TELK
OMNIKA
T
elecommunication,
Computing,
Electr
onics
and
Contr
ol
V
ol.
23,
No.
6,
December
2025,
pp.
1626
∼
1634
ISSN:
1693-6930,
DOI:
10.12928/TELK
OMNIKA.v23i6.26893
❒
1626
Analyzing
temporal
pr
operties
of
speech
trajectory
using
graph
structur
es
to
wards
speech
r
ecognition
P
arabattina
Bhagath
1
,
Malempati
Shanmukha
2
,
Gnana
Nagasri
Puthi
2
1
Department
of
Articial
Intelligence
and
Data
Science,
Lakireddy
Bali
Reddy
Colle
ge
of
Engineering,
Myla
v
aram,
India
2
Department
of
Computer
Science
and
Engineering,
Lakireddy
Bali
Reddy
Colle
ge
of
Engineering,
Myla
v
aram,
India
Article
Inf
o
Article
history:
Recei
v
ed
Jan
4,
2025
Re
vised
Aug
25,
2025
Accepted
Sep
10,
2025
K
eyw
ords:
Graph
eigen
v
alues
Graph
signal
processing
Speech
analysis
Structural
processing
Speech
trajectory
ABSTRA
CT
Speech
signal
analysis
aims
to
identify
patterns
within
data
to
de
v
elop
ef
fec-
ti
v
e
recognition
algorithms.
This
process
primarily
utilizes
feature
e
xtraction
techniques
such
as
linear
predicti
v
e
coding
(LPC),
linear
predicti
v
e
cepstral
co-
ef
cients
(LPCCs),
and
Mel-frequenc
y
cepstral
coef
cients
(MFCCs).
These
features
are
crucial
for
constructing
recognition
algorithms
that
le
v
erage
both
statistical
and
deep
learning
methods.
While
deep
learning
models
require
e
x-
tensi
v
e
datasets,
the
y
often
pro
v
e
unsuitable
for
lo
w-resource
languages.
The
Hidden
Mark
o
v
model
(HMM)
is
the
most
widely
adopt
ed
statistical
frame
w
ork
in
speech
processing.
Ho
we
v
er
,
HMMs
are
characterized
by
state-dependent
models,
where
each
state
interacts
only
with
its
neighboring
states.
This
limita-
tion
restricts
HMMs
from
capturing
long-term
signal
properties,
highlighting
the
need
for
addressing
these
constraints
at
the
fe
ature
e
xtraction
stage.
Most
feature
e
xtraction
methods
rely
on
short-term
signal
processing,
which
further
limits
the
comprehension
of
spe
ech
utterances.
T
o
o
v
ercome
the
se
limitations,
alter
-
nati
v
e
methods
are
necessary
to
capture
more
comprehensi
v
e
patterns.
This
pa-
per
presents
a
graph-based
approach
for
analyzing
speech
trajectories
and
their
temporal
properties,
which
are
subsequently
v
alidated
using
HMMs
in
speech
recognition
tasks.
Graph-based
representations
on
a
lo
w-resource
T
elugu
dataset
impro
v
e
recognition
accurac
y
by
13%
while
reducing
processing
time
compared
to
traditional
LPC.
This
is
an
open
access
article
under
the
CC
BY
-SA
license
.
Corresponding
A
uthor:
P
arabattina
Bhag
ath
Department
of
Articial
Intelligence
and
Data
Science,
Lakireddy
Bali
Reddy
Colle
ge
of
Engineering
Myla
v
aram,
Andhra
Pradesh,
India
Email:
bhag
ath.2014@alumni.iitg.ac.in
1.
INTR
ODUCTION
Speech
signal
analysis
is
a
e
ld
that
ai
ms
to
u
nder
stand
speech
signals
and
pro
vi
de
v
aluable
i
nsights
for
resea
rchers
w
orking
on
de
v
eloping
recogniti
on
algorithms
and
feature
e
xtraction
techniques.
This
area
has
progressed
by
taking
inspiration
from
v
arious
domains
such
as
signal
processing,
cogniti
v
e
science
[1],
linguistics
[2],
[3],
and
deep
learning
[4].
R
ecently
,
graph
signal
processing
(GSP)
[5]
has
emer
ged
as
a
ne
w
area
of
research
that
focuses
on
graph-based
analysis
for
understanding
patterns
of
underlying
data.
It
pro
vides
a
wide
v
ariety
of
tools
and
techniques
for
computer
netw
orks,
speech
processing,
and
data
science.
Graphs
are
an
ef
cient
w
ay
to
represent
multi-dimensional
properties
of
data,
which
in
turn
helps
to
e
xtract
appropriate
features.
T
raditional
automatic
speech
recognit
ion
(ASR)
systems
use
v
arious
techniques
such
as
linear
predicti
v
e
coef
cients
(LPCs)
and
Mel-frequenc
y
cepst
ral
coef
cients
(MFCCs)
to
e
xtract
features
from
J
ournal
homepage:
http://journal.uad.ac.id/inde
x.php/TELK
OMNIKA
Evaluation Warning : The document was created with Spire.PDF for Python.
TELK
OMNIKA
T
elecommun
Comput
El
Control
❒
1627
ra
w
audio
signals
and
transform
them
into
representati
v
e
feature
v
ectors
that
capture
essential
characteristics
of
speech
[6].
One
promising
a
v
enue
for
f
eature
e
xtraction
is
the
use
of
graph
eigen
v
alues,
which
pro
vide
insights
into
the
comple
x
temporal
patterns
present
in
s
p
ok
en
utterances.
By
representing
speech
signals
as
graphs
and
utilizing
the
eigen
v
alues
of
these
graphs,
we
can
encode
v
aluable
information
about
the
underlying
structure
and
dynamics
of
speech
patterns.
In
this
paper
,
we
in
v
estig
ate
the
graph-based
representation
of
speech
signals
to
gi
v
e
crucial
features
for
speech
recognition
tasks.
