Detecting disfluency in spontaneous speech

Autor: Lickley, Robin J.
Rok vydání: 1994
Předmět:
Druh dokumentu: Electronic Thesis or Dissertation
Popis: Disfluent speech presents problems for both computational and psycholinguistic models of speech processing. The surface strings produced when speech is interrupted by disfluency require complex editing processes from computational models in order to produce well-formed strings for parsers. There is little empirical evidence about how the human speech processing mechanism deals with disfluences, but our everyday experience of listening to speech suggests that we can deal with disfluencies very smoothly and efficiently. One of the first problems for a speech processor is to detect that disfluency has occurred. No reliable acoustic or prosodic cues have been identified which signal the presence of a discontinuity. In this thesis, the main aims are to address this problem by first establishing detection points for a set of disfluent utterances and then finding out what acoustic and prosodic cues are available at these points. The main part of the study consists of a series of 5 perceptual experiments, followed by acoustic and prosodic analyses. The first 3 experiments establish detection points for disfluencies and relate these points to recognition points of the words in the vicinity of the interruption. The last 2 experiments examine the rôle of prosodic information in detecting disfluency, first over whole utterances and then focussing in on the region of the interruption. The acoustic and prosodic analyses of the experimental stimuli match responses indicating disfluency detection to events in the speech signal which might act as cues. The results of the first 3 experiments show that disfluency can be recognised very early, usually within the first word after the interruption point. Importantly, it is also shown that the detection of disfluency can be achieved before the word is recognised - non-syntactic information is used.
Databáze: Networked Digital Library of Theses & Dissertations