An Improved Model for Recognizing Disfluencies in Conversational Speech

This paper presents a novel metadata extraction (MDE) system for automatically detecting edited words, fillers, and self-interruption points in conversational speech. Our edit word detection sub-system combines a Tree Adjoining Grammar (TAG) noisy channel model, a statistical syntactic language model, and a MaxEnt reranker. Hand-built, deterministic rules are used to detect fillers. Self-interruption points are explicitly determined by detected fillers and edited words. We have evaluated our system for these three tasks on two types of input: manually annotated words and automatically recognized speech-to-text tokens. In all six cases, our system has improved the state-of-the-art, as measured in a recent blind evaluation.