m
etry con
s
ta
nts (RI
C
s) [3
2], [33
]. Except for the mutual co
heren
ce and cumula
tive
coh
e
re
nce, none of the
s
e mea
s
u
r
e
s
can be ef
fi
ciently cal
c
ul
ated for an
arbitra
r
y given
diction
a
ry.
3.1. Mutual Coher
e
nc
e o
f
a Diction
a
r
y
The perfo
rma
n
ce of sp
arse
approximatio
n
algorithm
s
depe
nd
s on the mutual co
here
n
ce
of the dictionary
de
fi
n
ed a
s
the maximum absol
ute inner p
r
od
uct
between an
y two di
ff
ere
n
t
atoms.
≝m
a
x
〈
,
〉
(5)
The mutual
coheren
ce of a dict
ion
a
ry measures th
e simila
ri
ty betwee
n
the dictiona
ry's
atoms. F
o
r
an o
r
thog
on
al matrix
,
0
. For a
n
ove
r
-compl
ete ma
trix
we
necessa
rily h
a
ve
0
. There i
s
an intere
st i
n
diction
a
ri
es with
as
sm
all as p
o
ssibl
e
for spa
r
se re
pre
s
entatio
n purp
o
ses. If
1
,
it implies the existence of two parallel atoms,
and this cau
s
es ambi
guity in the con
s
tru
c
tion of sp
a
r
se atom comp
osition
s
. In [34] it was sho
w
n
that for a full rank di
ction
a
ry of size
1
(6)
and equ
ality is obtaine
d for a family of dictionari
e
s called Grassmannia
n
fra
m
es. For
≫
the mutual coheren
ce we can exp
e
ct
to
have is thus
of the orde
r o
f
1
√
⁄
.
3.2. Cumulativ
e
Coheren
ce of a Dic
t
i
onar
y
A re
fi
nem
ent
of the cohe
rence pa
ram
e
ter is th
e
c
u
mu
la
tive
co
he
r
e
nc
e fu
nc
tio
n
[3
5
]
,
[36]. It measure
s
ho
w m
u
ch a
coll
ect
i
on of
atom
s can resem
b
le a
fi
xe
d, distin
ct atom
.
Formally [1]
≝m
a
x
|
|
max
∉
|〈
,
〉|
∈
(7)
We pla
c
e the
convention t
hat
0
0
. The sub
s
cript on
serv
es as a mn
e
m
onic that
the cumulativ
e
cohe
ren
c
e
is an absol
ute
sum, and it distinguishe
s
the function
from the
numbe
r
μ
. When the cumulat
i
ve coheren
ce grows slo
w
l
y
,
we say info
rmally that th
e dictiona
ry is
inco
herent or
qua
si-in
c
o
herent.
3.3. Spark of a Dictionary
The spark of
a diction
a
ry
is the sm
a
llest num
ber
of colum
n
s t
hat form a li
nearly
depe
ndent
se
t [37]. In-spite the
simil
a
r
de
fi
nition, n
o
te that
spa
r
k i
s
m
a
rkedly
di
fferent fro
m
t
he
matrix
ran
k
, being
the greatest num
b
e
r of
line
a
rly
inde
pen
dent
col
u
mn
s. A
trivial
relati
on
betwe
en the
spa
r
k
and the mutual co
h
e
ren
c
e
is
[37].
1
1
(8)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 25
02-4
752
IJEECS
Vol.
2, No. 1, April 2016 : 161 –
167
164
4. Speech Pr
ocessin
g
Ba
sed on Spar
se Modeling
Sparse mo
de
ling ha
s ubiq
u
itous a
ppli
c
ations
in
spe
e
ch a
nd au
di
o pro
c
e
s
sing
area
s,
inclu
d
ing di
mensi
onality
redu
ction,
model re
gulari
z
atio
n, spee
ch
compressio
n
and
recon
s
tru
c
tio
n
, acou
stic/a
udio fe
ature
sele
ction,
acousti
c m
odeli
ng, spee
ch
reco
gnition,
bl
ind
sou
r
ce sepa
ration, and m
any others. T
h
is sect
io
n prese
n
ts some
efforts on
sp
eech processing
usin
g sp
arse
modelin
g.
4.1. Speaker
Identifica
tio
n
The
autho
rs i
n
[38] i
n
trod
u
c
ed
a
novel
method
for spea
ker ide
n
ti
fi
cation or
det
ermini
ng
an un
kno
w
n
spe
a
ker’
s ide
n
tity based o
n
a sp
arse
si
gnal mo
del a
nd the u
s
e o
f
Comp
re
sse
d
Sensin
g (CS
)
. The use
of CS permit
s
the use of
less tra
n
sm
issi
on po
we
r for the sen
s
or
recording
the
voice. A
dditionally, this
method
had
been
sho
w
n
to be
rob
u
st
to noi
se
in
the
recorded
spe
e
ch
sign
al. This is e
n
cour
aging a
nd wa
rra
nts furth
e
r
investigatio
n.
