A Linear Discriminant Analysis (LDA) is proposed to be used as a tool for verification of the usefulness of additional modeling of prediction error signal in a context-dependent Hidden Control Neural Network (HCNN-CDF) system. While showing that the squared prediction error of the HCNN-CDF system is not a white noise, the LDA also identified the most important components of the error vector signal which could be used to support the discrimination among predictive models of the system. This information is then used to enhance the context-dependent HCNN model to become more discriminant by modeling dynamics in acoustics and in error signal space. Our preliminary speaker-dependent tests with vocabulary size of 100 words confirmed an increase in word recognition rate and in discrimination power of the system. Recently, continuous speech recognition experiments at perplexity 100 have shown an increase in word accuracy from 76% to 87% on the test set database.
Keywords: Automatic Speech Recognition, context-dependent Hidden Control Neural Network, large vocabulary countinuous speech recognition, Lin- ear Discriminant Analysis.