Information Technology Reference
In-Depth Information
0
,
1
Fig. 8.10.
Terminal states
^
`
k
XX
for ten instances of speech signals “zero” and
“one”. First three instances are used to build prototype vectors A0 and A1, used during the
test phase to decide the class membership
k
The remaining terminal state patterns are used to test the classification per-
formance. The classifier computes for each particular pattern
X
two scalar values
1
as follows:
h
0
h
h
0
SUM
A
0.*
X
and
h
1
SUM
A
1.*
X
where
.
denotes a point-wise multiplication operator (similar to the one used in
Matlab) while SUM() evaluates the sum of the resulting matrix. In other words,
only the elements of the prototype matrices corresponding to 1's in the
X
matrix
are preserved and summed.
The decision is taken as follows:
IF
1
X
zero
X
one
ELSE
h
!
0
h
With only six training samples,
an accuracy of
85.7%
is obtained
on the
remaining 14 samples (2 errors in 14). The result, which can still be improved by
choosing a larger training set, a different classifier, or by doing other refinements
(e.g. splitting the sound sequence into several sub-sequences) demonstrates that
Search WWH ::
Custom Search