In this paper, a neural network architecture is presented for speech recognition. A library of two words is created by capturing male and female 8-bit voice data using a SoundBlaster sound card on an IBM-PC. The captured data are sent through four different band-pass filters for feature extraction of each word presented. Then the band-pass filter outputs are put through two cascaded low-pass filters resulting in down-sampling. The low-pass filter outputs are presented to an error backpropagation neural network for training. MATLAB, a mathematical software environment, is used for the training and testing of the algorithm. The trained neural network performs reliably distinguishing the two words regardless of speaker and gender.
 

• (No web site available)
• Return to 1994 Project List
• Return to main Senior Project page