The emergence of phonology from the interplay of speech comprehension and production ; A distributed connectionist approach

How do infantslearnto understandandproducespoken language?Despitedecadesof intensi ve investigation,the answerto this questionremainslargely a mystery. This is, in part, because, althoughthe use of languageseemsstraightforward to adultnative speakers,speechrecognitionandproductionpresentthe infantwith numerousdifficult computational problems(seeLively, Pisoni,& Goldinger , 1994).First of all, processingspeechis difficult bothbecauseit is extendedin time andbecauseit is subjectto considerable variability acrossspeakersandcontexts. Moreover, even with an accuraterepresentationof the underlying phoneticcontentof a heardutterance,mappingthis representationontoits meaningis difficult because the relationshipof spoken wordsto their meaningsis essentiallyarbitrary. On the outputside,the infant mustlearnto producecomprehensiblespeechin theabsenceof any directfeedbackfrom caretakersor the environmentasto whatarticulatorymovementsarerequiredto produceparticularsoundpatterns.Finally, theprocessesof learningto understandspeechandlearningto produceit mustbecloselyrelated(although certainly not synchronized;Benedict,1979) to ensurethat they eventually settleon mutually consistent solutions. In the currentwork, we formulatea generalframework for understandinghow the infant surmounts thesechallenges, andwepresentacomputational simulationof theframework thatlearnsto understandand producespoken wordsin the absenceof explicit articulatoryfeedback.Our initial focusis on addressing the relevant computationalissues;we postponeconsiderationof how the approachaccountsfor specific empiricalphenomenauntil theGeneralDiscussion. The framework, depictedin abstractform in Figure 1, is basedon connectionist/parall e distributed