A voting and predictive Neural Network system for use in a new artificial Larynx

Abstract

A new artificial Larynx is currently under development at the University of the Witwatersrand, Johannesburg. This device uses dynamic tongue movement from a palatometer system to infer what the user is trying to say. Feature selection algorithms extract information from the palatometer data and are then used as input to a Multi-Layer Perceptron Neural Network. This paper deals with improving the success rate of the Neural Networks by using a voting system as well as a word prediction system. By using a voting system unknown non-rejected input words were correctly identified 93.5% of the time, while the system has a rejection rate of 17.36%. A set of grammar rules were developed for the word set and this improved the number of correct unknown, non-rejected words to 94.14% but increased the rejection rate to 17.74%.