Bootstrapping Word Boundaries: A Bottom-up Corpus-Based Approach to Speech Segmentation

Leading up the lexical garden path: Segmentation and ambiguity in spoken word recognition.

“…The reverse approach-that infrequent sequences are assumed to straddle word boundaries-has also been proposed. A variety of symbolic and connectionist systems have demonstrated that this form of phonotactic knowledge provides a plausible prelexical segmentation strategy (Aslin, Woodward, LaMendola, & Bever, 1996;Brent & Cartwright, 1996;Cairns, Shillcock, Chater, & Levy, 1997;Christiansen, Allen, & Seidenberg, 1998;Gaskell, 1994;Harrington, Watson, & Cooper, 1989). In support of these accounts, word-spotting experiments have showing that words bounded by phonotactically illegal sequences are detected more easily than words ending in a legal sequence (in adults, McQueen, 1998; in 9-month-old infants, Mattys, Jusczyk, Luce, & Morgan, 1999).…”

“…At the prelexical level, a variety of cues have been proposed, allowing segmentation through the use of acoustic cues to word onsets (Lehiste, 1960;Nakatani & Dukes, 1977) or from knowledge of statistical regularities of lexical items (such as distributional regularity, phonotactics, or metrical stress; see Brent & Cartwright, 1996;Cairns, Shillcock, Chater, & Levy, 1997;Cutler & Norris, 1988). However, because not all words can be segmented in this way, accounts of spoken word recognition also incorporate mechanisms by which lexical identification can contribute to speech segmentation.…”

Lexical viability constraints on speech segmentation by infants

“…Considerable empirical evidence supports the view that, in languages such as English and Dutch (which has the same predominant word-stress pattern), listeners indeed follow the MSS in word segmentation (Cutler & Butterfield, 1992;Cutler & Norris, 1988;McQueen, Norris, & Cutler, 1994;Vroomen & de Gelder, 1995). Similarly, knowledge of native language phonotactic patterns has been suggested as another source of information about potential word boundaries in fluent speech (Brent & Cartwright, 1996;Cairns, Shillcock, Chater, & Levy, 1997;van der Lugt, 2001;McQueen, 1998;Vitevitch & Luce, 1998, 1999. For example, certain phonotactic patterns occur much more frequently between words than within the words of a language.…”

Ambiguity, competition, and blending in spoken word recognition

“…Recent experimental and computational research has demonstrated the value of statistical and distributional information, in both the development of language abilities, and the adult system (Brent & Cartwright, 1996;Cairns, Shillcock, Chater & Levy, 1997;Saffran, Aslin & Newport, 1996). The use of a simple recurrent network allows the model to pick up statistical information during training and to reflect conditional probabilities during states of ambiguity (e.g., before a word's UP).…”