754.10
ELICITING ADULT PLASTICITY: BOTH ADAPTIVE AND NON-ADAPTIVE TRAINING
IMPROVES JAPANESE ADULTS' IDENTIFICATION OF ENGLISH /R/ AND /L/.
B.D. McCandliss1, J.A. Fiez1, M. Conway1,
A. Protopapas3 and J. L. McClelland2.
Departments of Psychology & Center for the Neural Basis of
Cognition, 1University of Pittsburgh, Pittsburgh, PA, 15260
and 2Carnegie Mellon University; and 3Scientific
Learning Corp.
Merzenich et al. (Science, 1996, 271:77-81) describe a training
procedure that enhances the phoneme discrimination abilities of children
who fail to learn from natural speech input. We tested (a) if a similar
approach could be used to teach Japanese adults to discriminate English /r/
and /l/, and (b) if success depends on adaptively exaggerating the relevant
differences between /r/ and /l/ stimuli (e.g., road-load) based on
performance. We used a speech resynthesis technique to create several
multi-step continua ranging from stimuli beginning with exaggerated /r/ to
exaggerated /l/. In the fixed training condition, subjects attempted to
categorize two fixed examples selected to be reliably identified as
beginning with /r/ vs. /l/ by U.S. natives, yet difficult for these
Japanese subjects to identify during an initial test (50-70% correct). For
the adaptive condition, subjects began training with an exaggerated /r/
example and an exaggerated /l/ example, and the degree of exaggeration was
reduced after 8 correct trials or increased after each error. The adaptive
training procedure produced dramatic improvements across 3 twenty-minute
training sessions that generalized partially to another resynthesized
continuum (e.g. rock-lock) based on speech samples from the same speaker.
However, we also found comparable improvement in the fixed condition. A
variant of Kohonen's self-organizing map model (Thomas & McClelland, Soc.
Neur. Abs., 1997:97.2) accounts for Japanese adults' initial failure to
learn from natural speech, while also providing a mechanism for learning
the discrimination when two contrasting stimuli are repeatedly presented in
isolation within either the adaptive or fixed training condition.
This work was supported by NSF LIS 97238.