Speech recognition by using the extended associative memory neural network (EAMNN)

The paper concerns noisy speech recognition by using the extended bi-directional associative memory neural network which consists of a MLP and a connected feedback network. For feedback associative memory, we forward a fast gradient descent algorithm of error backpropagation which improves associative memory ability and picks up training speed. Through changing the error energy function update weights according to the output error, and deducting appropriate fast learning algorithms, the training speed of the network is increased by 300-500%. The whole system has a high adaptive, robust, error-tolerating and associative memory ability for weak noisy speech signals.