A piecewise approach to connectionist networks for speech recognition

A three-layer two-dimensional connectionist network is described. The network is structured in such a way that the trained net can be extended in both dimensions without losing its existing knowledge. Noncontributing or low-contributing units can be identified and pruned from the net to improve efficiency. These features enable the size of the net to be tailored to a particular task without the need for time-consuming retraining from scratch. A net of this type has been applied to the task of speaker-independent isolated-word speech recognition, with a vocabulary consisting of the digits `zero´ to `nine´ plus silence