A neural predictive HMM architecture for speech and speaker recognition

A speech recognition system is developed utilizing a neural predictive hidden Markov model architecture. In this framework multi-layered feed-forward neural networks are employed to implement accurate nonlinear speech frame prediction. A Markov chain is used to control changes in the weight parameters of these predictive networks. We show that speech recognition accuracy is closely linked to the capability of the predictive model in representing long-term temporal correlations in the speech signal.;Analytical expressions are derived for the correlation functions of various types of predictive models (linear, nonlinear, and jointly linear/nonlinear) in order to determine the faithfulness of these models to the actual speech data. This analysis, computer simulation and speech recognition experiments suggest that when both nonlinear and linear prediction are jointly performed within the hidden layer of the neural network predictor, the model is better at capturing long-term speech data correlations, and consequently, improving speech recognition accuracy. A convergence analysis was carried out indicating the advantages, in terms of better conditioning, of including explicit linear nodes within the hidden layer of neural predictive models.;Performance estimates are also obtained for the neural predictive HMM algorithm when mapped to a parallel systolic-based architecture, indicating its real time speech recognition performance capabilities.;A corrective training technique based on the Maximum Mutual Information criterion is then developed for training this class of models. Speech recognition systems trained using this approach were shown to significantly outperform their noncorrective Maximum Likelihood trained counterparts in several speech recognition tasks.;Finally, the Neural Predictive Hidden Markov Model (NP-HMM) is successfully used to implement an accurate and robust fixed-text speaker identification system.