260 Gitte Kristiansen
Labov’s research on a single individual’s productions of vowel tokens (La-
bov 1994, inter alia) demonstrates that individual exemplars of one pho-
neme will be included in the phonetic range of another phoneme: for exam-
ple, some exemplars of /æ/ will occur in the range of exemplars of /ɛ/. How
does a speaker know that those tokens are exemplars of /æ/ and not /ɛ/?
This question cannot be answered in a purely segment-based approach to
phonological representation. If one begins with segments, one must have a
definition of those segments that is either ultimately phonetic, or else purely
arbitrary (i.e., a particular exemplar is stipulated to be an exemplar of /æ/
even if its actual realization is [ɛ] in purely phonetic terms). On the other
hand, if one begins with words as phonological units, then the question can
be answered and the paradox is solved. The phonetically outlying token is
an exemplar of /æ/ because it is part of a specific word, and other occur-
rences of that word contain exemplars that cluster around the central pho-
netic tendency for /æ/. How is the word identified as the same word? The
word is of course identified as the same by its meaning in the context of
use, linked to prior occurrences of the word with that meaning in similar
contexts of use.Vihman and Croft thus argue in favor of an approach to phonological pro-
totype categories (cf. exemplars that cluster around the central phonetic
tendency of /æ/) according to which phonological classification is deter-
mined by units, or constructions, larger than its components. In the case of
lects, however, I would like to argue that the lexicon facilitates disambigua-
tion, but fails to determine it. To illustrate the difference, imagine the fol-
lowing situation:
On a university campus you catch a crowded bus and notice a group of
students engaging in lively conversation at the rear end. You cannot really
hear what they say because of the background noise, but you capture
sounds and tones that tentatively identify the group of students as speakers
of, say, Icelandic or Dutch (or any other language you are familiar with). It
is only when you reach the rear end of the bus to get out at your destination
that you realize that the students in question are not speaking in their moth-
er tongue but in the native language of the country you are in (say, England
or Spain). Like the children in experiment 2 you were able to identify the
L1 language of the speakers, but in order to do so you did not rely on struc-
tural constructions at the level of words or clauses. The link went directly
from a stored linguistic stereotype to the unclassified speech tokens that
you heard.
If we can agree that this is not a far-fetched example, but rather a situa-
tion that many of us have actually experienced, we draw the tentative con-