| Summary: | 碩士 === 國立海洋大學 === 電子工程學系 === 81 === Recent advances in massivelly parallel neural networks have
made it plausible to use matched filter banks as pattern
classifiers for spatiotemporal pattern (STP) such as speech,
sonar,radar. This thesis begins with the introduction of
spatiotemporal neural network that implements a matched filter
bank which could process STP. The definition of STP is also
defined. For the purpose of better understanding of STNN,the
speech signal is chosen as example of STP for simulation of
speech recognition by STNN. Feature extraction of speech is
processed by three signal processing methords for comparison:
power spectrum,LPC and cepstrum. The STNN dynamic equation used
in the thesis was referenced from the books of Freeman and
Hecht-Nielsen, but we also modify the dynamic equation for
reasons explained in the thesis.The learning algorithms for
STNN we used in simulation are kohonen rule and kosko/klopf
learning rule, for spatial part and temporal connection part,
separately.
|
|---|
