Word Recognition for Tractor Control

Abstract

This investigation is to develop an isolated word recognition technique in the presence of large amounts of engine noise for the purpose of controlling a tractor. The profiles of consonants and vowels are measured as spectra integrals and maxima of cepstrum respectively. These profiles are introduced to a neural network to identify the categories of words uttered by an operator. A high-pass cutoff needed for the spectrum integral in order to reduce the noise interference, is also discussed. To confirm the effectiveness of the system developed, online control experiments were carried out with the tractor at rest and favorable results were achieved.