A fundamental question in first language acquisition is the "word segmentation problem" -- how it is that infants find words in running speech with no direct instruction. However, most computational models of word segmentation assume unrealistic degrees of invariance in the input provided to infants.
I have extended one such model of word segmentation to handle input automatically derived from speech, more closely approximating the auditory input available to infants. This extended model is robust to the subsegmental variability found in typical speech, but degrades in performance when given highly-variable input. This is consistent with findings that infants' abilities to perform word segmentation degrade in noisy environments. However, by focusing on the clearest parts of the signal, the model improves its performance significantly.
Anton Rytting is a recent graduate from the Ohio State University Linguistics Department, where he studied computational linguistics and automatic speech processing with Chris Brew and Eric Fosler-Lussier. His interests include language acquisition, psycho-computational models of language processing, and perception of suprasegmentals (such as stress and tone). Languages of interest include Greek, Arabic, and Vietnamese.
This talk is part of the CLIP Colloquium Series, organized by Jimmy Lin (jimmylin -at- umd .dot. edu). For the complete schedule, please visit http://www.umiacs.umd.edu/research/CLIP/colloq/.