The
paper
is
or
g
anized
as:
in
section
2,
we
present
the
historical
de
v
elopments
of
GSP
and
related
research
for
time
series
analysis.
The
section
3
discusses
the
proposed
methodology
.
In
section
4,
we
discuss
the
en
vironmental
setup
used
for
de
v
eloping
the
proposed
approach.
Finally
,
section
4
elaborates
on
the
ndings
whereas
the
paper
is
concluded
in
section
5.
2.
LITERA
TURE
REVIEW
The
proposed
approach
is
moti
v
ated
by
the
f
act
that
a
signal
can
be
represented
as
a
graph.
GSP
has
e
v
olv
ed
as
a
research
domain
studying
v
arious
areas
such
as
signal
processing,
communication
netw
orks,
and
big
data
[7].
GSP
adapted
the
methods
from
graph
theory
and
digital
signal
processing
and
it
is
pro
viding
tools
and
techniques
to
solv
e
dif
ferent
problems
in
the
afore-mentioned
areas.
Graphs
are
generally
used
to
represent
relationships
between
components
in
a
ph
ysical
netw
ork
or
data
elements
[8].
The
usage
of
graphs
to
represent
the
relationships
is
tw
o-fold:
causality
and
in
v
ariance.
The
causality
property
helps
to
retain
the
relationship
between
the
data
components
whereas
the
in
v
ariance
pro
vides
a
w
ay
to
hold
the
relationships
in
a
v
arying
en
vironment.
These
properties
are
further
useful
in
speech
signal
analysis
for
tw
o
reasons.
The
y
are
useful
properties
to
capture
the
temporal
dynamics
and
represent
the
relationships
between
v
arious
acoustic
e
v
ents
[9].
Researchers
are
trying
to
use
graph-based
methods
to
study
speech
signals
a
nd
use
the
concepts
such
as
temporal
GSP
,
graph
fourier
transform
(GFT)
[10],
and
speech
modeling
techniques.
A
graph
is
a
component
that
has
a
set
of
v
ertices,
edges,
and
weights.
It
can
be
represented
as
G
=
(
V
,
E
,
W
)
,
where
V
is
a
set
of
v
ertices,
and
E
represents
the
edges
between
the
v
ertices.
The
weights
associated
with
each
edge
are
gi
v
en
by
W
.
F
or
a
signal
s
represented
by
a
random
v
ector
gi
v
en
by
S
=
x
1
,
x
2
,
x
3
,
...,
x
n
with
n
samples,
there
are
tw
o
approaches
for
constructing
the
graphs
for
studying
the
underl
ying
patterns.
The
rst
one
is
a
data-dri
v
en
approach
in
which
the
signal
is
modelled
by
understanding
the
statistical
distrib
ution.
In
this
method,
statistical
components
such
as
mean
and
co
v
ariance
are
commonly
used.
The
second
method
for
graph
construction
is
an
intuiti
v
e
method
in
which
a
set
of
heuristic
rules
are
dened
for
choosing
the
components
of
a
graph
[11].
A
causal
graph
w
as
proposed
by
Mei
and
Moura
[12]
t
o
estimate
the
time
series.
F
or
a
discrete-time
signal
S
[
n
]
,
a
causal
graph
is
gi
v
en
by
(1).
Here
y
[
n
]
represents
the
statistical
noise,
P
i
is
the
lter
on
matrix
A
,
and
N
is
the
order
of
the
model.
S
[
n
]
=
y
[
n
]
+
N
X
i
=1
P
i
(
A
)
S
[
n
−
i
]
(1)
GFT
is
a
technique
that
helps
in
unders
tanding
the
frequenc
y
of
graphs.
GFT
for
a
matrix
A
is
dened
based
on
laplacian
L
which
is
gi
v
en
by
(2).
Here
U
is
the
matrix
for
which
the
columns
are
eigen
v
ectors
of
A
and
V
is
the
diagonal
matrix
of
eigen
v
alues
[13].
L
=
U
V
U
T
(2)
Stationarity
and
harmonics
of
a
signal
are
tw
o
important
properties
to
nd
out
the
relationship
between
a
time
series
and
its
associated
graph.
Understanding
these
components
together
pro
vides
necessary
clues
to
the
temporal
and
harmonic
properties
of
a
signal.
Loukas
and
F
oucard
[14]
dened
joint
fourier
transform
(JFT)
that
uses
the
principles
of
discrete
fourier
transform
(DFT)
and
GFT
.
JFT
gi
v
es
the
frequenc
y
content
of
a
signal
S
[
n
]
w
.r
.t
both
time
and
graph
domains.
The
relationship
of
a
graph
represented
as
G
=
V
,
E
,
W
and
a
signal
S
[
n
]
JFT
is
gi
v
en
by
(3)
where
V
G
is
the
N
×
N
left
eigen
v
ector
matrix
and
X
V
T
T
is
the
result
of
D
F
T
{
X
}
.
GFT
pro
vides
a
w
ay
to
estimate
the
v
arying
nature
of
a
signal
o
v
er
the
graph
and
is
a
useful
w
ay
of
measuring
the
abrupt
changes
in
a
temporal
graph.
J
F
T
(
S
,
G
)
=
V
G
S
V
T
T
(3)
The
study
of
graph
structures
in
the
conte
xt
of
speech
signal
analysis
is
pro
v
en
to
be
ef
fecti
v
e
and
it
w
as
used
for
identifying
phoneme
boundaries.
A
method
to
se
gment
the
phonetic
units
using
the
graph
Analyzing
tempor
al
pr
operties
of
speec
h
tr
ajectory
using
gr
aph
structur
es
towar
ds
...
(P
ar
abattina
Bha
gath)
Evaluation Warning : The document was created with Spire.PDF for Python.
1628
❒
ISSN:
1693-6930
eigen
v
alues
is
proposed
by
Bhag
ath
and
Das
[15]
and
it
w
as
pro
v
ed
to
capture
the
temporal
properties
of
speech
signals.
