Segment-Based Acoustic Models for Continuous Speech Recognition

摘要

This research aims to develop new and more accurate stochastic models for speaker-independent continuous speech recognition by extending previous work in segment-based modeling, by introducing a new hierarchical approach to representing intra-utterance statistical dependencies, and by developing language models that capture topic dependencies. These techniques, which have high computational costs because of the large search space associated with higher order models, are made feasible through a multi-pass search strategy that involves rescoring a constrained space given by an HMM decoding. We expect these different modeling techniques to result in improved recognition performance over that achieved by current systems, which handle only frame-based observations and assume that these observations are independent given an underlying state sequence. The primary research efforts and results over the past quarter have included: experimentation with a new approach to continuous density parameter adaptation; improved the language modeling software to handle more general vocabularies and reduce storage requirements, and implemented a course-grained parallel training algorithm; extended the training algorithm for discrete distribution dependence trees (our model of intra-utterance correlation) to handle missing observations with the EM algorithm; implemented a dynamic programming algorithm for word lattice rescoring algorithm and demonstrated performance comparable to N-best rescoring with the SSM. (AN).