Analysis of the correlation structure for a neural predictive model with application to speech recognition

-A speech recogmzer ts developed usmg a layered feedforward neural network to implement speech-frame predwtlon. A Markov cham ts used to control changes in the network's wetght parameters. We postulate that speech recogmtion accuracy ts closely hnked to the capabthty of the predictive model m representing long-term temporal correlattons in speech data. Analyttcal expresstons are obtamed for the correlatton functions for various types of predwttve models (hnear, compressively nonhnear, and )omtly hnear and compresstvely nonhnear) to determme the fatthfulness of the models to the actual speech data Analyttcal results, computer simulattons, and speech recognttton experiments suggest that when compresstve nonhnear predictton and hnear prediction are jomtly performed withm the same layer of the neural network, the model ts better at capturing long-term data correlattons and consequently lmprovmg speech recogmtton performance Keywords--Temporal correlations, Joint linear/nonlinear prediction, Multllayer perceptron, HMM.