Gaussian
graph
signals
use
the
cumulati
v
e
sum
(CUSUM)
procedure
to
acquire
kno
wledge
of
an
e
v
ent
by
estimating
i
ts
prior
probability
and
posterior
probability
.
Generalized
lik
elihood
ratio
(GLR)
is
used
for
estimating
post-changes
[16],
[17].
The
domain
of
speech
signal
analysis
also
adapts
to
graph
structures
to
understand
and
model
the
speech
data.
The
methods
ha
v
e
been
sho
wn
to
be
ef
fecti
v
e
in
noisy
en
vironments
as
well.
An
i
mpro
v
ed
W
iener
lter
approach
is
proposed
by
W
ang
et
al.
[18]
for
noisy
speech
signal
s.
In
addition
to
using
the
graph
structures
for
representation,
the
y
ha
v
e
been
used
for
modeling
using
deep
neural
netw
ork
architectures.
Shirian
and
Guha
[19]
proposed
a
graph-based
method
to
recognize
emotions
using
graph
neural
netw
orks
(GNNs).
The
study
of
graph
structures
in
the
conte
xt
of
speech
signal
analysis
has
been
pro
v
en
to
be
ef
fecti
v
e,
and
i
t
w
as
used
for
identifying
phoneme
boundaries.
A
method
to
se
gment
the
phonetic
units
using
the
graph
eigen
v
alues
is
proposed
by
Bhag
ath
and
Das
[20],
and
i
t
w
as
pro
v
ed
to
capture
the
temporal
properties
of
speech
signals.
The
ASR
problem
for
under
-resourced
languages
is
challenging,
especially
for
Indian
languages.
W
e
can
nd
a
lot
of
researchers
w
orking
in
these
languages,
de
v
eloping
both
the
corpus
and
methods.
In
the
literature,
there
ha
v
e
been
numerous
research
w
orks
found
to
address
the
speech
recognition
problem
for
lo
w-resourced
languages
[21].
Ho
we
v
er
,
these
de
v
elopments
are
limited
to
a
fe
w
languages
and
domains.
T
elugu
is
a
language
for
which
there
a
minimal
contrib
ution
in
the
eld
of
speech
recognition.
T
elugu
has
multiple
dialects
for
dif
ferent
geographical
areas,
such
as
Andhra,
T
elang
ana,
and
Coas
tal.
De
v
eloping
dialect-specic
spok
en
models
is
a
challenging
task
because
of
the
a
v
ailability
of
the
datasets.
Y
ada
v
alli
et
al.
[22]
proposed
a
unied
model
to
recognize
w
ords
irrespecti
v
e
of
the
dialect.
In
the
research
w
ork,
T
ransformer
-based
language
models
were
used
with
block
size
16.
The
proposed
approach
is
discussed
in
the
ne
xt
section.
In
the
present
paper
,
we
de
v
elop
a
lo
w-cost
feature
e
xtraction
method
that
uses
graph
structures.
3.
PR
OPOSED
FRAMEW
ORK
FOR
SPEECH
SIGN
AL
AN
AL
YSIS
The
proposed
approach
consists
of
four
major
components:
identifying
structural
components,
graph
construction,
eigen
v
alue
computation,
and
modeling
as
outlined
in
Figure
1.
At
rst,
a
speech
signal
is
nor
-
malized
to
ensure
that
the
amplitudes
are
uniform
across
the
dataset.
Each
speech
se
gment
is
further
analyzed
to
e
xtract
the
primary
components
of
a
graph’
s
peaks
and
v
alle
ys.
The
amplitude
changes
of
a
spok
en
utter
-
ance
form
a
structural
representation
in
a
w
a
v
eform.
This
essentially
helps
to
understand
the
properties
of
the
signal’
s
shape
and
can
be
used
as
features
for
modeling.
Each
peak
and
v
alle
y
is
associated
with
the
location,
which
tells
the
sequence
of
ho
w
the
y
are
presented
in
a
w
a
v
eform.
A
series
of
graphs
will
be
constructed
for
a
spok
en
utterance.
The
construction
of
the
graph
is
e
xplained
in
the
ne
xt
subsection.
F
or
e
xample,
p
0
refers
to
the
peak
presented
at
location
0
in
the
respecti
v
e
w
a
v
eform.
This
is
true
for
the
v
alle
y
also.
The
presence
of
a
peak
and
a
v
alle
y
in
a
location
is
mutually
e
xclusi
v
e,
which
means
a
peak
and
a
v
alle
y
cannot
be
in
the
same
location.
W
ith
these
components,
a
graph
is
constructed
for
each
speec
h
se
gment,
usually
ha
ving
a
10-ms
to
20-ms
duration.
T
o
understand
the
procedure,
we
rst
dene
the
primiti
v
e
elements
of
a
graph:
peaks
and
v
alle
ys.
A
peak
in
a
speech
se
gment
is
a
sample
s
p
that
satises
(4),
where
p
is
an
inte
ger
and
the
position
of
a
peak.
A
v
alle
y
in
a
speech
se
gment
is
a
sample
s
v
that
satises
(5),
where
v
is
an
inte
ger
and
the
position
of
a
v
alle
y
.
A
graph
designated
as
G
is
a
set
of
v
ertices
V
and
edges
E
.
Each
v
erte
x
in
a
graph
can
be
a
peak
or
a
v
alle
y
.
The
edges
between
the
v
ertices
are
dened
as
a
link
between
a
peak
a
n
d
a
v
alle
y
,
designated
as
p
i
and
v
i
.
The
indi
vidual
steps
in
the
proposed
approach
are
e
xplained
in
the
subsequent
subsections.
s
p
−
1
<
s
p
<
s
p
+1
(4)
s
v
−
1
>
s
v
>
s
v
+1
(5)
Figure
1.