4.2. Speech
Compre
ssio
n
In [39],
the author present
ed the Molecular Ma
tchin
g
Pursuit (M
MP) algorith
m
that is
suitabl
e for
spee
ch
codin
g
.
The main
g
oal of MMP i
s
to ma
ke a
pra
c
tical
de
compo
s
ition
such
that at every iteration the
algorithm id
enti
fi
e
s
an
d remove
s a
whol
e clu
s
ter of (orthog
o
nal)
atoms. At the cost of a sli
g
ht
sub optim
a
lity in the approximation e
r
ror
rate, this
offers a n
u
m
ber
of advantage
s, most
notably it is signi
fi
cantly faster since the i
nner p
r
odu
cts update step
is
made for a la
rge num
be
r of atoms at every iterati
on. Also, the use
of a Modified Discrete Co
si
ne
Tran
sfo
r
m (MDCT) for
spee
ch codin
g
wa
s inve
st
igated in [40
]. This appro
a
ch p
r
od
uce
s
a
spa
r
ser
de
co
mpositio
n tha
n
the tra
d
ition
a
l MDCT
-ba
s
ed o
r
thog
ona
l tran
sform
a
nd allo
ws bet
ter
codi
ng ef
fi
cie
n
cy at low bitrates. Contra
ry to state-
of-the-a
r
t low bit
r
ate co
de
rs,
whi
c
h are ba
sed
on pure pa
ra
metric o
r
hybrid rep
r
e
s
entat
ions,
the ap
proach is abl
e to provide tran
spa
r
en
cy.
4.3. Blind Source Separ
a
tion
Und
e
rd
etermi
ned
spee
ch
sep
a
ratio
n
is a chall
engi
n
g
pro
b
lem th
at has b
een
studie
d
extensively in recent yea
r
s. The a
u
th
or in
[56] prese
n
ted a promisin
g met
hod to the Blind
Sourc
e
Separation (BSS) for s
p
eech s
i
gnals
ba
sed on spars
e
represent
ation with adaptive
diction
a
ry le
arnin
g
. In
a
nother
work [58
], the
author
sho
w
ed that
the
use
of sp
arse
decompo
sitio
n
in a p
r
op
er si
gnal
dictionary p
r
ovides
high
-qu
a
lity blind so
urce sepa
rati
on.
More
over, h
e
proved th
at the maxi
mum a
po
st
erio
ri fra
m
e
w
ork
gives the mo
st g
e
neral
approa
ch,
which i
n
cl
ude
s the situ
atio
n of mo
re
source
s tha
n
sen
s
o
r
s. In
[41], the aut
hor
addressed the convol
utiv
e BSS issue
and suggest
ed a soluti
on using sparse
Independent
Comp
one
nt Analysi
s
(ICA).
4.4. Speech
Enhanceme
n
t
Re
cently, sp
arse
rep
r
e
s
e
n
tation i
s
wi
dely u
s
ed
f
o
r
sp
ee
ch
p
r
ocessin
g
in
noi
sy
environ
ment
s; howeve
r
,
many proble
m
s n
eed to
be solved b
e
ca
use of th
e pa
rticula
r
it
y of
spe
e
ch. In [4
2], a n
o
vel vi
ew fo
r th
e e
n
han
ceme
nt of
sig
nal
s
wa
s
applie
d
su
cce
ssfully to
spe
e
ch
usin
g the K
-
Singula
r
Val
ue Decomp
o
s
ition Alg
o
rit
h
m (K
-SVD)
[22]. The K-SVD algo
rith
m is
desi
gne
d for
training
an
ov
er-com
plete
diction
a
ry
tha
t
best
suits
a
set of
given
signal
s. Anoth
e
r
spe
e
ch e
nha
ncem
ent te
ch
nique
was su
gge
sted i
n
[5
7] wh
en th
e a
u
thor
propo
sed a
n
exem
pl
ar-
based te
chni
que for th
e n
o
isy spee
ch.
The te
chniq
u
e
wo
rks by
fi
nding
a sparse rep
r
e
s
e
n
tation
of the n
o
isy
spee
ch i
n
a
di
ctiona
ry cont
aining
both
spee
ch
and
n
o
ise
exempl
a
r
s,
and
u
s
e
s
the
activated di
ctionary atom
s to cre
a
te a time-varyin
g
fi
lte
r
to enhan
ce t
he noi
sy spe
e
ch.
A good
effort
wa
s d
one i
n
[43]; the aut
hor
pro
p
o
s
ed
an effe
ctive dual
-chan
ne
l noise
redu
ction
alg
o
rithm b
a
sed
on sp
arse
re
pre
s
entatio
ns. The algo
rith
m is compo
s
ed of four
ste
p
s.
Firstly, ove
r
la
pping
pat
che
s
sam
p
led
from two
cha
n
nels togeth
e
r inste
a
d
of e
a
ch
chan
nel
one
by one
are
trained to
be
a dictio
na
ry via K-SVD.
