Third International Conference on Spoken Language Processing (ICSLP 94)

Yokohama, Japan
September 18-22, 1994

Phonetic Prototypes: Modelling the Effects of Speaking Rate on the Internal Structure of a Voiceless Category Using Recurrent Neural Networks

Mukhlis Abu-Bakar (1), Nick Chater (2)

(1) Departments of Linguistics & Psychology, University of Wales, Bangor, Gwynedd, UK
(2) Department of Psychology, University of Edinburgh, Edinburgh, UK

This is the last in a series of studies that attempts to account for rate effects in phonetic perception using a recurrent neural network. We present new findings that extend this research, emphasising particularly the effects of syllable duration on the internal structure of stimuli representing the voiceless /p/ category. In a series of simulations, network performance reveals a systematic rate effect that closely parallels that found in studies of human categorization. First, the internal structure of the voiceless category undergoes extensive alteration with changes in syllable duration. Second, in a selective adaptation type of experiment, the maximally effective adaptor shifts toward a longer VOT as syllable duration increases. The performance can be traced to the network's sensitivity to the training frequency of the individual stimulus, consistent with exemplar-based accounts of human category formation.

Full Paper

Bibliographic reference.  Abu-Bakar, Mukhlis / Chater, Nick (1994): "Phonetic prototypes: modelling the effects of speaking rate on the internal structure of a voiceless category using recurrent neural networks", In ICSLP-1994, 511-514.