COST278 and ISCA Tutorial and Research Workshop (ITRW) on Robustness Issues in Conversational Interaction
University of East Anglia, Norwich, UK
In this paper two studies are presented in which the detection of speech recognition errors on the word level was examined. In the first study, memory-based and transformation-based machine learning was used for the task, using confidence, lexical, contextual and discourse features. In the second study, we investigated which factors humans benefit from when detecting errors. Information from the speech recogniser (i.e. word confidence scores and 5-best lists) and contextual information were the factors investigated. The results show that word confidence scores are useful and that lexical and contextual (both from the utterance and from the discourse) features further improve performance.
Bibliographic reference. Skantze, Gabriel / Edlund, Jens (2004): "Early error detection on word level", In Robust2004, paper 17.