Third International Conference on Spoken Language Processing (ICSLP 94)

Yokohama, Japan
September 18-22, 1994

Using Prediction to Learn Pre-Linguistic Speech Characteristics: A Connectionist Model

John Nienart, J. Devin McAuley

Computer Science Department, Cognitive Science Program, Indiana University, Bloomington, IN, USA

We describe initial results in the development of a connectionist model of pre-linguistic categorization in infant language acquisition. We assume that early listeners are sensitized to salient contour properties of natural sounds. To model this learning process, we trained a recurrent connectionist network to predict its future inputs, under the assumption that rising or falling sounds generate an expectation in the hearer that they will continue in the given direction. The training and testing sets represented rising and falling tone sequences, as found in formant transitions. To learn the task, the network developed attractors for each input and positioned these attractors tonotopically in state space. Principal components analysis of the memory-layer activations showed that this positioning of attractors enabled the network to implicitly categorize sweeps as high or low, and rising or falling.

Full Paper

Bibliographic reference.  Nienart, John / McAuley, J. Devin (1994): "Using prediction to learn pre-linguistic speech characteristics: a connectionist model", In ICSLP-1994, 1715-1718.