Phonotactic and Lexical Constraints in Speech Recognition

Book One

Volume

Proceedings of the AAAI Conference on Artificial Intelligence, 3

Natural Language

We demonstrate a method for partitioning a large lexicon into small equivalence classes, based on sequential phonetic and prosodic constraints. The representation is attractive for speech recognition systems because it allows all but a small number word candidates to be excluded, using only gross phonetic and prosodic information. The approach is a robust one in that the representation is relatively insensitive to phonetic variability and recognition error.