In A. Sorace, C. Heycock & R. Shillcock (Eds.), Proceedings of GALA 1997: Language Acquisition: Knowledge Representation and Processing (pp. 327-332). University of Edinburgh.


Coping with Variation in Speech Segmentation

Morten H. Christiansen & Joe Allen



Abstract

This paper presents results from word segmentation simulations in which a Simple Recurrent Network (SRN) was exposed to speech input incorporating high degrees of variation. In Experiment 1, a network trained on a speech corpus transcribed to include variation in terms of coarticulation was compared with a network trained on a citation form version of the same corpus. The results show that the network accommodates this variation without significant impairment to its performance on the segmentation task. Experiment 2 involved a novel approach to the modeling of segmental variation in which feature values were systematically varied according to a predetermined probability schedule. Results demonstrate that following training the networks were able to withstand a very high degree of segmental variation within words and still able to locate word boundaries in the input. Together the experiments indicate that the SRN provides a robust mechanism for the modeling of early speech segmentation.


Click here to download the full paper.

rule.gif (155 bytes)

Home | People | Research | Links | Contact | Publications | Presentations
Cognitive Neuroscience Laboratory

Please email suggestions/errors to mhc27@cornell.edu