Combining Multiple Knowledge Sources for Continuous Speech Recognition.

摘要

The objective of this project has been to develop methods and techniques to coordinate the many sources of knowledge in the decision for a continuous speech recognition system. This effort includes finding methods for effectively combining information from various knowledge sources, and for developing recognition search strategies that find the most likely word sequence, given the input speech. These search strategies must consider a very large number of word-sequence hypotheses in a computationally efficient manner. To develop and demonstrate these techniques, we designed and implemented a complete word recognition system for continuous speech which is capable of incorporating knowledge from several sources, including lexical, phonetic, phonological, and grammatical knowledge. The complete system called BYBLOS, has been shown to achieve the highest recognition accuracy to date on standard government tests using a 1000-word continuous speech corpus. Keywords: Continuous speech recognition; Hidden Markov models; Language modelling; Speech database; Speaker adaptation. (KT)