Digital Signal Processing Reference
In-Depth Information
where
It is not difficult to conclude from Figure 8-4 that the
and the
are given by:
In the training mode, Equations (9a,9b) and their simplified
approximations have been used to obtain the weights for the Weiner filter.
These weights were then stored into the MCMAC memory.
In the recall mode, however, all the memory elements, pointed by the
and the as address indices, are added together with
respect to a neighborhood basis function. A neighborhood basis function is
needed to restrict the impact of various memory locations in the computation
of the final winning neuron.
The neighborhood function
f (x)
can be a simple average or an
algorithm based on a uniform distribution of errors in a specific region of
memory weights. In practice, spline functions have been the most popular
basis functions for higher-order CMAC systems.
In our experiments, we have chosen a Gaussian neighborhood function as
the basis function:
where is the variance of a Gaussian distribution and is the
distance between a cell with coordinates
(i,j)
and the input with the same
coordinates. The details of the learning rule and the associated neighborhood
function and their implementations can be found in [7].
3.
ASE-MCMAC ALGORITHM
In Figure 8-5, we propose a speech enhancement system including an
MCMAC where the Weiner filter coefficients are constantly updated. The
