<s>
An	O
acoustic	B-Algorithm
fingerprint	I-Algorithm
is	O
a	O
condensed	O
digital	O
summary	O
,	O
a	O
fingerprint	B-Algorithm
,	O
deterministically	B-General_Concept
generated	O
from	O
an	O
audio	O
signal	O
,	O
that	O
can	O
be	O
used	O
to	O
identify	O
an	O
audio	O
sample	O
or	O
quickly	O
locate	O
similar	O
items	O
in	O
an	O
audio	O
database	O
.	O
</s>
<s>
Practical	O
uses	O
of	O
acoustic	B-Algorithm
fingerprinting	I-Algorithm
include	O
identifying	O
songs	O
,	O
melodies	O
,	O
tunes	O
,	O
or	O
advertisements	O
;	O
sound	O
effect	O
library	O
management	O
;	O
and	O
video	O
file	O
identification	O
.	O
</s>
<s>
Media	O
identification	O
using	O
acoustic	B-Algorithm
fingerprints	I-Algorithm
can	O
be	O
used	O
to	O
monitor	O
the	O
use	O
of	O
specific	O
musical	O
works	O
and	O
performances	O
on	O
radio	O
broadcast	O
,	O
records	O
,	O
CDs	B-Device
,	O
streaming	O
media	O
and	O
peer-to-peer	O
networks	O
.	O
</s>
<s>
A	O
robust	O
acoustic	B-Algorithm
fingerprint	I-Algorithm
algorithm	O
must	O
take	O
into	O
account	O
the	O
perceptual	O
characteristics	O
of	O
the	O
audio	O
.	O
</s>
<s>
If	O
two	O
files	O
sound	O
alike	O
to	O
the	O
human	O
ear	O
,	O
their	O
acoustic	B-Algorithm
fingerprints	I-Algorithm
should	O
match	O
,	O
even	O
if	O
their	O
binary	O
representations	O
are	O
quite	O
different	O
.	O
</s>
<s>
Acoustic	B-Algorithm
fingerprints	I-Algorithm
are	O
not	O
hash	B-Error_Name
functions	I-Error_Name
,	O
which	O
must	O
be	O
sensitive	O
to	O
any	O
small	O
changes	O
in	O
the	O
data	O
.	O
</s>
<s>
Acoustic	B-Algorithm
fingerprints	I-Algorithm
are	O
more	O
analogous	O
to	O
human	O
fingerprints	B-Algorithm
where	O
small	O
variations	O
that	O
are	O
insignificant	O
to	O
the	O
features	O
the	O
fingerprint	B-Algorithm
uses	O
are	O
tolerated	O
.	O
</s>
<s>
One	O
can	O
imagine	O
the	O
case	O
of	O
a	O
smeared	O
human	O
fingerprint	B-Algorithm
impression	O
which	O
can	O
accurately	O
be	O
matched	O
to	O
another	O
fingerprint	B-Algorithm
sample	O
in	O
a	O
reference	O
database	O
;	O
acoustic	B-Algorithm
fingerprints	I-Algorithm
work	O
in	O
a	O
similar	O
way	O
.	O
</s>
<s>
Perceptual	O
characteristics	O
often	O
exploited	O
by	O
audio	B-Algorithm
fingerprints	I-Algorithm
include	O
average	O
zero	B-Algorithm
crossing	I-Algorithm
rate	O
,	O
estimated	O
tempo	B-Application
,	O
average	O
spectrum	O
,	O
spectral	O
flatness	O
,	O
prominent	O
tones	O
across	O
a	O
set	O
of	O
frequency	B-Algorithm
bands	I-Algorithm
,	O
and	O
bandwidth	B-Algorithm
.	O
</s>
<s>
A	O
robust	O
acoustic	B-Algorithm
fingerprint	I-Algorithm
will	O
allow	O
a	O
recording	O
to	O
be	O
identified	O
after	O
it	O
has	O
gone	O
through	O
such	O
compression	O
,	O
even	O
if	O
the	O
audio	O
quality	O
has	O
been	O
reduced	O
significantly	O
.	O
</s>
<s>
For	O
use	O
in	O
radio	O
broadcast	O
monitoring	O
,	O
acoustic	B-Algorithm
fingerprints	I-Algorithm
should	O
also	O
be	O
insensitive	O
to	O
analog	O
transmission	O
artifacts	O
.	O
</s>
<s>
Generating	O
a	O
signature	O
from	O
the	O
audio	O
is	O
essential	O
for	O
searching	B-Application
by	I-Application
sound	I-Application
.	O
</s>
<s>
Shazam	B-Application
's	O
algorithm	O
picks	O
out	O
points	O
where	O
there	O
are	O
peaks	O
in	O
the	O
spectrogram	O
which	O
represent	O
higher	O
energy	O
content	O
.	O
</s>
<s>
Shazam	B-Application
builds	O
their	O
fingerprint	B-Algorithm
catalog	O
out	O
as	O
a	O
hash	B-Algorithm
table	I-Algorithm
,	O
where	O
the	O
key	O
is	O
the	O
frequency	O
.	O
</s>
<s>
So	O
their	O
database	O
key	O
is	O
not	O
just	O
a	O
single	O
frequency	O
,	O
it	O
is	O
a	O
hash	B-Error_Name
of	O
the	O
frequencies	O
of	O
both	O
points	O
.	O
</s>
<s>
This	O
leads	O
to	O
fewer	O
hash	B-Algorithm
collisions	I-Algorithm
improving	O
the	O
performance	O
of	O
the	O
hash	B-Algorithm
table	I-Algorithm
.	O
</s>