Secon
d
ly, O
M
P re
con
s
tru
c
tion al
go
rith
m is
applie
d to
ob
tain the
sp
arse
co
ef
fi
cient
s of
pat
che
s
usin
g the
di
ctionary. T
h
irdly, the de
noisi
n
g
spe
e
ch can
b
e
obtain
ed b
y
the upd
ate
d
co
ef
fi
cient
s. Lastly, the
above th
ree
step
s a
r
e ite
r
ated
to get clea
rer spee
ch u
n
til some
con
d
itions a
r
e rea
c
hed. Experim
ental re
sults
sho
w
that this
algorith
m
perf
o
rm
s better t
han that with
singl
e ch
ann
el.
Another spe
e
ch
den
oisi
n
g
metho
d
b
a
s
ed
on
gree
dy ortho
gon
a
l
adaptive
di
ctiona
r
y
learni
ng was propo
se
d in
[25]. The algorithm
con
s
truct
s
a use
r
-de
fi
ne
d co
m
p
lete dictio
na
ry,
who
s
e
atom
s clea
rly en
co
de lo
cal p
r
o
p
e
rties of
the sign
al.
The p
e
rform
a
n
c
e o
f
the
algo
rithm
Evaluation Warning : The document was created with Spire.PDF for Python.
IJEECS
ISSN:
2502-4
752
Sparse Mod
e
ling with Appli
c
ation
s
to Speech Pro
c
e
s
sing: A Surve
y
(AN Om
ara
)
165
wa
s compa
r
e
d
to that of the Prin
cipal
Compon
ent
An
alysis
(PCA
) method, an
d
it was fou
nd t
o
give good
sig
nal app
roxim
a
tions, eve
n
as the
numb
e
r
of atom
s in t
he re
co
nst
r
u
c
tions d
e
crea
ses
con
s
id
era
b
ly; it was also o
b
se
rved that the al
gorithm
has goo
d toleran
c
e to noi
se, com
p
a
r
ab
le
to that afforded by PCA.
The enh
an
ce
ment of spe
e
c
h de
gra
ded
by non-
statio
nary interfe
r
e
r
s
wa
s add
re
ssed in
[55]. The a
u
thor
presente
d
a m
ona
ural
sp
ee
ch
e
n
h
ancement m
e
thod ba
sed on spa
r
se co
ding
of noisy spee
ch si
gnal
s in
a com
p
o
s
ite diction
a
ry
, co
nsi
s
ting of th
e con
c
ate
nati
on of a sp
ee
ch
and i
n
terfe
r
e
r
dictio
na
ry, b
o
th bei
ng
po
ssi
bly ove
r
-complete. T
h
e
sp
ee
ch
dicti
onary
is lea
r
ned
off-line o
n
a t
r
aining
co
rp
us, while
an
env
ironm
ent spe
c
i
fi
c int
e
rfe
r
er diction
a
ry i
s
l
earn
ed
on-li
n
e
durin
g sp
ee
ch pau
se
s.
4.5. Speech
Reco
gnition
Most of auto
m
atic spee
ch
recognitio
n
(
ASR) technol
ogie
s
are ba
sed on hi
dde
n
Markov
model
s (HM
M
s),
whi
c
h m
odel a time
-v
arying
spe
e
ch sig
nal u
s
in
g a sequ
en
ce
of states, e
a
c
h of
whi
c
h i
s
a
s
so
ciated
with a
distrib
u
tion of
aco
u
sti
c
feat
ure
s
. While HMMs
rea
c
h a
relatively hig
h
perfo
rman
ce
in
go
od co
n
d
itions,
th
ey have
p
r
obl
e
m
s i
n
mo
deli
ng
wide
vari
ances in
nat
ural
spe
e
ch
sign
als,
su
ch
a
s
spe
e
ch in
natu
r
al
en
vironme
n
ts
whi
c
h i
s
often inte
rfered
by
environ
menta
l
noise
s.
Re
cently, so
me studi
es
[44-45], an
d
[51-54] h
a
ve aimed at
ASR usin
g
spa
r
se
rep
r
e
s
entatio
ns of sp
ee
ch.
In them, a time-fre
que
ncy
representati
on of spe
e
ch is as a
weig
h
t
ed
linear
com
b
i
nation of sp
eech atom
s. Bene
fi
t
s
of the existing
system
s ra
n
ge from imp
r
oved
recognitio
n
a
c
cura
cy to
an
ea
sy in
co
rpo
r
ation
of
robu
stne
ss to
add
itive noises.
Some of
the
s
e
system
s con
s
tru
c
t the di
ctiona
ry of atoms to
be
use
d
in the
spa
r
se rep
r
ese
n
tation from
exempla
r
s of
sp
ee
ch,
whi
c
h
are
realizations of
sp
e
e
ch
in th
e training
data,
spa
nnin
g
mul
t
iple
time frames
[54].
