The role of prosodic boundary cues in auditory speech processing

It has long been known that although words are not separated from each other by pauses in continuous speech, they are demarcated by several prosodic and phonetic cues (Trubetzkoy 1939). Moreover, it has been shown experimentally that listeners are sensitive to these cues (cf., among others, Nakatani & Shaffer 1978; Rietveld 1980; Quené 1991; De Pijper & Sanderman 1994). What has not been shown, however, is whether these cues are salient enough for listeners to use them online for the purposes of lexical access. There appears to be no a priori reason why explicit boundary information should be used for speech segmentation. In fact, in current models of continuous speech recognition, segmentation emerges purely as a byproduct of word identification. The first of these models to be proposed was TRACE (McClelland & Elman 1986). In TRACE, the input consists of a sequence of segments. All the words that are compatible with the input are simultaneously activated. Crucially, a competition process between words that share one or more segments from the input ensures that the model converges on a solution in which each segment belongs to one and only one word. In addition to this competition process, Shortlist (Norris 1994) implements language-specific cues to word segmentation. For English in particular, a distinction is made between candidate words that begin with a strong syllable and those that begin with a weak syllable. According to a proposal by Anne Cutler and colleagues known as the Metrical Segmentation Strategy (MSS), English listeners use the distinction between strong syllables, containing a full vowel, and weak syllables, containing a reduced vowel, in order to segment speech. That is, given that the large majority of English content words begin with a strong syllable, listeners posit a word boundary before every strong syllable. (Cutler & Norris 1988; McQueen, Norris & Cutler 1994). In Shortlist, then, candidate words that are compatible with the input segmental string and that conform to the MSS receive extra activation. Thus, TRACE does not use any prosodic information at all, while Shortlist implements probabilistic information concerning the presence of word boundaries. Neither one of these models, however, build in explicit prosodic boundary cues that are present in the signal. Still, under the assumption that during speech processing, listeners do not neglect information to which they have access, we would expect that prosodic boundary cues are exploited in addition to the mechanisms proposed in TRACE and Shortlist. In this paper, we explore this issue; specifically, we investigate whether the presence of prosodic boundaries suffices to resolve lexical ambiguities on-line. The organization of the paper is as follows. We first review some earlier studies based on the cross-modal priming paradigm. These studies found no evidence for the hypothesis that word boundaries are available on-line during auditory speech processing. We then report on a phoneme and a word detection experiment in which we investigated the same topic. Crucially, while the results of the phoneme detection experiment are in accordance with the cross-modal priming studies, the results of the word detection experiment will be shown to contrast with the previous studies. In fact, in this experiment, prosodic boundary cues appeared to be used before lexical access is completed. We conclude by discussing possible interpretations of these contrasting results as well as consequences for models of speech recognition.