ANALYSIS OF THE CORRELATION STRUCTURE FOR A NEURAL PREDICTIVE MODEL WITH APPLICATION TO SPEECH RECOGNITION

A speech recognizer is developed using a layered feedforward neural network to implement speech-frame prediction. A Markov chain is used to control changes in the network's weight parameters. We postulate that speech recognition accuracy is closely linked to the capability of the predictive model in representing long-term temporal correlations in speech data. Analytical expressions are obtained for the correlation functions for various types of predictive models (linear, compressively nonlinear, and jointly linear and compressively nonlinear) to determine the faithfulness of the models to the actual speech data. Analytical results, computer simulations, and speech recognition experiments suggest that when compressive nonlinear prediction and linear prediction are jointly performed within the same layer of the neural network, the model is better at capturing long-term data correlations and consequently improving speech recognition performance.