<s>
An	O
acoustic	B-General_Concept
model	I-General_Concept
is	O
used	O
in	O
automatic	B-Application
speech	I-Application
recognition	I-Application
to	O
represent	O
the	O
relationship	O
between	O
an	O
audio	O
signal	O
and	O
the	O
phonemes	B-Language
or	O
other	O
linguistic	O
units	O
that	O
make	O
up	O
speech	O
.	O
</s>
<s>
Modern	O
speech	B-Application
recognition	I-Application
systems	O
use	O
both	O
an	O
acoustic	B-General_Concept
model	I-General_Concept
and	O
a	O
language	B-Language
model	I-Language
to	O
represent	O
the	O
statistical	O
properties	O
of	O
speech	O
.	O
</s>
<s>
The	O
acoustic	B-General_Concept
model	I-General_Concept
models	O
the	O
relationship	O
between	O
the	O
audio	O
signal	O
and	O
the	O
phonetic	O
units	O
in	O
the	O
language	O
.	O
</s>
<s>
The	O
language	B-Language
model	I-Language
is	O
responsible	O
for	O
modeling	O
the	O
word	O
sequences	O
in	O
the	O
language	O
.	O
</s>
<s>
Most	O
modern	O
speech	B-Application
recognition	I-Application
systems	O
operate	O
on	O
the	O
audio	O
in	O
small	O
chunks	O
known	O
as	O
frames	O
with	O
an	O
approximate	O
duration	O
of	O
10ms	O
per	O
frame	O
.	O
</s>
<s>
The	O
raw	O
audio	O
signal	O
from	O
each	O
frame	O
can	O
be	O
transformed	O
by	O
applying	O
the	O
mel-frequency	B-Algorithm
cepstrum	I-Algorithm
.	O
</s>
<s>
The	O
coefficients	O
from	O
this	O
transformation	O
are	O
commonly	O
known	O
as	O
mel	B-Algorithm
frequency	I-Algorithm
cepstral	I-Algorithm
coefficients	I-Algorithm
(	O
MFCC	O
)	O
s	O
and	O
are	O
used	O
as	O
an	O
input	O
to	O
the	O
acoustic	B-General_Concept
model	I-General_Concept
along	O
with	O
other	O
features	O
.	O
</s>
<s>
Recently	O
,	O
the	O
use	O
of	O
Convolutional	B-Architecture
Neural	I-Architecture
Networks	I-Architecture
has	O
led	O
to	O
big	O
improvements	O
in	O
acoustic	O
modeling	O
.	O
</s>
<s>
Audio	O
can	O
be	O
encoded	B-Algorithm
at	O
different	O
sampling	O
rates	O
(	O
i.e.	O
</s>
<s>
Speech	B-Application
recognition	I-Application
engines	O
work	O
best	O
if	O
the	O
acoustic	B-General_Concept
model	I-General_Concept
they	O
use	O
was	O
trained	O
with	O
speech	O
audio	O
which	O
was	O
recorded	O
at	O
the	O
same	O
sampling	O
rate/bits	O
per	O
sample	O
as	O
the	O
speech	O
being	O
recognized	O
.	O
</s>
<s>
The	O
limiting	O
factor	O
for	O
telephony	O
based	O
speech	B-Application
recognition	I-Application
is	O
the	O
bandwidth	O
at	O
which	O
speech	O
can	O
be	O
transmitted	O
.	O
</s>
<s>
Therefore	O
,	O
for	O
telephony	O
based	O
speech	B-Application
recognition	I-Application
,	O
acoustic	B-General_Concept
models	I-General_Concept
should	O
be	O
trained	O
with	O
8kHz/8	O
-bit	O
speech	O
audio	O
files	O
.	O
</s>
<s>
In	O
the	O
case	O
of	O
Voice	B-Application
over	I-Application
IP	I-Application
,	O
the	O
codec	B-General_Concept
determines	O
the	O
sampling	O
rate/bits	O
per	O
sample	O
of	O
speech	O
transmission	O
.	O
</s>
<s>
Codecs	B-General_Concept
with	O
a	O
higher	O
sampling	O
rate/bits	O
per	O
sample	O
for	O
speech	O
transmission	O
(	O
which	O
improve	O
the	O
sound	O
quality	O
)	O
necessitate	O
acoustic	B-General_Concept
models	I-General_Concept
trained	O
with	O
audio	O
data	O
that	O
matches	O
that	O
sampling	O
rate/bits	O
per	O
sample	O
.	O
</s>
<s>
For	O
speech	B-Application
recognition	I-Application
on	O
a	O
standard	O
desktop	O
PC	O
,	O
the	O
limiting	O
factor	O
is	O
the	O
sound	B-Device
card	I-Device
.	O
</s>
<s>
Most	O
sound	B-Device
cards	I-Device
today	O
can	O
record	O
at	O
sampling	O
rates	O
of	O
between	O
16kHz-48kHz	O
of	O
audio	O
,	O
with	O
bit	O
rates	O
of	O
8	O
to	O
16-bits	O
per	O
sample	O
,	O
and	O
playback	O
at	O
up	O
to	O
96kHz	O
.	O
</s>
<s>
As	O
a	O
general	O
rule	O
,	O
a	O
speech	B-Application
recognition	I-Application
engine	O
works	O
better	O
with	O
acoustic	B-General_Concept
models	I-General_Concept
trained	O
with	O
speech	O
audio	O
data	O
recorded	O
at	O
higher	O
sampling	O
rates/bits	O
per	O
sample	O
.	O
</s>
<s>
Thus	O
for	O
desktop	O
speech	B-Application
recognition	I-Application
,	O
the	O
current	O
standard	O
is	O
acoustic	B-General_Concept
models	I-General_Concept
trained	O
with	O
speech	O
audio	O
data	O
recorded	O
at	O
sampling	O
rates	O
of	O
16kHz/16bits	O
per	O
sample	O
.	O
</s>
