BibTeX
@MISC{Hough_andliterature,
author = {Julian Hough and Matthew Purver},
title = {and Literature},
year = {}
}
Share
 |
 |
 |
 |
||
OpenURL
Abstract
We present STIR (STrongly Incremen-tal Repair detection), a system that de-tects speech repairs and edit terms on transcripts incrementally with minimal la-tency. STIR uses information-theoretic measures from n-gram models as its prin-cipal decision features in a pipeline of classifiers detecting the the different stages of repairs. Results on the Switchboard dis-fluency tagged corpus show utterance-final accuracy on a par with state-of-the-art in-cremental repair detection methods, but with better incremental accuracy, faster time-to-detection and less computational overhead. We evaluate its performance us-ing incremental metrics and propose new repair processing evaluation standards. 1
Keyphrases
n-gram model edit term prin-cipal decision feature incremen-tal repair detection different stage performance us-ing incremental metric corpus show utterance-final accuracy information-theoretic measure minimal la-tency evaluation standard switchboard dis-fluency new repair speech repair computational overhead state-of-the-art in-cremental repair detection method incremental accuracy
