The discrimination of short acoustic events by artificial neural networks

The discrimination of short acoustic events poses special problems, particularly if the events have few distinguishing features and experience long term variation. A neural network is described which offers significant improvement in discrimination rates compared with traditional artificial neural networks and parametric techniques. The new network, called a cluster network, is described and discrimination performance for short vocal events is compared with those obtained using other techniques. The techniques described are applicable to the discrimination of other types of short, acoustically similar events besides the spoken letters of the alphabet