The Interplay of Perception and Production in Phonological Development: Beginnings of a Connectionist Model Trained on Real Speech

Three forward models are presented that map articulatory positions onto acoustic outputs for a single speaker of the MOCHA speech database. Backpropagation learning was used to train the forward models on a database of 460 TIMIT sentences. Efficacy of the trained models was assessed by subjecting the model outputs to speech intelligibility tests. The results of these tests showed that enough phonetic information was captured by the models to support fairly high rates of word identification in sentences. These forward models provide the first step toward building a connectionist model of spoken word acquisition trained on real speech. The design of this model is based on a theory of phonological development in which distributed codes are learned in the service of spoken word perception, production, and comprehension.