Proposed
frame
w
ork
using
graph
eigen
v
alues
TELK
OMNIKA
T
elecommun
Comput
El
Control,
V
ol.
23,
No.
6,
December
2025:
1626–1634
Evaluation Warning : The document was created with Spire.PDF for Python.
TELK
OMNIKA
T
elecommun
Comput
El
Control
❒
1629
3.1.
Graph
construction
At
rst,
a
set
of
peaks
and
v
alle
ys
are
found
using
the
primiti
v
es
gi
v
en
by
(4)
and
(5).
An
e
xample
graph
structure
constructed
this
w
ay
is
sho
wn
in
Figure
2.
In
this
structure,
a
link
between
a
peak
and
a
v
alle
y
is
created
if
the
y
are
both
positioned
in
subsequent
locations.
Based
on
the
starting
component
of
a
graph,
there
are
three
possible
structures
that
contain
a
v
aried
number
of
peaks
and
v
alle
ys.
There
are
three
types
of
edges
possible
in
a
graph
as
follo
ws:
Figure
2.
Proposed
graph
structure
-
An
edge
is
dened
between
tw
o
consecuti
v
e
peaks,
dened
as
(
p
i
,
p
i
+1
)
-
An
edge
is
dened
between
tw
o
consecuti
v
e
v
alle
ys
dened
as
(
v
i
,
v
i
+1
)
-
An
edge
is
dened
between
a
peak
and
a
v
alle
y
positioned
in
consecuti
v
e
locations
dened
as
(
p
i
,
v
i
)
Once
the
graphs
are
constructed
for
a
speech
signal,
graph
eigen
v
alues
are
computed
for
each
graph.
The
procedure
to
compute
graph
eigen
v
alues
is
discussed
in
the
ne
xt
subsection.
3.2.
Graph
eigen
v
alue
computation
The
proce
d
ur
e
of
eigen
v
alue
computation
in
v
olv
es
three
important
matrices
such
as
the
adjacenc
y
matrix,
the
diagonal
matrix,
and
the
Laplacian
matrix.
F
or
a
graph
G
,
the
adjacenc
y
matrix
designated
as
A
G
is
dened
as
the
matrix
consisting
of
either
’0’
or
’1’.
A
G
contains
1
in
A
G
(
i,
j
)
if
node
i
and
node
j
are
adja
cent.
D
G
is
the
diagonal
matrix,
and
it
contains
t
he
v
erte
x
de
grees
of
graph
G
.
The
Laplacian
matrix
L
G
is
gi
v
en
by
(6)
and
(7).
Finally
,
the
eigen
v
alues
of
G
can
be
computed
by
solving
(8).
These
steps
are
outlined
in
Algorithm
1.
The
solution
to
equations
gi
v
e
the
eigen
v
alues
that
represent
a
speech
signal.
F
or
a
speech
signal
S
[
n
]
with
samples
n
and
windo
w
length
k
,
n
k
number
of
frames
can
be
formed
which
is
designated
as
n
f
r
ames
in
the
algorithm.
This
f
actor
decides
the
number
of
graphs
|
G
|
formed
from
a
source
signal.
Subsequently
,
each
graph
is
used
to
compute
an
eigen
v
alue
that
represents
a
feature
for
each
speech
se
gment.
The
graph
eigen
v
alues
are
computed
for
all
the
corresponding
graphs
using
lines
8
to
10
in
the
algorithm.
Once
this
step
is
completed,
we
need
to
model
each
spok
en
w
ord
with
a
modeling
technique.
A
sample
set
of
graph
eigen
v
alues
computed
for
digits
zero
through
eight
are
sho
wn
in
Figure
3.
Algorithm
1
Graph
eigen
v
alue
computation
1:
pr
ocedur
e
G
E
V
(
G
k
)
▷
G
k
is
a
series
of
graphs
2:
i
←
0
3:
while
i
̸
=
n
f
r
ames
do
▷
n
f
r
ames
is
the
number
of
se
gments
4:
G
i
←
Construct
graph(
s
i
)
▷
Map
each
s
i
to
g
i
5:
end
while
6:
i
←
0
7:
while
i
̸
=
|
G
|
do
8:
L
(
g
i
)
←
D
(
g
i
)
−
A
(
g
i
)
▷
Laplacian
of
graph
g
i
9:
10:
g
e
i
←
U
T
(
g
i
)
A
(
g
i
)
U
(
g
i
)
U
T
(
g
i
)
U
(
g
i
)
11:
end
while
12:
r
etur
n
g
e
i
▷
Set
of
graph
eigen
v
alues
for
S[n]
13:
end
pr
ocedur
e
F
or
v
erifying
the
proposed
features,
we
used
a
Hidden
Mark
o
v
model
(HMM)
for
modeling
the
spok
en
data,
which
is
discussed
in
the
ne
xt
subsection.
L
G
=
D
G
−
A
G
(6)
Analyzing
tempor
al
pr
operties
of
speec
h
tr
ajectory
using
gr
aph
structur
es
towar
ds
...
(P
ar
abattina
Bha
gath)
Evaluation Warning : The document was created with Spire.PDF for Python.
1630
❒
ISSN:
1693-6930
L
G
(
i,
j
)
=
deg
r
ee
i
,
if
i=j
−
A
G
(
i,
j
)
,
if
i
̸
=
j
(7)
|
L
G
−
λ
(
G
)
|
=
0
(8)
Figure
3.
Graph
eigen
v
alues
for
the
digits
from
’zero’
through
’nine’
(one
speak
er)
3.3.
Hidden
Mark
o
v
model
modeling
The
HMM
are
nite
state
machines
in
which
a
process
is
represented
by
a
set
of
components.
The
elements
of
HMM
are
a
set
of
states,
a
transition
matrix,
and
a
special
start
state.
HMM
is
often
called
a
stochastic
process,
which
means
an
unkno
wn
process
can
be
observ
ed
through
a
kno
wn
process
[23].
