CBSU bibliography search
To request a reprint of a CBSU publication, please click here to send us an email (reprints may not be available for all publications)
A dynamic-net model of human speech recognition.
In G.T.M. Altmann (Ed.), Cognitive Models of Speech Processing: Psycholinguistic and Computational Perspectives (pp.87-105). Cambridge, MA: MIT Press.
Year of publication:
Speech is a time-varying signal, but conventional feed-forward networks have no way of handling the temporal component of speech directly. Consequently, such networks tend to be unable to generalise appropriately in the time domain. The present paper demonstrates how this problem can be overcome by the use of a dynamic net with a very simple architecture. Unlike TRACE, for example, the net can identify words in continuous mock-speech input without the need for duplication of word or phoneme nodes. Also, the net is very tolerant of temporal variability in the signal. In contrast to feed-forward nets it generalises well to instances of previously learned patterns presented at different rates. Finally, the strongly left-to-right nature of the dynamic net means that it provides a very natural model of the left-to-right, or 'cohort', behaviour observed in human speech recognition. The interesting psychological characteristics of the net follow automatically from an architecture designed to recognize patterns which unfold over time.