Segmentation of Continuous Speech Using Phonotactics

Abstract

Listeners appear to use phonotactic constraints in the segmentation of continuous speech. Because some strings of phonemes (such as [lv] and [mr] in Dutch) never occur within the same syllable, they cue syllable boundaries, and hence possible word boundaries. Dutch listeners found it easier to detect words at the beginnings of nonsense sequences when the words were aligned with a phonotactic boundary (e.g.,pil,pill, in [pil.vrem]) than when they were misaligned (e.g.,pilin [pilm.rem]). This effect was stronger for words at the ends of nonsense sequences (e.g.,rok,skirt, in [fim.rɐk] and [fi.drɐk]). Two control experiments indicated that although part of this effect can be attributed to the role of phonotactics in speech production, there remains a significant perceptual component; the legality of sound sequences appears to be computed on-line during recognition. It is argued that segmentation is achieved via competition between candidate words, that competition is modulated by knowledge about where in the input candidates are unlikely to begin or end, and that phonotactic constraints are one of several information sources used in this segmentation process.