Whe
n
the we
ights of the
sparse rep
r
e
s
entati
on a
r
e
use
d
directly
in the re
cog
n
ition,
a
fundame
n
tal
probl
em is t
he asso
ciati
on of high
er-level inform
ation with th
e atoms in
the
diction
a
ry to enabl
e the reco
gnition. In
[45],
the author trai
ned a
neural n
e
twork to ma
p the
weig
hts
of th
e atom
s
dire
ctly to phon
em
e cl
asse
s.
Where
a
s in [
5
3
], the autho
r
asso
ciated
e
a
ch
atom with
on
e pho
netic
cl
as
s, and
re
co
gnition
wa
s d
one by
fi
ndi
n
g
the ph
one
me cla
s
s
with
the
highe
st sum
of weig
hts. A
l
so in [5
2], the autho
r
u
s
e
d
a di
ctiona
ry con
s
istin
g
of both a
c
ou
stic
informatio
n a
nd hig
her-lev
el pho
netic i
n
formation.
Bu
t in [51], the
author u
s
ed t
he ind
e
x of the
spe
e
ch atom
with the hi
ghe
st weig
ht as an
a
ddit
i
onal feature
for their Dy
namic Baye
sian
Network re
co
gnizer.
Beside
the fo
regoi
ng effo
rts, there a
r
e
mo
re re
se
arches on spe
e
c
h re
cog
n
ition
ba
sed
on spa
r
se re
pre
s
entatio
ns [46-50]. In [46], t
he author enh
an
ced
the Least Ab
solute Shri
nkage
and Selection Operator (LASSO) algo
ri
thm for improving the speech
recognition rates. In [50],
the auth
o
r
used the
sp
arse
rep
r
e
s
ent
atio
n to e
s
timate
the missin
g (unreli
able
)
co
efficients of t
h
e
spe
e
ch si
gn
al. In [47], the auth
o
r
h
ad evalu
a
te
d
the sparsity assum
p
tion
s in
co
rpo
r
ate
d
in
spa
r
se com
p
onent analy
s
i
s
in the fram
ewo
r
k of De
gene
rate Un-mixing Estimation Techni
que
(DUET) fo
r spee
ch recog
n
ition in a m
u
lti-sp
ea
ke
r e
n
vironm
ent. In [48], the author p
r
op
osed a
state-b
a
sed l
abelin
g for a
c
ou
stic patterns
of spee
ch
and
a m
e
th
od for u
s
ing
this la
beling
in
noise robu
st
automatic sp
eech recogni
tion. In
[49],
a fram
ework
for an
exem
plar-ba
s
e
d
, d
e
-
convol
utive spee
ch re
co
gn
it
ion system
wa
s pre
s
e
n
te
d.
5. Conclusio
n
This
study
sh
eds light o
n
t
he a
pplicatio
ns
of
the
spa
r
se
mo
deling
in the field
of
sp
ee
ch
pro
c
e
ssi
ng. A
l
though
the
sparse
mod
e
li
ng i
s
ju
st
a
si
gnal
de
comp
osition
techni
que, thi
s
su
rvey
sho
w
e
d
the importa
nce of this strategy
in the
spee
ch sou
r
ce sep
a
ration, spee
ch comp
ressi
on,
spe
a
ker id
ent
ification, spe
e
ch
re
cog
n
ition an
d noi
se
redu
ction.
Not only the sp
arsene
ss of t
h
e
rep
r
e
s
entatio
n plays a
role
in the
s
e
wid
e
appli
c
atio
ns, but also the
choi
ce
of the
diction
a
ry pl
a
y
s
an imopo
rtant
role in this variation of the
application
s
.
Referen
ces
[1]
JA
T
r
opp.
T
opics in sp
ar
se appr
o
x
imat
ion.
Ph.D. Di
ssertation, Co
mputatio
nal a
nd App
lie
d
Mathematics, T
he
Universit
y
of
T
e
xas at Au
stin. 2004.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 25
02-4
752
IJEECS
Vol.
2, No. 1, April 2016 : 161 –
167
166
[2]
P F
ö
ldiák. F
o
rming spars
e
repres
ent
ati
ons
b
y
loc
a
l anti-
Heb
b
ia
n lear
ni
ng.
Biol
ogic
a
l Cyber
netics
.
199
0; 64: 165-
170.
[3]
Le
w
i
ck
i MS, Sejn
o
w
ski T
J
. "Lear
nin
g
ov
ercompl
e
te repr
e
s
entatio
ns".
N
eura
l
Co
mput
. 200
0; 12(2)
:
337-
65.
[4]
W
Dai, T
Xu, a
nd W
W
ang.
"
Simulta
neo
us
Cod
e
w
o
rd Opti
mi
z
a
t
i
o
n
(Si
m
C
O
) for Diction
a
r
y Lear
nin
g
".
in Proc.