This
approach
is
suitable
for
solving
the
speech
recognition
problem
for
one
rea
son.
In
speech
recognition,
the
spok
en
utterance
is
kno
wn,
and
the
v
ocal
tract
system
that
utters
the
m
essage
is
hidden.
By
using
the
HMM,
we
are
supposed
to
estimate
the
properties
of
the
v
ocal
tract
system.
The
HMM
solv
es
a
speech
recognition
problem
in
three
stages.
The
present
approach
uses
a
4-state
HMM
model
in
the
recognition
module
as
sho
wn
in
Figure
4.
Figure
4.
4-state
HMM
model
The
rst
step
focus
es
on
kno
wing
the
parameters
of
the
model
by
understanding
the
patterns
hidden
in
the
observ
ation
sequence,
often
called
as
features.
In
the
second
stage,
the
states
of
the
HMM
are
decoded
to
nd
out
the
optimal
state
sequence
that
produces
the
observ
ation
sequence.
Finally
,
the
model
is
e
v
aluated
TELK
OMNIKA
T
elecommun
Comput
El
Control,
V
ol.
23,
No.
6,
December
2025:
1626–1634
Evaluation Warning : The document was created with Spire.PDF for Python.
TELK
OMNIKA
T
elecommun
Comput
El
Control
❒
1631
by
estimating
ho
w
well
the
model
can
generate
the
gi
v
en
observ
ation
sequence.
This
is
usually
computed
by
maximum
lik
elihood
estimation.
In
general,
the
number
of
states
in
an
HMM
is
decided
based
on
the
number
of
phonetic
units
in
the
tar
geted
w
ords.
The
proposed
approach
is
tested
on
a
T
elugu
Digit
dataset,
and
the
number
of
states
in
the
HMM
is
appropriate.
4.
EXPERIMENT
AL
SETUP
AND
RESUL
TS
In
this
section,
we
detail
the
en
vironment
used
for
module
de
v
elopment
and
outline
our
obser
v
ations
from
the
e
xperiments.
The
complete
dataset
comprises
17,600
speech
samples
of
ten
distinct
w
ords,
spok
en
by
88
dif
ferent
speak
ers.
These
w
ords
represent
T
elugu
digits,
ranging
from
zero
(0)
to
nine
(9),
with
each
digit
being
articulated
20
times.
The
dataset
is
a
v
ailable
at
the
follo
wing
link:
https://tinyurl.com/
57rsz5xs
.
In
our
study
,
70%
of
the
dataset
w
as
designated
for
training,
while
the
remaining
30%
w
as
utilized
for
model
e
v
aluation.
Each
speech
utterance
underwent
pre-processing,
which
in
v
olv
ed
remo
ving
the
direct
current
(DC)
component
and
normalizing
the
audio
to
a
range
of
{
-1,
+1
}
.
This
normalization
ensures
consistent
amplitude
v
alues
across
all
speak
ers.
The
details
of
the
dataset
are
summarized
in
T
able
1.
W
e
emplo
yed
se
v
eral
libraries,
including
scientic
Python
(SciPy),
Python
graph
signal
processing
(PyGSP)
[24],
and
Netw
orkX
[25],
t
o
enhance
our
w
ork
on
HMM
modeling,
the
implementation
of
graph
structures,
and
the
computation
of
graph
eigen
v
alues.
SciPy
pro
vides
a
range
of
functions
for
processing
matrix
data
structures
with
its
b
uilt-in
capabili
ties.
PyGSP
is
a
l
ibrary
that
e
n
a
bles
the
appli
cation
of
si
gn
a
l
processing
c
o
nc
epts
to
graphs,
whereas
Netw
orkX
f
acilitates
the
representation
of
graph
structures.
T
able
1.
Summary
of
the
dataset
S.
no
Attrib
utes
V
alues
1
T
otal
#
of
distinct
w
ords
10
2
T
otal
#
of
speak
ers
88
3
T
otal
#
of
sessions
80
4
T
otal
#
of
les
17,600
5
T
otal
duration
of
all
the
les
587
mins
(9.8
hours)
The
recognition
rates
of
3
dif
ferent
features
b
uilt
with
HMM
are
sho
wn
in
T
able
2.
Our
ndings
re
v
eal
that
the
recognition
rate
for
the
linear
predict
i
v
e
coding
(LPC)+HMM
model
ranges
from
24%
to
76%,
linear
predicti
v
e
cepstral
coef
cients
(LPCC)+HMM
model
gi
v
es
recognition
rates
v
arying
from
22%
to
82%
whereas
the
recognition
rates
for
the
graph
EigenV
alues
(GEV)+HMM
model
span
from
37%
to
100%.
These
e
xperiments
illustrate
that
the
proposed
approach
achie
v
es
a
superior
recognition
rate
when
utilizing
structural
features.
In
addition
to
the
better
recognition
rates,
there
is
a
reduction
in
the
computational
cost
since
a
single
eigen
v
alue
needs
to
be
computed
for
each
speech
se
gment.
In
contrast,
the
con
v
entional
LPC
and
LPCC
methods
compute
13
coef
cients
for
each
speech
se
gment,
resulting
in
a
higher
dimensionality
.
This
ne
w
method
reduces
the
length
of
the
feature
v
ector
by
a
f
actor
of
12,
of
fering
signicant
computational
adv
antages
during
the
modeling
phase.
The
results
are
presented
in
Figure
5.
T
able
2.
