Kingston, John, and Elliott Moreton (2001).  How do listeners learn
foreign vowel categories: as disjunctive sets, through selective
attention, or as prototypes?  In: R. Smits, J. Kingston, T. M. Nearey,
and R. Zondervan (eds.), Proceedings of the Workshop on Speech
Recognition as Pattern Classification, Nijmegen, The Netherlands, July
11-13, 2001.

In exemplar models, category learning is a secondary process that
depends on which dimensions of the observers' attention is selectively
focused on.  We tested adult American English speakers' ability to
categorize the nonlow rounded vowels of German, which contrast for the
features [high], [tense], and [back].  Earlier work showed that high
speaker or context variability impaired learning, but didn't test
variability across the different phonemes belonging to the natural
class defined by a distinctive feature value.  Here, the number of
vowels in each natural class was manipulated in the training stimuli.
If each class is a disjunctive set of exemplars, more traiing vowels
should, like more speakers or contexts, impair generalization to a new
vowel pair.  if learning a natural class is instead discovering and
attending selectively to the class's defining properties, more traiing
vowels should improved generalization.  Finally, if learners abstract
a feature-based prototype, learning the class is discovering the
relevant feature, and more training vowels whould improve learning
only so longa s they narrow down the choice between the class's
defining feature value vs. values that vary within the class.  Results
show that listeners learned feature-based prototypes for [high] and
[back], and did not distinguish between featur-based prototypes and
attention-weighted exemplars for [tense].