49th
Annu
al A
llerto
n
Co
nfere
n
ce
on C
o
mmu
nic
a
tion, Contro
l, and
Com
putin
g
(AL
L
ERT
O
N
201
1), Montice
llo, Illin
ois, US
A. 2011.
[5]
S Mallat. A W
a
velet T
our of Signa
l Processi
n
g
. 2nd e
d
. Ne
w
York: Academi
c
. 1998.
[6]
David
L Do
no
ho an
d Mich
a
e
l Ela
d
.
Optima
lly sp
arse r
epres
entati
on
in ge
ner
al (no
northo
g
o
nal
)
dictionaries via
¹
mini
mi
z
a
t
i
on
. Proc. Natl. Acad. Sci. USA 100. 200
3.
[7]
JJ F
u
chs. “On
sparse r
epr
ese
n
tations
in
arb
i
trar
y
r
edu
nd
an
t bases”.
IEEE
T
r
ans. Inf. T
heory
. 2
004
;
50: 134
1–
13
44
.
[8]
Bruno A Olsha
u
sen, Dav
i
d J F
i
eld. Emerg
e
n
ce of
simple-
c
ell rece
ptive fi
eld pr
operti
es b
y
l
ear
nin
g
a
sparse co
de fo
r natural im
age
s.
Nature
. 199
6; 381(6
5
8
3
): 607-9.
[9]
Olshaus
en BA,
F
i
eld DJ. “Sp
a
rse Co
din
g
w
i
th an Overcom
p
lete Bas
i
s Set
:
A Strateg
y
E
m
plo
y
e
d
b
y
V1?".
Visi
on R
e
searc
h
. 199
7; 37: 331
1-33
25
.
[10]
David
L Don
o
h
o
. “F
or most large un
derd
e
ter
m
ine
d
s
y
stems
of linear e
qua
tions, the mini
mal 1-n
o
r
m
near-s
oluti
on appr
o
x
imates
the
sp
arsest
n
ear-sol
ut
io
n,”
Dept. of Statis
ti
cs, Stanford
Univ. 2
0
0
4
,
T
e
ch. Rep. 2004-1
1
.
[11]
Huggins,
PS,
Zucker,
SW. "Greedy
Basis
Pursuit".
IEEE
Transactions on Signal Proc
essing
. 20
07;
55(7): 37
60-
37
72.
[12]
Mallat
SG, Z
h
i
f
eng Z
h
an
g. “
M
atchin
g p
u
rs
uits w
i
t
h
time-f
requ
enc
y
dicti
onar
ies”.
IEEE
T
r
ansacti
ons
on Sig
nal Pr
oc
essin
g
. 199
3; 41(12): 33
97-
34
15.
[13]
W
e
i Dai; Mi
le
n
k
ovic O. "Sub
space P
u
rsuit
for Compr
e
ssi
ve Sens
in
g Si
gna
l Rec
onstr
uction".
IE
EE
T
r
ansactio
n
s o
n
Information T
heory
. 20
09; 5
5
(5): 223
0-2
2
4
9
.
[14]
D Ne
ede
ll a
nd
JA T
r
opp Co S
a
MP. Iterative
sign
al rec
o
ver
y
from incom
p
le
te and
in
accur
a
te samp
le
s
Appl
ied a
nd C
o
mputati
o
n
a
l H
a
rmonic A
nal
ysis. 2009
[15]
T
i
bshirani R.
Regr
essio
n
sh
rinka
ge an
d s
e
lecti
on via th
e Lasso.
J. Roy. Statist.Soc.
Ser. B
. 1996;
58(1): 26
7-2
8
8
.
[16]
Gorodn
itsk
y
IF
, Rao BD. "S
parse s
i
gn
al r
e
constructi
on f
r
om limite
d
d
a
t
a usin
g F
O
CUSS: a re-
w
e
ighted minimum norm algorithm".
IEEE Transactions on Signal Proc
essing
. 19
97; 45
(3): 600-6
16.
[17]
Blume
n
sath
T
,
Dav
i
es ME.
"
G
radie
n
t Purs
uits".
IEEE Tr
ansactions
on
Signal Proc
essing
. 20
08;
56(6): 23
70-
23
82.
[18]
Seles
n
ick IW
, Baran
i
uk RG,
Kings
bur
y N
C
. "T
he dual-tre
e
compl
e
x
w
a
v
e
let transform".
IEEE Signal
Processi
ng Ma
ga
z
i
n
e
. 20
05; 22(6): 12
3- 15
1.
[19]
EJ Cand
es, DL Don
oho et
al. Curvel
ets:
A surprisin
g
ly
effective non
a
daptiv
e repres
entatio
n for
obj
ects w
i
th edges
. Dep
a
rtme
nt of Statistics,
Stanford Un
ive
r
sit
y
. 19
99.
[2
0
]
MN
D
o
a
n
d
M Ve
tte
rl
i
.
