One of the tasks of the language acquisition process
is to optimize strategies for comprehension. For speech perception,
this means that the learner has to establish an accurate mapping
from acoustically detailed input to discrete phonological categories.
As an example, this paper considers the development of the perception
of the English vowels /I/ and /i/ in native speakers.

Production-wise, the two vowels differ in various respects.
Here, we limit ourselves to considering duration and the first formant (F1).
It turns out that the use of these two acoustic dimensions in production
depends on the dialect at hand:
for Scottish English speakers, /I/ and /i/ differ much more in F1 and
much less in duration than for Southern English speakers.

We hypothesize that humans have an optimal perception strategy
that minimizes the probability of confusion and that
there is a knowledge that underlies the implementation of this strategy.
We model the knowledge behind speech perception as
an Optimality-Theoretic perception grammar, and we model the acquisition
of this knowledge with the Gradual Learning Algorithm.
Using an environment based on real production data, we simulate
the development of a Scottish and a Southern English listener,
and show that the Scot comes to rely almost exclusively on height (F1)
when distinguishing /I/ and /i/, whereas the Southerner comes to rely on
both height and duration, so the model indeed implements an optimal strategy
for acoustic cue integration. Perception experiments show that real Scots and
real Southerners also use this optimal strategy in their own environments.

We find, therefore, that perceptual strategies depend on the
production environment, and that we can successfully model this dependency
within the framework of stochastic Optimality Theory, thus bringing
speech-processing systems within the reach of formal phonological theory.