Recognition
rates
(%)
obtained
for
dif
ferent
speak
ers
S.
no
LPC
LPCC
GEV
S.
no
LPC
LPCC
GEV
S.
no
LPC
LPCC
GEV
1
44
64
46
13
41
50
62
25
48
48
54
2
52
52.66
46
14
54
52
66
26
62
68
70
3
54
60
54
15
37
36
68
27
56
48
52
4
52
22
48
16
54
36
88
28
46
58
56
5
46
82
52
17
52
30
88
29
34
34
37
6
54
82
58
18
24
24
96
30
44
50
46
7
60
62
52
19
76
48
100
31
52
48
66
8
54
30
54
20
44
56
46
32
54
72
88
9
38
48
54
21
60
48
90
33
48
28
40
10
58
58
54
22
52
48
54
34
56
18
96
11
56
46
56
23
54
44
100
35
36
38
44
12
52
40
58
24
60
36
90
36
46
40
66
Analyzing
tempor
al
pr
operties
of
speec
h
tr
ajectory
using
gr
aph
structur
es
towar
ds
...
(P
ar
abattina
Bha
gath)
Evaluation Warning : The document was created with Spire.PDF for Python.
1632
❒
ISSN:
1693-6930
Figure
5.
Comparison
between
LPC
and
GEV
features
for
40
speak
ers
5.
CONCLUSION
This
paper
introduces
a
graph-based
method
for
the
analysi
s
of
speech
signals.
The
methodology
em-
plo
ys
graph
structures
as
feature
representations
within
the
temporal
domain,
ef
fecti
v
ely
capturing
the
inherent
patterns
of
a
speech
signal
through
this
graphical
representation.
The
proposed
feature
e
xtraction
method
w
as
tested
in
recognition
tasks
using
HMM
to
assess
its
ef
cac
y
.
The
performance
of
the
model
s
de
v
eloped
for
the
dataset
w
as
compared
to
traditional
approaches
using
LPC+HMM
and
LPCC+HMM.
The
results
indi-
cate
a
13%
impro
v
ement
in
recognition
rates
o
v
er
LPC
and
LPCC
features,
along
with
a
12-fold
reduction
in
computational
costs.
Future
w
ork
will
focus
on
analyzing
the
scalability
of
the
proposed
features
with
lar
ger
datasets.
FUNDING
INFORMA
TION
Authors
state
no
funding
in
v
olv
ed.
A
UTHOR
CONTRIB
UTIONS
ST
A
TEMENT
This
journal
uses
the
C
on
t
rib
utor
Roles
T
axonomy
(CRediT)
to
recognize
indi
vidual
author
contrib
u-
tions,
reduce
authorship
disputes,
and
f
acilitate
collaboration.
Name
of
A
uthor
C
M
So
V
a
F
o
I
R
D
O
E
V
i
Su
P
Fu
P
arabattina
Bhag
ath
✓
✓
✓
✓
✓
✓
✓
Malempati
Shanmukha
✓
✓
✓
✓
✓
✓
Gnana
Nag
asri
Puthi
✓
✓
✓
✓
✓
C
:
C
onceptualization
I
:
I
n
v
estig
ation
V
i
:
V
i
sualization
M
:
M
ethodology
R
:
R
esources
Su
:
Su
pervision
So
:
So
ftw
are
D
:
D
ata
Curation
P
:
P
roject
Administration
V
a
:
V
a
lidation
O
:
Writing
-
O
riginal
Draft
Fu
:
Fu
nding
Acquisition
F
o
:
F
o
rmal
Analysis
E
:
Writing
-
Re
vie
w
&
E
diting
CONFLICT
OF
INTEREST
ST
A
TEMENT
Authors
state
no
conict
of
interest.
D
A
T
A
A
V
AILABILITY
The
data
that
support
the
ndings
of
this
study
are
a
v
ailable
on
request
from
the
corr
esponding
author
,
[initials:
PB].
The
data,
which
contain
information
that
could
compromise
the
pri
v
ac
y
of
research
participants,
TELK
OMNIKA
T
elecommun
Comput
El
Control,
V
ol.
23,
No.
6,
December
2025:
1626–1634
Evaluation Warning : The document was created with Spire.PDF for Python.
TELK
OMNIKA
T
elecommun
Comput
El
Control
❒
1633
are
not
publicly
a
v
ailable
due
to
certain
restrictions.
REFERENCES
[1]
R.
V
oleti,
J.
M.
Liss,
and
V
.
Berisha,
“
A
Re
vie
w
of
Automated
Speech
and
Language
Features
for
Assessment
of
Cogniti
v
e
and
Thought
Disorders,
”
IEEE
J
ournal
of
Selected
T
opics
in
Signal
Pr
ocessing
,
v
ol.
14,
no.
2,
pp.
282–298,
Feb
.
2020,
doi:
10.1109/JSTSP
.2019.2952087.
[2]
L.
Xue,
F
.
K.
Soong,
S.
Zhang,
and
L.
Xie,
“P
araTTS:
Learning
Linguistic
and
Prosodic
Cross-Sentence
Information
in
P
aragraph-Based
TTS,
”
IEEE/A
CM
T
r
ansactions
on
A
udio,
Speec
h,
and
Langua
g
e
Pr
ocessing
,
v
ol.
30,
pp.
2854–2864,
2022,
doi:
10.1109/T
ASLP
.2022.3202126.
[3]
Z.
Zhang,
E.
Coutinho,
J.
Deng,
and
B.
Schuller
,
“Distrib
uting
recognition
in
computational
paralinguistics,
”
IEEE
T
r
ansactions
on
Af
fective
Computing
,
v
ol.
5,
no.
4,
pp.
406–417,
Oct.
2014,
doi:
10.1109/T
AFFC.2014.2359655.
[4]
D.
F
.
O.
Onah
and
A.
Ibrahim,
“Ev
aluating
Speech
Emotion
Recognition
through
the
lens
of
CNN;
LSTM
Deep
Learning
Models,
”
in
2023
IEEE
International
Confer
ence
on
Big
Data
(BigData),
IEEE
,
Dec.
2023,
pp.
6245–6247,
doi:
10.1109/Big-
Data59044.2023.10386881.
[5]
J.
W
ang
and
X.
