Dynamic synapse neural networks with a Gauss–Newton learning method for speech processing
Hassan H. Namarvar, Alireza Afshordi Dibazar, Theodore W. Berger · The Journal of the Acoustical Society of America · 2002
A continuous implementation of the biologically-based dynamic synapse neural network (DSNN) (H. H. Namarvar et al., 2001) is created by replacing the discrete nonlinear function of the synaptic cleft mechanism, which represent the neurotransmitter release in the discrete DSNN, with a continuous nonlinear function. A Gauss–Newton learning algorithm is introduced and is shown to efficiently determine the optimal parameters of a continuous DSNN being applied to a nonlinear problem in nonstationary speech processing. The continuous DSNN incorporates a new feedback architecture to model biological inhibitory mechanisms. Optimality is determined by an objective error function on the continuous DSNN output. This network has been successfully applied to the task of phoneme recognition in continuous speech. Preliminary results demonstrate that a phoneme recognizer utilizing a continuous DSNN may be successfully used as a phone recognition module in future automatic speech recognition systems. [Work supported by DARPA CBS, NASA, and ONR.]