<s>
Word	B-General_Concept
error	I-General_Concept
rate	I-General_Concept
(	O
WER	O
)	O
is	O
a	O
common	O
metric	O
of	O
the	O
performance	O
of	O
a	O
speech	B-Application
recognition	I-Application
or	O
machine	B-Application
translation	I-Application
system	I-Application
.	O
</s>
<s>
The	O
WER	O
is	O
derived	O
from	O
the	O
Levenshtein	O
distance	O
,	O
working	O
at	O
the	O
word	O
level	O
instead	O
of	O
the	O
phoneme	B-Language
level	O
.	O
</s>
<s>
Examination	O
of	O
this	O
issue	O
is	O
seen	O
through	O
a	O
theory	O
called	O
the	O
power	O
law	O
that	O
states	O
the	O
correlation	O
between	O
perplexity	O
and	O
word	B-General_Concept
error	I-General_Concept
rate	I-General_Concept
.	O
</s>
<s>
Word	B-General_Concept
error	I-General_Concept
rate	I-General_Concept
can	O
then	O
be	O
computed	O
as	O
:	O
</s>
<s>
When	O
reporting	O
the	O
performance	O
of	O
a	O
speech	B-Application
recognition	I-Application
system	O
,	O
sometimes	O
word	O
accuracy	O
(	O
WAcc	O
)	O
is	O
used	O
instead	O
:	O
</s>
<s>
Note	O
that	O
since	O
N	O
is	O
the	O
number	O
of	O
words	O
in	O
the	O
reference	O
,	O
the	O
word	B-General_Concept
error	I-General_Concept
rate	I-General_Concept
can	O
be	O
larger	O
than	O
1.0	O
,	O
and	O
thus	O
,	O
the	O
word	O
accuracy	O
can	O
be	O
smaller	O
than	O
0.0	O
.	O
</s>
<s>
It	O
is	O
commonly	O
believed	O
that	O
a	O
lower	O
word	B-General_Concept
error	I-General_Concept
rate	I-General_Concept
shows	O
superior	O
accuracy	O
in	O
recognition	O
of	O
speech	O
,	O
compared	O
with	O
a	O
higher	O
word	B-General_Concept
error	I-General_Concept
rate	I-General_Concept
.	O
</s>
<s>
In	O
a	O
Microsoft	O
Research	O
experiment	O
,	O
it	O
was	O
shown	O
that	O
,	O
if	O
people	O
were	O
trained	O
under	O
"	O
that	O
matches	O
the	O
optimization	O
objective	O
for	O
understanding	O
"	O
,	O
(	O
Wang	O
,	O
Acero	O
and	O
Chelba	O
,	O
2003	O
)	O
they	O
would	O
show	O
a	O
higher	O
accuracy	O
in	O
understanding	O
of	O
language	O
than	O
other	O
people	O
who	O
demonstrated	O
a	O
lower	O
word	B-General_Concept
error	I-General_Concept
rate	I-General_Concept
,	O
showing	O
that	O
true	O
understanding	O
of	O
spoken	O
language	O
relies	O
on	O
more	O
than	O
just	O
high	O
word	O
recognition	O
accuracy	O
.	O
</s>
<s>
These	O
factors	O
are	O
likely	O
to	O
be	O
specific	O
to	O
the	O
syntax	B-Application
being	O
tested	O
.	O
</s>
<s>
A	O
further	O
complication	O
is	O
added	O
by	O
whether	O
a	O
given	O
syntax	B-Application
allows	O
for	O
error	O
correction	O
and	O
,	O
if	O
it	O
does	O
,	O
how	O
easy	O
that	O
process	O
is	O
for	O
the	O
user	O
.	O
</s>
<s>
For	O
text	O
dictation	O
it	O
is	O
generally	O
agreed	O
that	O
performance	O
accuracy	O
at	O
a	O
rate	O
below	O
95%	O
is	O
not	O
acceptable	O
,	O
but	O
this	O
again	O
may	O
be	O
syntax	B-Application
and/or	O
domain	O
specific	O
,	O
e.g.	O
</s>
<s>
The	O
term	O
"	O
Single	O
Word	B-General_Concept
Error	I-General_Concept
Rate	I-General_Concept
"	O
is	O
sometimes	O
referred	O
to	O
as	O
the	O
percentage	O
of	O
incorrect	O
recognitions	O
for	O
each	O
different	O
word	O
in	O
the	O
system	O
vocabulary	O
.	O
</s>
<s>
The	O
word	B-General_Concept
error	I-General_Concept
rate	I-General_Concept
may	O
also	O
be	O
referred	O
to	O
as	O
the	O
length	O
normalized	O
edit	O
distance	O
.	O
</s>
