Noise robust speech recognition using dynamic synapse neural networks.

摘要

Training parameters of Dynamic Synapse Neural Networks (DSNNs) is an ill-conditioned problem, due to nonlinearity, high dimensionality, and large-scale information processing. In addition to convergence problem, a trained neural network might be unstable which causes a poor generalization for unseen data. Therefore, dynamic networks have been lacking proper learning algorithm to extract complex information from spatio-temporal patterns.; In this dissertation, a new discrete version of DSNN based on the Wavelet filter bank and the Genetic Algorithms (GAs) learning method is introduced to optimize the DSNN parameters. Despite the success of GAs on training of a small-scaled DSNN, training of a large-scale DSNN with GAs becomes a challenging problem. To address this issue, a general class of DSNNs from biological and mathematical perspective is introduced to model the average local population activity of neurons. Efficient learning methods to optimize the DSNN parameters with point-process I/O are developed based on the Gauss-Newton learning and an interior trust-region (TR) nonlinear optimization method.; In order to study the robustness of the DSNNs, the following assumptions are made to simplify the problem: (a) continuous DSNNs and (b) removing the dynamic characteristics of I/O signals---working in steady state mode. These assumptions reduce DSNNs to their sub-classes, i.e., static neural network. A robust command recognition system is developed using frequency domain analysis and radial basis function networks with the kernel learning. A system for multiword spotting in continuous speech has been developed. A proof-of-concept of this system is illustrated by a demo which is implemented to identify 27 words for a drive-through MacDonald automatic system. The sensitivity, specificity and correct classification rate of the proposed system for a slow rate speaking are 82.6%, 69.1 %, and 72.4%, respectively.; Moreover, to construct a hybrid noise robust speech recognition system, a speaker-independent phoneme recognition system is developed based on the wavelet de-noising preprocessing at front-end. The improvement of 16.31% to 20.20% on the accuracy of continuous phoneme recognition under SNR levels of 20 to -5 dB for the TIMIT corpus was obtained.