Pseudo-Syntactic Language Modeling for Disfluent Speech Recognition

Abstract Language models for speech recognition are generally trained on text corpora. Since these corpora do not contain the disfluencies found in natural speech, there is a train/test mismatch when these models are applied to conversational speech. In this work we investigate a language model (LM) designed to model these disfluencies as a syntactic process. By modeling selfcorrections we obtain an improvement over our baseline syntactic model. We also obtain a 30% relative reduction in perplexity from the best performing standard N-gram model when we interpolate it with our syntactically derived models.