Advanced computational models and learning theories for spoken language processing

Various methods for fast search through finite-state machines, Bayesian solutions for modeling and classification of speech, and a training approach for minimizing errors in large vocabulary continuous speech recognition (LVCSR) technology are discussed. The development of effective speech recognition decoders requires understanding software programming skills and sufficient understanding of LVCSR technology. The weighted finite-state transducer (WFST) framework provides an alternative to LVCSR and enables efficient global optimization of the search space and a one-pass decoding over the speech input using all knowledge simultaneously. The MCE-based training framework is extended to make full use of WFST to obtain directly model word accuracy.