T
h
e
co
n
t
ou
rl
e
t
tra
n
s
fo
rm: a
n
effi
ci
e
n
t
d
i
re
ction
a
l
mul
t
i
r
e
s
ol
uti
o
n
i
m
age
repres
entati
on.
IEEE Transactions on Im
age
Processing
. 20
05; 14(1
2
): 209
1–2
10
6.
[21]
Le P
enn
ec E,
Mallat S.
"Sp
a
r
s
e g
eomet
ric
i
m
age
repr
ese
n
t
ations
w
i
th
ba
nde
lets".
IEEE Transactions
on Image Proc
essin
g
. 200
5; 14(4): 423-
43
8.
[22]
M Ahar
on, M
Elad and A
B
r
uckstein. "T
he K-SVD: A
n
algorith
m
for
designing of
overcomplet
e
dictio
nari
e
s for sparse re
pres
e
n
tation
".
IEEE Transactions on Signal Proc
essing
. 2
006; 5
4
(11): 4
311-
432
2.
[23]
W
i
pf DP, Rao BD. "Sparse Ba
yesi
an le
ar
nin
g
for basis
selectio
n".
IEEE Transactions on Signal
Processi
ng
. 20
04; 52(8): 2
153
- 2164.
[24]
MD
Plumbl
e
y
.
Diction
ary Le
arni
ng for L1-
Exact Sparse
Codi
ng
. In ME Davies, CJ
James, SA
Abda
lla
h an
d
MD Plumb
l
e
y
(
eds.), Procee
di
ngs of
the 7th
Internatio
na
l C
onf
ere
n
ce o
n
Inde
pe
nde
nt
Comp
one
nt An
al
ysis
an
d Si
gn
al S
epar
atio
n, ICA 2
007,
Lo
nd
on, UK,
9-12
S
ept 2
0
0
7
, LN
C
S
46
66, p
p
406-
413, Spr
i
n
ger, Berli
n
.
[25]
MG Jafari and MD Plumbl
e
y
.
Speec
h de
nois
i
ng b
a
sed o
n
a
greedy
a
dapti
v
e dictio
nary al
gorith
m
. I
n
Proc. 17th E
u
rope
an S
i
gn
al
Processi
ng
C
onfere
n
ce
(E
U
S
IPCO 200
9)
Glasgo
w
,
Sc
ot
lan
d
. 20
09:
142
3-14
26.
[26]
Rémi Grib
onv
al, Karin Sc
h
nass: Dictio
n
a
r
y
id
e
n
tificati
on: spars
e
matrix-factor
i
z
a
tion vi
a l1-
minimiz
a
tion.
IEEE Transacti
ons on Infor
m
a
t
ion Theory
. 2
0
10; 56(7): 3
523
-353
9.
[27]
Q Geng, H W
ang, a
nd J W
r
ight. "
On the
Loca
l
Correct
ness of L
^
1
Mini
mi
z
a
t
i
o
n
for Dictio
nar
y
Lear
nin
g
". Pre
s
ented at C
o
R
R
. 2011.
[28]
Barchi
esi D, Pl
umbl
e
y
M
D
. "L
earn
i
ng
inco
he
rent
dicti
onar
ie
s for sparse a
p
p
ro
ximati
on
usi
ng iter
ativ
e
proj
ections a
n
d
rotations".
IEEE Transactions
on Signal Proc
essing
. 201
3; (99): 1, 0.
[29]
Don
oho
DL, H
uo
X. "U
ncerta
int
y
pri
n
cip
l
es and id
eal atom
ic
dec
ompos
iti
on".
IEEE Transactions o
n
Information Theory
. 200
1; 47(
7): 2845-
28
62.
[30]
T
r
opp JA. "Just rela
x: conve
x
programm
i
ng
me
thods for i
d
entif
yin
g
spars
e
sign
als i
n
no
ise".
IEEE
T
r
ansactio
n
s o
n
Information T
heory
. 20
06; 5
2
(3): 103
0-1
0
5
1
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IJEECS
ISSN:
2502-4
752
Sparse Mod
e
ling with Appli
c
ation
s
to Speech Pro
c
e
s
sing: A Surve
y
(AN Om
ara
)
167
[31]
Don
oho D
L
, Elad M,
T
e
ml
y
a
k
o
v VN. "Stable
recove
r
y
of sparse overcom
p
lete repr
esenta
t
ions in t
h
e
prese
n
ce of no
ise".
IEEE Transactions o
n
Informatio
n
Theor
y
. 2006; 52(
1): 6- 18.
[32]
Can
des EJ
a
n
d
T
ao T
.
T
he Dantzi
g se
lect
or: stat
i
s
ti
ca
l esti
ma
ti
o
n
w
hen
p i
s
mu
ch l
a
rg
e
r
tha
n
n.
Anna
ls of Statistics
. 2007; 35(
6): 2313-
23
51.
