Lab Meeting: Ewan Dunbar (Context in phonological development: Computational approaches)

The Polinsky Lab meeting Wednesday 2/12/14, 5:15 pm, will be:
 
Context in phonological development: Computational approaches
 
presented by: Ewan Dunbar
                     ( Laboratoire de Sciences Cognitives et Psycholinguistique, ENS / EHESS / CNRS )
 
Location: 2 Arrow Street, 4th floor conference room
 
See below for relevant background readings.
 
ABSTRACT:
Early cognitive development in the systems implied in perceiving and producing speech (phonological processing) is well-studied experimentally and observationally.  Recently, statistical machine learning tools have begun to address three fundamental problems in phonological development: perceptual learning, morpheme segmentation, and lexical organization. In this talk, I address an issue that cross-cuts all three problems: contextual adaptation. I focus primarily on the perceptual learning problem. Using Bayesian models of category learning, I argue against a widespread assumption in developmental psycholinguistics, in which early speech perceptual development largely consists of the formation of context-independent segmental categories, while sensitivity to coarticulatory and allophonic effects of speech context must wait for a later stage of more abstract development. I propose a new model, in which phonetic units are inherently context-dependent. The model is analogous to standard approaches to talker normalization in speech recognition, has support from the cross-linguistic typology of allophony and coarticulation, and is grossly consistent with the behavioral evidence about the role of phonetic context in speech perception. I also briefly discuss an approach to lexicon organization and word segmentation in which the ideal learner is characterized as adhering to a measure of overall lexicon "coherence", by which top-down effects of semantic context can influence phonological development. Finally, I outline the very recent emergence of collaboration and methodological convergence between developmental psycholinguistic modeling and unsupervised or low-resource automatic speech recognition (ASR) research, where models must be trained without any phoneme labels. This engineering problem corresponds perfectly to the learning problem the infant faces. I argue that, in ASR, too, handling contextual variability is a sine qua non, but that standard approaches are difficult or impossible to recreate in the unsupervised setting; effective methods for contextual adaptation in the unsupervised setting are thus a defining issue for the coming decades of ASR research. I suggest a few ways in which domain knowledge from speech sciences may be able to help us build better models both for ASR and for answering developmental questions.
 
BACKGROUND READINGS:
(1)  [LINKDillon, Dunbar, Idsardi. A Single-Stage Approach to Learning Phonological Categories: Insights From Inuktitut
(2)  [LINKJansen, Dunbar, et. al. A SUMMARY OF THE 2012 JHU CLSP WORKSHOP ON ZERO RESOURCE SPEECH TECHNOLOGIES AND MODELS OF EARLY LANGUAGE ACQUISITION
        
        
Location: 
2 Arrow Street, 4th floor conference room
Date: 
Wednesday, February 12, 2014 - 5:15pm to 7:00pm
Event category: