<s>
The	O
audio	B-Algorithm
deepfake	I-Algorithm
(	O
also	O
known	O
as	O
voice	O
cloning	O
)	O
is	O
a	O
type	O
of	O
artificial	B-Application
intelligence	I-Application
used	O
to	O
create	O
convincing	O
speech	O
sentences	O
that	O
sound	O
like	O
specific	O
people	O
saying	O
things	O
they	O
did	O
not	O
say	O
.	O
</s>
<s>
This	O
technology	O
can	O
also	O
create	O
more	O
personalized	O
digital	O
assistants	O
and	O
natural-sounding	O
text-to-speech	B-Application
as	O
well	O
as	O
speech	O
translation	O
services	O
.	O
</s>
<s>
Audio	B-Algorithm
deepfakes	I-Algorithm
,	O
recently	O
called	O
audio	O
manipulations	O
,	O
are	O
becoming	O
widely	O
accessible	O
using	O
simple	O
mobile	O
devices	O
or	O
personal	O
PCs	B-Device
.	O
</s>
<s>
This	O
has	O
led	O
to	O
cybersecurity	O
concerns	O
among	O
the	O
global	O
public	O
about	O
the	O
side	O
effects	O
of	O
using	O
audio	B-Algorithm
deepfakes	I-Algorithm
.	O
</s>
<s>
Audio	B-Algorithm
deepfake	I-Algorithm
attackers	O
have	O
targeted	O
individuals	O
and	O
organizations	O
,	O
including	O
politicians	O
and	O
governments	O
.	O
</s>
<s>
Audio	B-Algorithm
deepfakes	I-Algorithm
can	O
be	O
divided	O
into	O
three	O
different	O
categories	O
:	O
</s>
<s>
Replay-based	O
deepfakes	B-Application
are	O
malicious	O
works	O
that	O
aim	O
to	O
reproduce	O
a	O
recording	O
of	O
the	O
interlocutor	O
's	O
voice	O
.	O
</s>
<s>
Text-dependent	O
speaker	B-Application
verification	I-Application
can	O
be	O
used	O
to	O
defend	O
against	O
replay-based	O
attacks	O
.	O
</s>
<s>
A	O
current	O
technique	O
that	O
detects	O
end-to-end	O
replay	O
attacks	O
is	O
the	O
use	O
of	O
deep	B-Architecture
convolutional	I-Architecture
neural	I-Architecture
networks	I-Architecture
.	O
</s>
<s>
The	O
category	O
based	O
on	O
speech	B-Application
synthesis	I-Application
refers	O
to	O
the	O
artificial	O
production	O
of	O
human	O
speech	O
,	O
using	O
software	O
or	O
hardware	O
system	O
programs	O
.	O
</s>
<s>
Speech	B-Application
synthesis	I-Application
includes	O
Text-To-Speech	B-Application
,	O
which	O
aims	O
to	O
transform	O
the	O
text	O
into	O
acceptable	O
and	O
natural	O
speech	O
in	O
real-time	O
,	O
making	O
the	O
speech	O
sound	O
in	O
line	O
with	O
the	O
text	O
input	O
,	O
using	O
the	O
rules	O
of	O
linguistic	O
description	O
of	O
the	O
text	O
.	O
</s>
<s>
A	O
classical	O
system	O
of	O
this	O
type	O
consists	O
of	O
three	O
modules	O
:	O
a	O
text	O
analysis	O
model	O
,	O
an	O
acoustic	O
model	O
,	O
and	O
a	O
vocoder	B-Application
.	O
</s>
<s>
Second	O
,	O
the	O
Text-To-Speech	B-Application
model	O
must	O
be	O
trained	O
using	O
these	O
data	O
to	O
build	O
a	O
synthetic	O
audio	O
generation	O
model	O
.	O
</s>
<s>
Finally	O
,	O
the	O
vocoder	B-Application
learns	O
to	O
create	O
vocal	O
waveforms	O
based	O
on	O
the	O
parameters	O
of	O
the	O
acoustic	O
features	O
.	O
</s>
<s>
The	O
first	O
breakthrough	O
in	O
this	O
regard	O
was	O
introduced	O
by	O
WaveNet	B-Application
,	O
a	O
neural	B-Architecture
network	I-Architecture
for	O
generating	O
raw	O
audio	O
waveforms	O
capable	O
of	O
emulating	O
the	O
characteristics	O
of	O
many	O
different	O
speakers	O
.	O
</s>
<s>
Unfortunately	O
,	O
Text-To-Speech	B-Application
is	O
highly	O
dependent	O
on	O
the	O
quality	O
of	O
the	O
voice	O
corpus	O
used	O
to	O
realize	O
the	O
system	O
,	O
and	O
creating	O
an	O
entire	O
voice	O
corpus	O
is	O
expensive	O
.	O
</s>
<s>
Another	O
disadvantage	O
is	O
that	O
speech	B-Application
synthesis	I-Application
systems	O
do	O
not	O
recognize	O
periods	O
or	O
special	O
characters	O
.	O
</s>
<s>
Audio	B-Algorithm
deepfake	I-Algorithm
based	O
on	O
imitation	O
is	O
a	O
way	O
of	O
transforming	O
an	O
original	O
speech	O
from	O
one	O
speaker	O
-	O
the	O
original	O
-	O
so	O
that	O
it	O
sounds	O
spoken	O
like	O
another	O
speaker	O
-	O
the	O
target	O
one	O
.	O
</s>
<s>
In	O
recent	O
years	O
,	O
the	O
most	O
popular	O
approach	O
involves	O
the	O
use	O
of	O
particular	O
neural	B-Architecture
networks	I-Architecture
called	O
Generative	B-Algorithm
Adversarial	I-Algorithm
Networks	I-Algorithm
(	O
GAN	O
)	O
due	O
to	O
their	O
flexibility	O
as	O
well	O
as	O
high-quality	O
results	O
.	O
</s>
<s>
The	O
audio	B-Algorithm
deepfake	I-Algorithm
detection	O
task	O
determines	O
whether	O
the	O
given	O
speech	O
audio	O
is	O
real	O
or	O
fake	O
.	O
</s>
<s>
In	O
general	O
,	O
deepfake	B-Application
detection	O
methods	O
can	O
be	O
divided	O
into	O
two	O
categories	O
based	O
on	O
the	O
aspect	O
they	O
leverage	O
to	O
perform	O
the	O
detection	O
task	O
.	O
</s>
<s>
Many	O
machine	O
learning	O
and	O
deep	B-Algorithm
learning	I-Algorithm
models	O
have	O
been	O
developed	O
using	O
different	O
strategies	O
to	O
detect	O
fake	O
audio	O
.	O
</s>
<s>
Over	O
the	O
years	O
,	O
many	O
researchers	O
have	O
shown	O
that	O
machine	O
learning	O
approaches	O
are	O
more	O
accurate	O
than	O
deep	B-Algorithm
learning	I-Algorithm
methods	O
,	O
regardless	O
of	O
the	O
features	O
used	O
.	O
</s>
<s>
Instead	O
,	O
when	O
deep	B-Algorithm
learning	I-Algorithm
algorithms	O
are	O
used	O
,	O
specific	O
transformations	O
are	O
required	O
on	O
the	O
audio	O
files	O
to	O
ensure	O
that	O
the	O
algorithms	O
can	O
handle	O
them	O
.	O
</s>
<s>
There	O
are	O
several	O
open-source	O
implementations	O
of	O
different	O
detection	O
methods	O
,	O
and	O
usually	O
many	O
research	O
groups	O
release	O
them	O
on	O
a	O
public	O
hosting	O
service	O
like	O
GitHub	B-Application
.	O
</s>
<s>
The	O
audio	B-Algorithm
deepfake	I-Algorithm
is	O
a	O
very	O
recent	O
field	O
of	O
research	O
.	O
</s>
<s>
Regarding	O
the	O
generation	O
,	O
the	O
most	O
significant	O
aspect	O
is	O
the	O
credibility	O
of	O
the	O
victim	O
,	O
i.e.	O
,	O
the	O
perceptual	O
quality	O
of	O
the	O
audio	B-Algorithm
deepfake	I-Algorithm
.	O
</s>
<s>
Several	O
metrics	O
determine	O
the	O
level	O
of	O
accuracy	O
of	O
audio	B-Algorithm
deepfake	I-Algorithm
generation	O
,	O
and	O
the	O
most	O
widely	O
used	O
is	O
the	O
MOS	O
(	O
Mean	O
Opinion	O
Score	O
)	O
,	O
which	O
is	O
the	O
arithmetic	O
average	O
of	O
user	O
ratings	O
.	O
</s>
<s>
The	O
sampling	B-Algorithm
rate	I-Algorithm
also	O
plays	O
an	O
essential	O
role	O
in	O
detecting	O
and	O
generating	O
audio	B-Algorithm
deepfakes	I-Algorithm
.	O
</s>
<s>
Currently	O
,	O
available	O
datasets	O
have	O
a	O
sampling	B-Algorithm
rate	I-Algorithm
of	O
around	O
16kHz	O
,	O
significantly	O
reducing	O
speech	O
quality	O
.	O
</s>
<s>
An	O
increase	O
in	O
the	O
sampling	B-Algorithm
rate	I-Algorithm
could	O
lead	O
to	O
higher	O
quality	O
generation	O
.	O
</s>
<s>
Most	O
studies	O
focus	O
on	O
detecting	O
audio	B-Algorithm
deepfake	I-Algorithm
in	O
the	O
English	O
language	O
,	O
not	O
paying	O
much	O
attention	O
to	O
the	O
most	O
spoken	O
languages	O
like	O