P
an,
“The
signal
feature
e
xtraction
of
graph
F
ourier
transform
on
the
constructed
graph,
”
in
Pr
oceedings
-
2021
6th
International
Confer
ence
on
Communication,
Ima
g
e
and
Signal
Pr
ocessings,
CCISP
2021,
IEEE
,
No
v
.
2021,
pp.
379–383,
doi:
10.1109/CCISP52774.2021.9639270.
[6]
A.
Cho
wdhury
and
A.
Ross,
“Fusing
MFCC
and
LPC
Features
Using
1D
T
riplet
CNN
for
Speak
er
Recognition
in
Se
v
erely
De
graded
Audio
Signals,
”
IEEE
T
r
ansactions
on
Information
F
or
ens
ics
and
Security
,
v
ol.
15,
pp.
1616–1629,
2020,
doi:
10.1109/TIFS.2019.2941773.
[7]
A.
Sandryhaila
and
J.
M.
F
.
Moura,
“Big
data
analysis
with
signal
processing
on
graphs:
Representation
and
processing
of
massi
v
e
data
sets
with
irre
gular
structure,
”
IEEE
Signal
Pr
ocessing
Ma
gazine
,
v
ol.
31,
no.
5,
pp.
80–90,
Sep.
2014,
doi:
10.1109/MSP
.2014.2329213.
[8]
E.
Ceci
and
S.
Barbarossa,
“Graph
signal
processing
in
the
presence
of
topology
uncertainties,
”
IEEE
T
r
ansactions
on
Signal
Pr
ocessing
,
v
ol.
68,
pp.
1558–1573,
2020,
doi:
10.1109/TSP
.2020.2976583.
[9]
X.
He,
Y
.
Xie,
S.-M.
W
u,
and
F
.-C.
Lin,
“Sequential
Graph
Scanning
Statistic
for
Change-point
Detection,
”
in
2018
52nd
Asilomar
Confer
ence
on
Signals,
Systems,
and
Computer
s,
IEEE
,
Oct.
2018,
pp.
1317–1321,
doi:
10.1109/A
CSSC.2018.8645505.
[10]
A.
G.
V
eera
v
all
i,
W
.
D.
P
an,
R.
Adhami
and
P
.
G.
Cox,
”A
tutorial
on
using
hidden
Mark
o
v
models
for
phoneme
recognition,
”
Pr
o-
ceedings
of
the
Thirty-Se
venth
Southeastern
Symposium
on
System
Theory
,
2005,
pp.
154-157,
doi:
10.1109/SSST
.2005.1460896.
[11]
C.
Zhang,
D.
Florencio,
and
P
.
A.
Chou,
“Graph
Signal
Processing
–
A
Probabilistic
Frame
w
ork,
”
Micr
osoft
Resear
c
h
T
ec
hnical
Report
,
p.
MSR-TR-2015-31,
2015.
[12]
J.
Mei
and
J.
M.
F
.
Moura,
”Signal
Processing
on
Graphs:
Causal
Modeling
of
Unstructured
Data,
”
in
IEEE
T
r
ansactions
on
Signal
Pr
ocessing
,
v
ol.
65,
no.
8,
pp.
2077-2092,
2017.
[13]
N.
Perraudin,
A.
Loukas,
F
.
Grassi,
and
P
.
V
ander
ghe
ynst,
“T
o
w
ards
stationary
time-v
erte
x
signal
processing,
”
ICASSP
,
IEEE
International
Confer
ence
on
Acoustics,
Speec
h
and
Signal
Pr
ocessing
-
Pr
oceedings
,
pp.
3914–3918,
2017,
doi:
10.1109/ICASSP
.2017.7952890.
[14]
A.
Loukas
and
D.
F
oucard,
“Frequenc
y
analysis
of
time-v
arying
graph
signals,
”
in
2016
IEEE
Global
Confer
ence
on
Signal
and
Information
Pr
ocessing
(GlobalSIP),
IEEE
,
Dec.
2016,
pp.
346–350,
doi:
10.1109/GlobalSIP
.2016.7905861.
[15]
P
.
Bhag
ath
and
P
.
K.
Das,
“Phoneme
Boundary
Analysis
Using
Graphs,
”
in
IEEE
Re
gion
10
Annual
International
Confer
ence
,
Pr
oceedings/TENCON,
IEEE
,
Oct.
2019,
pp.
1764–1768,
doi:
10.1109/TENCON.2019.8929673.
[16]
A.
Buciulea
and
A.
G.
Marques,
“Graph
Learning
from
Gaussian
and
S
tationary
Graph
Signals,
”
in
ICASSP
,
IEEE
In-
ternational
Confer
ence
on
Acoustics,
Speec
h
and
Signal
Pr
ocessing
-
Pr
oceedings,
IEEE,
J
un.
2023
,
pp.
1–5,
doi:
10.1109/ICASSP49357.2023.10096413.
[17]
Y
.
Chen,
X.
Mao,
D.
Ling,
and
Y
.
Gu,
“Change-Poi
nt
Detection
of
Gaussian
Graph
Signals
with
P
artial
Information,
”
in
ICASSP
,
IEEE
International
Confer
ence
on
Acoustics,
Speec
h
and
Signal
Pr
ocessing
-
Pr
oceedings,
IEEE
,
Apr
.
2018,
pp.
3934–3938,
doi:
10.1109/ICASSP
.2018.8461397.
[18]
T
.
W
ang,
H.
Guo,
X.
Y
an,
and
Z.
Y
ang,
“Speech
signal
processing
on
graphs:
The
graph
frequenc
y
analysis
and
an
impro
v
ed
graph
W
iener
ltering
method,
”
Speec
h
Communication
,
v
ol.
127,
pp.
82–91,
Mar
.
2021,
doi:
10.1016/j.specom.2020.12.010.
[19]
A.
Shirian
and
T
.
Guha,
“Compact
graph
architecture
for
speech
emotion
recognition,
”
in
ICASSP
,
IEEE
Interna-
tional
Confer
ence
on
Acoustics,
Speec
h
and
Signal
Pr
ocessing
-
Pr
oceedings,
IEEE
,
Jun.
