Statistical Modeling for Continuous Speech Recognition

摘要

The authors' research into developing robust, high-performance continuous speech recognition systems for large-vocabulary tasks, such as battle management, has focused on the development of accurate mathematical models for the different phonemes that occur in English. The research performed in this project has been in three general areas: Hidden Markov Models, Stochastic Segment Models, and Rapid Speaker Adaptation. Hidden Markov models and stochastic segment models are two distinct methods of modeling phonetic coarticulation, i.e., the variation of phonemes in the context of other phonemes. The authors have tested the use of context-dependent hidden Markov models in BYBLOS, the BBN continuous speech recognition system, and report on word recognition accuracy in a 1000-word task domain. In contrast to hidden Markov modeling which models each part of a phoneme independently, stochastic segment modeling models each phoneme as a whole unit, and therefore has the promise of improved performance, as our preliminary experiments indicate.