A comparison of speech feature extraction employing autonomous neural network topologies

Describes results obtained from an experimental speech recognition system designed to assess the suitability of several different types of neural network when used for feature extraction. A number of independent speech samples were acquired using a commercial system (Micro Speech Laboratory) at a sampling rate of 10 kHz and encoded into 10 data-bits per sample. The data was further factorized by three common algorithms in order to extract alternative characteristics of feature structure, namely: (a) a 12-parameter fast-Fourier transform (FFT); (b) a 12-parameter FFT in association with the mean energy value per sample frame; and (c) a 12-parameter linear predictive coding (LPC) Cholesky-based method. The data obtained from these three factorizations was used to train each of the following neural network topologies: (a) Adaline; (b) Perceptron; and (c) Back-propagation