Abstract

Speech segmentation research considers how listeners identify word boundaries in the ongoing stream of sounds. Numerous linguistic cues to word boundaries have been identified, e.g. lexical stress, phonotactic transition probabilities, coarticulation, glottalisation, word-initial lengthening, and lexical information. I will begin by outlining a hierarchical framework (Mattys, White & Melhorn, 2005) which suggests that listeners integrate potential cues hierarchically, with descending weights allocated to lexical, segmental, and prosodic cues. Lower-level cues drive segmentation when the interpretive conditions are sub-optimal.

I will then present some recent studies examining the impact on segmentation of variations in interpretive conditions, including: (a) the state of linguistic knowledge of the listener, as seen in L2 speakers; (b) the nature of the speech input, in particular the contrast between careful read speech and natural spontaneous speech. With regard to (a), I will discuss a study examining how the level of competence of Hungarian L2 speakers of English impacts on their use of lexical knowledge and lexical stress as word boundary cues. With regard to (b), I will describe the ongoing development of a corpus of segmentation-oriented spontaneous speech and present some preliminary data on contrasts between read and spontaneous speech in terms of the availability of potential segmentation cues. Taken together, such studies converge towards integration of speaker and listerner variability within models of speech recognition.