Autoregressive hidden Markov models and the speech signal

This thesis introduces an autoregressive hidden Markov model (HMM) and demonstrates its application to the speech signal. This new variant of the HMM is built upon the mathematical structure of the HMM and linear prediction analysis of speech signals. By incorporating these two methods into one inference algorithm, linguistic structures are inferred from a given set of speech data. These results extend historic experiments in which the HMM is used to infer linguistic information from text-based information and from the speech signal directly. Given the added robustness of this new model, the autoregressive HMM is suggested as a starting point for unsupervised learning of speech recognition and synthesis in pursuit of modeling the process of language acquisition