Second International Conference on Spoken Language Processing (ICSLP'92)

Banff, Alberta, Canada
October 13-16, 1992

Processing Disfluent Speech: Recognising Disfluency Before Lexical Access

Robin J. Lickley (1,2), Ellen G. Bard (2,3)

(1) Centre for Speech Technology Research, University of Edinburgh, Edinburgh, Scotland, UK
(2) Department of Linguistics, University of Edinburgh, UK
(3) Human Communication Research Centre, University of Edinburgh, UK

As work on speech understanding moves towards the study of spontaneous rather than carefully prepared read speech, the problems posed by disfluency need to be addressed. The first problem for the processor is to detect that a disfluency has occurred. Previous experiments [11] have shown that listeners are usually able to detect disfluency within one word of the interruption. This paper presents results of a further experiment winch looks more closely at recognition points of disfluency and of the following word. It is found that listeners are able to detect that disfluency has occurred soon after the onset of the following word and prior to recognition of the word itself. Taken together with the results of an experiment with low-pass filtered speech [12], the results suggest that prosodic information may play a key role in the processing of disfluent speech.

Full Paper

Bibliographic reference.  Lickley, Robin J. / Bard, Ellen G. (1992): "Processing disfluent speech: recognising disfluency before lexical access", In ICSLP-1992, 935-938.