Moreton, Elliott (2004). Diachronically inaccessible grammars: a
diachronic-phonetic study of the English /ai/
alternations. Presentation at the workshop on "Redefining Elicitation:
Novel Data in Phonological Theory", New York University, April 9-11,
2004).  (Work done in collaboration with Erik R. Thomas, North Carolina
State University.)

Short abstract:

     Factorial typology sometimes overpredicts: Constraints needed to
model a real language can be re-ranked to yield an unattested one.
This is ordinarily taken as a sign that the constraint set has been
misapprehended.  Myers (2003), however, has proposed that some
phonologies permitted by Universal Grammar are diachronically
improbable or impossible.  This study presents evidence of a
systematic typological gap that is due to diachronic rather than
synchronic factors, and proposes a means of distinguishing diachronic
from synchronic gaps.

Details:

     The present study (a collaboration with Erik R. Thomas of North
Carolina State University) is a diachronic phonetic investigation of
such a case: the allophonic alternations of /ai/ conditioned by coda
voicing that are found in many dialects of English.  Three basic
patterns occur, in which (e.g.)  "ice" and "eyes" are pronounced [^Is
aIz], [aIs aEz], or [aIs az].  These patterns have been independently
re-innovated multiple times.  However, the constraints needed for them
also predict unattested patterns such as *[as aIz].

     Synchronically, English diphthongs show enhanced offglides before
voiceless codas and enhanced nuclei before voiced ones (Thomas 2000,
Moreton in press).  The attested /ai/ alternations appear to be
phonologizations of that phenomenon.  This view is corroborated by an
instrumental study of the re-innovation of the [^Is aIz] allophony in
Cuyahoga County, Ohio, based on recordings of 22 speakers born between
1880 and 1977.

     It follows that *[as aIz], though possible as far as UG goes, is
diachronically inaccessible: No potential immediate ancestor (i.e.,
minimally different phonological pattern) would have a phonetic
tendency to change into it.  For example, [aIs aIz] would not lose the
pre-voiceless offglide while keeping the pre-voiced one.

     These findings suggest that the set of possible phonologies is
smaller, in systematic ways, than factorial typology predicts, and
hence that a theory of Universal Grammar will be incomplete as a
theory of typology unless augmented by a theory of language
transmission.