Temporal sequence classification by memory neuron networks

摘要

A recurrent connectionist architecture, called memory neuron network (MNN), is applied for classification of temporal sequences. The network architecture allows a learnable parametric representation of the activation history of the units. It has been shown that the network is generalized version of network with time-delays. The learning protocol has been developed to train a collection of such networks as discriminant models for classes of temoral sequences. The design is tested in classification of voiced plosives /B/,/D/,/G/. Due to continuous movement of the articulators, the spectral characteristics of the speech signal change during transitions from one phoneme to the other. MNN has been used to model this dynamic behavior during the transitions from plosive sounds to vowels.