A new connectionist architecture for word spotting

A connectionist architecture and training procedure is proposed for the purpose of recognizing continuous speech using a word spotting approach. Standard backpropagation networks require a great deal of hardware to recognize the most fundamental features of speech signals. The proposed network architecture consists of units, each of which has a target input vector which represents the feature that fully activates the unit. The output of the unit is inversely related to the Euclidean distance of the unit´s actual input vector to its target input vector. The network is trained by gradient descent, using a procedure derived in the same manner as the standard backpropagation training procedure. Only preliminary tests have been run, using a single-speaker isolated-word database of spelled Spanish words, with a vocabulary consisting of the 29 letters of the Spanish alphabet. The recognition rate using the proposed architecture was 94.0%, compared with 92.5% for standard backpropagation