Paper

Effects of Duration, Locality, and Surprisal in Speech Disfluency Prediction in English Spontaneous Speech

---

Abstract

This study examines the role of two influential theories of language processing, Surprisal Theory and Dependency Locality Theory (DLT), in predicting disfluencies (fillers and reparandums) in the Switchboard corpus of English conversational speech. Using Generalized Linear Mixed Models for this task, we incorporate syntactic factors (DLT-inspired costs and syntactic surprisal) in addition to lexical surprisal and duration, thus going beyond the local lexical frequency and predictability used in previous work on modelling word durations in Switchboard speech. Our results indicate that compared to fluent words, words preceding disfluencies tend to have lower lexical surprisal (hence higher activation levels) and lower syntactic complexity (low DLT costs and low syntactic surprisal except for reparandums). Disfluencies tend to occur before upcoming difficulties, i.e., high lexical surprisal words (low activation levels) with high syntactic complexity (high DLT costs and high syntactic surprisal). Further, we see that reparandums behave almost similarly to disfluent fillers with differences possibly arising due to effects being present in the word choice of the reparandum, i.e., in the disfluency itself rather than surrounding it. Moreover, words preceding disfluencies tend to be function words and have longer durations compared to their fluent counterparts, and word duration is a very effective predictor of disfluencies. Overall, speakers may be leveraging the differences in access between content and function words during planning as part of a mechanism to adapt for disfluencies while coordinating between planning and articulation.

Keywords: psycholinguistics, disfluencies, speech, dependency locality theory, surprisal

How to Cite:

Dammalapati, S., Rajkumar, R., Ranjan, S. & Agarwal, S., (2021) “Effects of Duration, Locality, and Surprisal in Speech Disfluency Prediction in English Spontaneous Speech”, Society for Computation in Linguistics 4(1), 91-101. doi: https://doi.org/10.7275/s5yn-7d90