15th International Congress of Phonetic Sciences (ICPhS-15)
This paper illustrates how a computational linguistic model of phonology, incorporating statistical phonological/phonetic information, implemented in a multi-agent architecture, can contribute to robustness in speech technology applications. This work represents a novel extension to the Time Map model (Carson-Berndsen, 1998) retaining the main attributes of the model but recasting the architecture in terms of a multi-agent system. The Time Map model employs finite state methodology and an event logic to demonstrate how phonotactic constraints can play a role in speech recognition. These constraints model the permissible combinations of sounds for a language, and are used to interpret multilinear representations of speech utterances. Parsing takes place at both the phonological and phonetic feature levels, differing from mainstream techniques, in that the approach is non-segmental. Recasting the model in terms of a multi-agent system facilitates distributed parallel processing, constraint relaxation and output extrapolation in the case of underspecified utterances, and more elegant parsing.
Bibliographic reference. Walsh, Michael (2003): "Recasting the time map model as a multi-agent system", In ICPhS-15, 735-738.