<s>
Cepstral	B-General_Concept
mean	I-General_Concept
and	I-General_Concept
variance	I-General_Concept
normalization	I-General_Concept
(	O
CMVN	B-General_Concept
)	O
is	O
a	O
computationally	O
efficient	O
normalization	O
technique	O
for	O
robust	O
speech	B-Application
recognition	I-Application
.	O
</s>
<s>
The	O
performance	O
of	O
CMVN	B-General_Concept
is	O
known	O
to	O
degrade	O
for	O
short	O
utterances	O
.	O
</s>
<s>
CMVN	B-General_Concept
minimizes	O
distortion	O
by	O
noise	O
contamination	O
for	O
robust	O
feature	B-Algorithm
extraction	I-Algorithm
by	O
linearly	O
transforming	O
the	O
cepstral	B-Algorithm
coefficients	O
to	O
have	O
the	O
same	O
segmental	O
statistics	O
.	O
</s>
<s>
Cepstral	B-Algorithm
Normalization	O
has	O
been	O
effective	O
in	O
the	O
CMU	B-General_Concept
Sphinx	I-General_Concept
for	O
maintaining	O
a	O
high	O
level	O
of	O
recognition	O
accuracy	O
over	O
a	O
wide	O
variety	O
of	O
acoustical	O
environments	O
.	O
</s>
<s>
There	O
are	O
multiple	O
algorithms	O
that	O
achieve	O
Cepstral	B-Algorithm
Normalization	O
in	O
different	O
ways	O
.	O
</s>
<s>
While	O
environment	O
selection	O
for	O
the	O
compensation	O
vectors	O
of	O
MFCDCN	O
is	O
generally	O
performed	O
on	O
an	O
utterance-by-utterance	O
basis	O
,	O
IMFCFCN	O
improves	O
on	O
it	O
by	O
allowing	O
the	O
classification	O
process	O
to	O
make	O
use	O
of	O
cepstral	B-Algorithm
vectors	O
from	O
previous	O
utterances	O
in	O
a	O
given	O
session	O
.	O
</s>
<s>
Automatic	B-Application
speech	I-Application
recognition	I-Application
(	O
ASR	O
)	O
describes	O
the	O
steps	O
of	O
transcribing	O
speech	O
utterances	O
represented	O
as	O
acoustic	O
wave	O
forms	O
to	O
written	O
words	O
.	O
</s>
<s>
As	O
is	O
,	O
CMVN	B-General_Concept
has	O
been	O
used	O
in	O
different	O
applications	O
as	O
this	O
technique	O
has	O
proven	O
to	O
provide	O
better	O
speech	B-Application
recognitions	I-Application
results	O
in	O
different	O
environments	O
.	O
</s>
<s>
CMVN	B-General_Concept
has	O
the	O
capabilities	O
to	O
reduce	O
differences	O
between	O
test	O
and	O
training	O
data	O
produced	O
by	O
channel	O
distortions	O
and	O
colorizations	O
.	O
</s>
<s>
CMVN	B-General_Concept
has	O
also	O
been	O
found	O
to	O
be	O
able	O
to	O
reduce	O
differences	O
in	O
feature	O
representation	O
between	O
speakers	O
can	O
also	O
partly	O
reduce	O
the	O
influence	O
of	O
background	O
noise	O
.	O
</s>