[33]
EJ Can
des, J
K
Romb
erg a
nd T
T
ao. "Stabl
e
Sig
nal
R
e
cover
y
from
Incompl
e
te an
d Inaccur
a
t
e
Measur
ements
"
.
Communic
a
ti
ons on Pur
e
an
d Appl
ied M
a
th
ematics
. 200
6; LIX: 12
07
–12
2
3
.
[34]
T
Strohmer and RW
H
e
ath Jr. “
Grassma
n
n
i
an F
r
a
m
es w
i
th Ap
plicati
ons to
Codi
ng
an
d
Co
mmun
icati
o
ns
”. in Appl
ied
and C
o
mput
ati
ona
l Harmo
nic
Anal
ys
is. 200
3; 14(3): 257-
27
5.
[35]
T
r
opp JA. "Greed
is g
o
o
d
: alg
o
rithmic r
e
s
u
lts for sp
arse
appr
o
x
imati
o
n
"
.
IEEE Transactions
on
Information Theory
. 200
4; 50(
10): 223
1- 22
4
2
.
[36]
D D
ono
ho
a
n
d
M El
ad.
"
Maxi
ma
l
S
parsity R
epres
entati
on via l1
M
i
ni
mi
z
a
tion
". T
he Pr
oc
. Nat. Aca
.
Sci. 2003; 1
00:
2197-
22
02.
[37]
M Aharon, M
Elad, and AM
Bruckstein. "On the uniq
ueness of
overc
o
mplete dictionaries, and
a
practical
w
a
y
to retrieve them".
J. Linear Alg
ebra Ap
pl
. 20
0
6
; 416: 48-
67.
[38]
A Griffin, E Karamicha
li, a
nd
A Mouchtar
is. "
Spe
a
ker Ide
n
tificatio
n
usi
ng
Sparse
ly Excit
ed Sp
eec
h
Sign
als
a
nd Co
mpress
ed Sensi
n
g
". i
n
Proc. 18th E
u
rop
ean
Sig
n
a
l Proc
essin
g
Confer
enc
e
(EUSIPCO '
10), Aalborg, De
n
m
ark. 2010.
[39]
Dau
det L. "Sp
a
rse an
d stru
c
t
ured dec
omp
o
s
itions of si
gn
als
w
i
th th
e molec
u
lar match
i
ng p
u
rsuit"
.
IEEE Transactions on Au
di
o, Speec
h, and L
ang
ua
ge Proce
ssing
. 20
06; 14
(5): 1808-
18
16
.
[40]
Ravel
li E, Rich
a
rd G, Daud
et L.
"Unio
n
of MDCT
Bases for Audi
o Co
di
ng".
IEEE Transactions
on
Audi
o, Speec
h, and La
ng
ua
ge
Processin
g
. 2008; 16(
8): 136
1-13
72.
[41]
Jafari MG, Ab
dall
ah SA, Pl
u
m
ble
y
MD, D
a
vies ME.
Spar
se Co
din
g
for
Conv
olutiv
e
Blin
d Aud
i
o
Source Se
par
a
t
ion
. In ICA. 2006: 132-
13
9.
[42]
Z
hao N
an,
Xu
Xin
an
d Ya
n
g
Yi. "S
parse
Repr
ese
n
tatio
n
s for Sp
eec
h
Enh
ancem
ent
".
Chi
nes
e
Journ
a
l of Elec
tronics
. 201
1; 19(2).
[43]
Jianc
hen
g Z
o
u
,
Jingsai
Lia
ng,
Xi
n Yan
g
. "
Du
al-ch
ann
el n
o
is
e reducti
on vi
a
sprase re
pres
entatio
ns
".
201
2 IEEE 14th Internatio
na
l Workshop o
n
Multim
ed
ia Si
g
nal Proc
essin
g
(MMSP). 2012
: 221-22
5.
[44]
B King
and
L
E
Atlas. "
Sing
l
e
-cha
nne
l so
u
r
ce sep
a
ratio
n
usin
g si
mpl
i
fi
ed-trai
nin
g
co
mp
lex
matrix
factori
z
ation
". i
n
Proc. ICASSP. 2010: 42
06-
420
9.
[45]
GSVS Sivaram, SK Nem
a
la, M El
hilali,
T
D
T
r
an and
H Herm
ansk
y
.
"
Sp
arse c
odi
ng for
spe
e
c
h
recog
n
itio
n
". in Proc. ICASSP. 2010: 43
46-
434
9.
[46]
QF
T
an,
PG Georgi
ou,
a
n
d
SS Nara
ya
na
n. “Enha
nced
sparse im
putat
ion tech
ni
ques
for a robus
t
speec
h reco
gn
ition front-e
nd”
.
IEEE Trans.
Audi
o, Speec
h
,
Lang
uag
e Pr
ocess
. 20
11; 1
9
(8): 241
8-
242
9.
