The University of Sussex

Interactive recurrent nets for speech recognition: principles, algorithms and experiments

Xu Li-Qun, G.D. Tattersall

In this report we have investigated a kind of dynamic neural networks interactive recurrent nets (IRNs) for isolated words recognition under a speaker-independent Telecom Environment. Details are presented on the design and description of biologically plausibility of the net topologies, a comparative study of several different learning strategies for the nets on the basis of the gradient information to minimise a pre-defined error function. Analyses and simulations are undertaken to demonstrate the efficiency of the IRNs in learning the sequential properties of some typical symbolic sequences. Also discussed is the importance to design a set of appropriate objective functions for the IRNs to trace, and their influences on the network's classification ability with respect to a highly confusable speech database. A series of experiments have been run to show the specific meaning of different weight matrices of the IRNs in the representation of speech signals. Discussions are given and the results are compared with those of using conventional DTW technique on the same speech database, the superiority of the method is established.

This paper is not available online