Optimal Tailoring of Trajectories, Growing Training Sets and Recurrent Networks for Spoken Word Recognition

A novel system that efficiently integrates two types of neural networks for reliably performing isolated word recognition is described. The recognition system comprises of a feature extractor that includes a Self Organizing Map for an optimal tailoring of trajectory representations of words in reduced dimension feature spaces. Experimental results indicate that such lower dimensional trajectories can provide a reliable representation of spoken words, while reducing the training complexity for the recognition of the trajectory. A recurrent neural network is employed for performing trajectory recognition and a method that allows to progressively grow the training set is utilized for network training. The optimal tailoring of trajectories and growing training sets are two innovations that result in a superior training of the recurrent neural network, which in turn delivers a robust word recognition performance tolerating wide variations in the speech signal.