Pages 8

Ultimately, any model of speech segmentation must be able to deal with the high degree of variation that characterizes natural fluent speech. Our earlier work, as reported above (Allen & Christiansen, 1996; Christiansen, 1998; Christiansen et al., 1998) has established that SRNs constitute viable models of early speech segmentation. These models, like most other recent computational models of speech segmentation (e.g. Aslin et al., 1996; Brent, 1999; Brent & Cartwright, 1996; Perruchet & Vinter, 1998), were provided with idealized input. This is in part due to the use of corpora in which every instance of a word always has the same form (i.e. it is a so-called citation form). While this is a useful idealization, it abstracts away from the considerable variation in the speech input that a child is faced with in language acquisition. We therefore now present simulations involving a phonetically transcribed speech corpus that encoded the contextual variation of a word, more closely approximating natural speech. More specifically, we gleaned the adult utterances from the Carterette & Jones (1974) corpus-a part of the CHILDES database (MacWhinney, 2000). These utterances consist of informal speech among American college-aged adults.7