2021,
pp.
6284–6288,
doi:
10.1109/ICASSP39728.2021.9413876.
[20]
P
.
Bhag
ath
and
P
.
K.
Das,
“Graph
Eigen
v
alue
based
Structural
Method
to
w
ards
Phonetic
Boundary
Detection,
”
in
I
EEE
Re
gion
10
Annual
International
Confer
ence
,
Pr
oceedings/TENCON,
IEEE
,
Dec.
2021,
pp.
591–596,
doi:
10.1109/TEN-
CON54134.2021.9707281.
[21]
N.
Choudhary
and
D.
G.
Rao,
“The
LDC-IL
Speech
Corpora,
”
in
Pr
oceedings
of
2020
23r
d
Confer
ence
of
the
Oriental
CO-
COSD
A
International
Committee
for
the
Co-Or
dination
and
Standar
disation
of
Speec
h
Databases
and
Assessment
T
ec
hniques,
O-COCOSD
A
2020,
IEEE
,
No
v
.
2020,
pp.
28–32,
doi:
10.1109/O-COCOSD
A50338.2020.9295011.
[22]
A.
Y
ada
v
alli,
G.
S.
Mirishkar
,
and
A.
K.
V
uppala,
“Multi-T
ask
End-to-End
Model
for
T
elugu
Dialect
and
Speech
Recognition,
”
in
Pr
oceedings
of
the
Annual
Confer
ence
of
the
International
Speec
h
Communication
Association,
INTERSPEECH,
ISCA:
ISCA,
Sep
.
2022,
pp.
1387–1391,
doi:
10.21437/Interspeech.2022-10739.
[23]
D.
Ali,
I.
T
ouqir
,
A.
M.
Siddiqui,
J.
Malik
and
M.
Imran,
”F
ace
Recognition
System
Based
on
F
our
Sta
te
Hidden
Mark
o
v
Model,
”
in
IEEE
Access
,
v
ol.
10,
pp.
74436-74448,
2022,
doi:
10.1109/A
CCESS.2022.3188717.
[24]
M.
Def
ferrard,
L.
Martin,
R.
Pena,
and
N.
Perraudin,
“PyGSP:
Graph
Signal
Processing
in
Python,
”
Zenodo
,
no.
V
ersion
v0.5.0,
2017,
doi:
10.5281/zenodo.1003158.
[25]
A.
A.
Hagber
g,
D.
A.
Schult,
and
P
.
J.
Sw
art,
“Exploring
Netw
ork
Structure,
Dynamics,
and
Function
using
Netw
orkX,
”
in
Pr
o-
ceedings
of
the
7th
Python
in
Science
Confer
ence
,
Jun.
2008,
pp.
11–15,
doi:
10.25080/TCWV9851.
Analyzing
tempor
al
pr
operties
of
speec
h
tr
ajectory
using
gr
aph
structur
es
towar
ds
...
(P
ar
abattina
Bha
gath)
Evaluation Warning : The document was created with Spire.PDF for Python.
1634
❒
ISSN:
1693-6930
BIOGRAPHIES
OF
A
UTHORS
P
arabattina
Bhagath
is
a
professor
i
n
the
Department
of
Articial
Intelligence
and
Data
Science
at
Lakireddy
Bali
Reddy
Colle
ge
of
Engineering
in
Myla
v
aram,
India.
He
earned
his
M.T
ech.
and
Ph.D.
from
the
Indian
Institute
of
T
echnology
Guw
ahati,
India,
in
2010
and
2020,
respecti
v
ely
.
He
specializes
in
speech
processing,
particularly
for
resource-constrained
languages
and
IoT
de
vices.
He
has
de
v
eloped
structural
processing
methods
to
understand
speech
signals
using
their
structural
properties.
These
procedures
serv
e
as
alternati
v
e
methods
and
procedures
to
address
lo
w-resource
data,
in
addition
to
con
v
entional
machine-learning
tec
hniques.
His
re
search
has
been
published
in
t
he
proceedings
of
v
arious
esteeme
d
conferences.
He
can
be
contacted
at
email:
bha-
g
ath.2014@alumni.iitg.ac.in.
Malempati
Shanmukha
completed
his
B.T
ech.
in
Computer
Science
and
Engineering
at
Lakireddy
Bali
Reddy
Colle
ge
of
Engineering,
located
in
Myla
v
aram,
Indi
a.
His
research
focuses
on
T
elugu
spok
en
language
modeling
for
classifying
gender
and
age
groups
using
con
v
olutional
neural
netw
orks
(CNNs).
By
analyzing
speech
patterns
and
le
v
eraging
techniques
lik
e
MFCC
e
xtraction,
he
has
contrib
uted
to
de
v
eloping
ef
fecti
v
e
solutions
for
speech-based
applications
in
re
gional
languages.
He
can
be
contacted
at
email:
malempatishanmukha@gmail.com.
Gnana
Nagasri
Puthi
completed
her
B.T
ech.
in
Comput
er
Science
and
Engineering
at
Lakireddy
Bali
Reddy
Colle
ge
of
Engineering,
located
in
Myla
v
aram,
India.
Her
resea
rch
focuses
on
T
elugu
spok
en
language
modeling
for
classifying
gender
and
age
groups
using
con
v
olutional
neural
netw
orks
(CNNs).
By
analysing
speech
patterns
and
le
v
eraging
techniques
such
as
MFCC
e
xtraction,
she
has
contrib
uted
to
the
de
v
elopment
of
ef
fecti
v
e
solutions
for
speech-based
applications
in
re
gional
languages.
He
can
be
contacted
at
email:
gnananag
asri29@gmail.com.
TELK
OMNIKA
T
elecommun
Comput
El
Control,
V
ol.
23,
No.
6,
December
2025:
1626–1634
Evaluation Warning : The document was created with Spire.PDF for Python.