[47]
Afsaneh As
ae
i
,
Hervé Bo
url
a
rd, an
d Phi
l
i
p
N Garner. “
Sparse c
o
mpo
nent a
nalys
is
for speec
h
recog
n
itio
n
in mu
lti-spe
a
ker envir
on
me
nt
". INT
E
RSPEECH. 2010: 1
704-
170
7.
[48]
T
Virtanen, JF
Gemmeke, a
n
d
A H
u
rma
lai
n
en. "
State-
bas
ed
lab
e
ll
in
g for
a
s
parse
repr
esentati
o
n
of
speec
h an
d its app
licati
on to r
obust spe
e
ch r
e
cog
n
itio
n
". in
Proc. IN
T
E
RSPEECH. 201
0: 893-
896.
[49]
Antti Hurmala
i
nen, Jort F
Ge
mmeke, T
uomas Vi
rtane
n: Non-n
egativ
e matrix
d
e
co
nvol
ution i
n
noi
s
e
robust spe
e
ch
recog
n
itio
n.
ICASSP
. 2011: 4
588-
459
1.
[50]
J Gemmeke,
B Cran
en.
"
Us
ing
spars
e
re
p
r
esentati
ons f
o
r missin
g
d
a
ta
i
m
put
ation
in
nois
e
ro
bust
speec
h reco
gni
tion
". Proc. of EUSIPCO. 200
8
[51]
Schul
ler B, Weni
nger F
,
W
ö
llm
er M, Sun Y, Rigoll G. "Non-ne
gativ
e mat
r
ix factorizati
o
n as noise-
robust featur
e
extract
o
r for s
peec
h reco
gn
ition".
201
0 IEE
E
Internatio
na
l
Confer
ence
o
n
Acoustics
Speec
h an
d Si
gna
l Processi
n
g
(ICASSP)
. 2010: 45
62-
456
5.
[52]
Maarten
Va
n S
egbr
oeck,
Hug
o
Va
n
Hamme.
Uns
uperv
i
se
d lear
nin
g
of
tim
e
-frequ
enc
y pa
tches
as
a
nois
e
-rob
u
st represe
n
tatio
n
of speech.
Sp
ee
ch Co
mmu
n
ica
t
ion
. 200
9; 51(
11): 112
4-1
138
.
[53]
Sain
ath T
N
, C
a
rmi A, Kanev
sk
y
D, Ram
a
b
hadr
an B. "Ba
y
es
ia
n compre
ssive sens
ing
for phon
et
i
c
classification".
2010 IEEE I
n
ternatio
nal Conference
on
Acoustics S
p
e
e
ch
an
d Si
gn
al Pr
ocessi
n
g
(ICASSP)
. 201
0: 4370-
43
73.
[54]
Gemmeke JF
, Virtan
en T
.
"Noise
ro
bust e
x
em
pl
ar-bas
ed
con
nected
d
i
git rec
ogn
itio
n"
.
201
0 IEE
E
Internatio
na
l C
onfere
n
ce o
n
Acoustics Spe
e
c
h
and Si
gn
al P
r
ocessi
ng (ICA
SSP)
. 2010;: 4
546-
454
9.
[55]
Sigg C
D
, Dikk
T
,
Buhmann
JM. "Speech
e
nha
nceme
n
t
w
i
th sparse c
o
d
i
ng in
lear
ne
d
dictio
nari
e
s".
in
2010 IEEE I
n
ternational Conferenc
e on
Acoustics Speec
h and Si
gnal P
r
ocessing (ICA
SSP)
. 201
0:
475
8-47
61.
[56]
T
ao Xu, W
e
n
w
u
W
a
n
g
. "M
ethods f
o
r le
a
r
ni
n
g
ad
aptiv
e
dictio
nar
y i
n
und
erd
e
termi
ned s
p
e
e
ch
separ
ation
"
. 20
11
IEEE Intern
ation
a
l Worksh
op o
n
Mach
ine
Lear
nin
g
for S
i
gn
al Proc
essi
ng (MLSP)
.
201
1: 1-6.
[57]
Gemmeke JF
, Virtane
n T
,
Hurmal
a
i
nen
A. "E
xempl
a
r-
Based S
pars
e
Repr
esent
atio
ns for No
is
e
Robust Automatic Speec
h Re
co
gn
i
t
i
o
n
"
. in
IEEE Transactions
on A
u
dio,
Speech,
and Language
Processi
ng
. 20
11; 19(7): 2
067
-208
0.
[58]
M Z
i
bul
evsk
y
,
BA Pe
arlmut
ter, P Bofil
l
and
P Kis
ilev
.
"
Blin
d S
our
ce Se
parati
o
n
by S
par
s
e
Deco
mpositi
o
n
". Chapter i
n
the bo
ok: SJ Roberts, and
R
M
Everson ed
s. Indepen
de
nt Compo
n
e
n
t
Anal
ys
is: Princi
ples a
nd Practi
ce, Cambri
dge.
2001.
Evaluation Warning : The document was created with Spire.PDF for Python.