The Emergence of Phonology from the Interplay of Speech Comprehension and Production: A Distributed Connectionist Approach

How do infants learn to understand and produce spoken language? Despite decades of intensive investigation, the answer to this question remains largely a mystery. This is, in part, because, although the use of language seems straightforward to adult native speakers, speech recognition and production present the infant with numerous difficult computational problems (see Lively, Pisoni, & Goldinger, 1994). First of all, processing speech is difficult both because it is extended in time and because it is subject to considerable variability across speakers and contexts. Moreover, even with an accurate representation of the underlying phonetic content of a heard utterance, mapping this representation onto its meaning is difficult because the relationship of spoken words to their meanings is essentially arbitrary. On the output side, the infant must learn to produce comprehensible speech in the absence of any direct feedback from caretakers or the environment as to what articulatory movements are required to produce particular sound patterns. Finally, the processes of learning to understand speech and learning to produce it must be closely related (although certainly not synchronized; Benedict, 1979) to ensure that they eventually settle on mutually consistent solutions. In the current work, we formulate a general framework for understanding how the infant surmounts these challenges, and we present a computational simulation of the framework that learns to understand and produce spoken words in the absence of explicit articulatory feedback. Our initial focus is on addressing the relevant computational issues; we postpone consideration of how the approach accounts for specific empirical phenomena until the General Discussion. The framework, depicted in abstract form in Figure 1, is based on connectionist/parallel distributed