Chinese	O
and	O
Spanish	O
,	O
as	O
well	O
as	O
Hindi	O
and	O
Arabic	O
.	O
</s>
<s>
In	O
other	O
fields	O
of	O
audio	O
,	O
such	O
as	O
speaker	B-Application
recognition	I-Application
,	O
the	O
accent	O
has	O
been	O
found	O
to	O
influence	O
the	O
performance	O
significantly	O
,	O
so	O
it	O
is	O
expected	O
that	O
this	O
feature	O
could	O
affect	O
the	O
models	O
 '	O
performance	O
even	O
in	O
this	O
detection	O
task	O
.	O
</s>
<s>
For	O
this	O
reason	O
,	O
many	O
researchers	O
have	O
suggested	O
following	O
a	O
Self-Supervised	B-General_Concept
Learning	I-General_Concept
approach	O
,	O
dealing	O
with	O
unlabeled	O
data	O
to	O
work	O
effectively	O
in	O
detection	O
tasks	O
and	O
improving	O
the	O
model	O
's	O
scalability	O
,	O
and	O
,	O
at	O
the	O
same	O
time	O
,	O
decreasing	O
the	O
computational	O
cost	O
.	O
</s>
<s>
In	O
addition	O
,	O
most	O
of	O
the	O
effort	O
is	O
focused	O
on	O
detecting	O
Synthetic-based	O
audio	B-Algorithm
deepfakes	I-Algorithm
,	O
and	O
few	O
studies	O
are	O
analyzing	O
imitation-based	O
due	O
to	O
their	O
intrinsic	O
difficulty	O
in	O
the	O
generation	O
process	O
.	O
</s>
<s>
Over	O
the	O
years	O
,	O
there	O
has	O
been	O
an	O
increase	O
in	O
techniques	O
aimed	O
at	O
defending	O
against	O
malicious	O
actions	O
that	O
audio	B-Algorithm
deepfake	I-Algorithm
could	O
bring	O
,	O
such	O
as	O
identity	O
theft	O
and	O
manipulation	O
of	O
speeches	O
by	O
the	O
nation	O
's	O
governors	O
.	O
</s>
<s>
To	O
prevent	O
deepfakes	B-Application
,	O
some	O
suggest	O
using	O
blockchain	O
and	O
other	O
distributed	O
ledger	O
technologies	O
(	O
DLT	O
)	O
to	O
identify	O
the	O
provenance	O
of	O
data	O
and	O
track	O
information	O
.	O
</s>
<s>
Extracting	O
and	O
comparing	O
affective	O
cues	O
corresponding	O
to	O
perceived	O
emotions	O
from	O
digital	O
content	O
has	O
also	O
been	O
proposed	O
to	O
combat	O
deepfakes	B-Application
.	O
</s>
<s>
That	O
way	O
,	O
those	O
who	O
create	O
the	O
generation	O
models	O
,	O
perhaps	O
for	O
nefarious	O
purposes	O
,	O
would	O
not	O
know	O
precisely	O
what	O
features	O
facilitate	O
the	O
detection	O
of	O
a	O
deepfake	B-Application
,	O
discouraging	O
possible	O
attackers	O
.	O
</s>
<s>
Numerous	O
research	O
groups	O
worldwide	O
are	O
working	O
to	O
recognize	O
media	O
manipulations	O
;	O
i.e.	O
,	O
audio	B-Algorithm
deepfakes	I-Algorithm
but	O
also	O
image	O
and	O
video	O
deepfake	B-Application
.	O
</s>
<s>
Another	O
research	O
program	O
is	O
the	O
Preserving	O
Media	O
Trustworthiness	O
in	O
the	B-Application
Artificial	I-Application
Intelligence	I-Application
Era	O
(	O
PREMIER	O
)	O
program	O
,	O
funded	O
by	O
the	O
Italian	O
Ministry	O
of	O
Education	O
,	O
University	O
and	O
Research	O
(	O
MIUR	O
)	O
and	O
run	O
by	O
five	O
Italian	O
universities	O
.	O
</s>
<s>
In	O
the	O
last	O
few	O
years	O
,	O
numerous	O
challenges	O
have	O
been	O
organized	O
to	O
push	O
this	O
field	O
of	O
audio	B-Algorithm
deepfake	I-Algorithm
research	O
even	O
further	O
.	O
</s>
<s>
The	O
most	O
famous	O
world	O
challenge	O
is	O
the	O
ASVspoof	O
,	O
the	O
Automatic	O
Speaker	B-Application
Verification	I-Application
Spoofing	O
and	O
Countermeasures	O
Challenge	O
.	O
</s>
<s>
Another	O
recent	O
challenge	O
is	O
the	O
ADD	O
—	O
Audio	B-Algorithm
Deepfake	I-Algorithm
Detection	O
—	O
which	O
considers	O
fake	O
situations	O
in	O
a	O
more	O
real-life	O
scenario	O
.	O
</s>
