Locally recurrent neural networks for efficient realization of a speech recognizer

The computational complexity of speech recognizers based on fully connected recurrent neural networks, i.e. the large number of connections, prevents a hardware realization. We introduced locally connected recurrent neural networks in order to keep the properties of recurrent neural networks and to reduce the connectivity density of the network. A special form of feature presentation and output coding is developed which reduces the computational complexity and allows learning of long-term dependencies. By applying all these methods a locally recurrent neural network results, which has only one third of the weights as a fully connected recurrent network. Thus, with this concept a speech recognition system can be realized on a single VLSI-Chip.