A neural network speech interface with the DOS commander

A user friendly interface with the operating system requires the interfacing of speech data entry with the DOS commander. This intelligent interface relieves the user from tidy typing of DOS commands. This task necessitates the users independent recognition of the spoken commands and interfacing them with the operating system shell. Traditional speech recognition techniques are not capable of extracting significant features and are not able to generalize. These problems are conveniently solved with neural network techniques. The paper introduces the application of a three layer neural network for classification of DOS commands. A supervised learning method which is based on the back propagation technique is described. A training set including different patterns for each command recorded under different conditions, is first preprocessed, and then spotted to guarantee invariance under translation in time. Since the learning phase of a neural network based classifier is a time consuming task, it was necessary to use a data reduction technique which can preserve the command information. A linear prediction technique is used to achieve a high degree of data reduction. Only nine linear prediction coefficients are proved to be sufficient for discrimination of DOS commands. The experimental results for the neural classifier indicated a high percentage of correct classification for a training set including twelve DOS commands. The most effective number of units in the hidden layer and the value of the learning rate are conducted through extensive experimental work