Digital Signal Processing Reference
In-Depth Information
Although this modification to CMAC has removed the need for a
classical controller, the training of the modified cerebellar articulation
controller (MCMAC) still requires a careful planning such that all of the
cells in the MCMAC memory have to be visited. The contents of the
MCMAC memory represent the plant characteristics to be controlled by the
neuro-controller.
From our graph theory knowledge, the MCMAC memory can be also
visualized as a 3-D characteristic surface, or the contour surface. The axis of
Figure 8-4.
Noise Canceller-based on Amplitude Spectral Estimation and Wiener Filtering.
this contour surface consists of the cell indices
(m,n)
representing the
locations with quantized values of the closed-loop error and the output
and respectively, and the content of each cell. This
information is subsequently used in the computation of the training rate for
both the CMAC learning rule and the modified learning rule MCMAC.
In the framework of speech enhancement using the ubiquitous amplitude
spectral estimation (ASE) techniques, we have employed the usual short-
time Fourier transform method (STFT) to estimate the power spectral density
for both the reference noise and the noisy speech [4, 5, 6]. To achieve that
we have utilized a back-to-back configuration of a stereo microphone pair, as
shown in Figure 8-4.
This block diagram and its numerous variations and extensions are very
well-known in the speech processing community and we wanted to test our
ideas in this framework. To retrofit the stereo microphone pair suitable to the
vehicular systems we have placed them back-to-back as it has been regularly
done in recording studios.
