A neural network isolated word recognition system for moderate sized databases

A neural net-based isolated word recognition system that was tested on a moderate sized database is presented. The system includes an acoustic preprocessor, feature maps and single layer feedforward networks which are used for classification of the input from the preprocessor. The feature maps are trained using the K-means algorithm. The single layer feedforward networks are trained using the backpropagation algorithm. Several methods are studied to partition a moderate sized database into smaller groups, so as to obtain high speaker dependent recognition rates