The application of neural networks to wordspotting

The application of dynamic neural networks for improving keyword spotter performance is discussed. A conventional wordspotter, which is based on template matching, provides an initial screening of incoming speech for possible keyword occurrences. The role of the neural network is to provide a second layer of discriminant analysis in which the final accept/reject decision is made. In experiments conducted on standard corpora for wordspotting, secondary scoring with temporally constrained Kohonen feature maps resulted in reduced false alarm rates. Preliminary results using a recurrent neural network are also presented.<>