260 Speech Production and Perception
encapsulated from such feedback; however, when the proces-
sor yields an ambiguous output, lexical knowledge is brought
to bear to resolve the ambiguity. In the first account, the effect
of the lexicon is on perceptual processing; in the second it is
on processing that follows perception of phones. The Ganong
paradigm has been used many times in creative attempts to
distinguish these interpretations (e.g., Fox, 1984; Miller &
Dexter, 1988; Newman, Sawusch, & Luce, 1997). However,
it remains unresolved.
A second finding of lexical effects is phonemic restoration
(e.g., Samuel, 1981, 1996; Warren, 1970). When the acoustic
consequences of a phoneme are excised from a word (in
Warren’s classic example, the /s/ noise of legislature) and are
replaced with noise that would mask the acoustic signal if it
were present, listeners report hearing the missing phoneme
and mislocate the noise. Samuel (1981) showed that when
two versions of these words are created, one in which the
acoustic consequences are present in the noise and one in
which they are absent, listeners asked to make a judgment
whether the phone is present or absent in the noise show
lower perceptual sensitivity to phones in words than in non-
words. That the effect occurs on the measure of perceptual
sensitivity (d) suggests that, here, lexical knowledge is ex-
erting its effect on phoneme perception itself. (However, that
dcan be so interpreted in word recognition experiments has
been challenged; see Norris, 1995.)
A final lexical effect occurs in experiments on compensa-
tion for coarticulation. Mann and Repp (1981) found compen-
sation for /s/ and /ʃ/ on members of a /ta/-to-/ka/ continuum
such that the more front /s/ fostered /ka/ responses, and the
more back /ʃ/ fostered /ta/ responses. Elman and McClelland
(1988) used compensation for coarticulation in a study that
seemingly demonstrated lexical feedback on perceptual pro-
cessing of consonants. They generated continua ranging from
/d/ to /g/ (e.g.,datestogates) and from /t/ to /k/ (e.g.,tapesto
capes). Continuum members followed words such asChrist-
masandSpanishin which the final fricatives of each word (or,
in another experiment, the entire final syllables) were re-
placed with the same ambiguous sound. Accordingly, the only
thing that made the final fricative ofChristmasan /s/ was the
listeners’ knowledge thatChristmasis a word andChristmash
is not. Lexical knowledge, too, was all that made the final
fricative ofSpanishan /ʃ/. Listeners showed compensation
for coarticulation appropriate for the lexically specified frica-
tives of the precursor words.
This result is ascribed to feedback effects on perception,
because compensation for coarticulation is quite evidently an
effect that occurs during perceptual processing of phones.
However, Pitt and McQueen (1998) challenged the feedback
interpretation with findings appearing to show that the effect
is not really lexical. It is an effect of listeners’ knowledge of
the relative frequencies of phone sequences in the language,
an effect that they identify as prelexical and at the same level
of processing as that on which phonemes are perceived. Pitt
and McQueen note that in English, /s/ is more likely to fol-
low the final vowel ofChristmasthan is /ʃ/, and /ʃ/ is more
common than /s/ following the final vowel ofSpanish. (If
readers find these vowels—ostensibly /ə/ and /I/ according to
Pitt and McQueen—rather subtly distinct, they are quite
right.) These investigators directly pitted lexical identity
against phone sequence frequency and found compensation
for coarticulation fostered only by the transition probability
variable. Lately, however, Samuel (2000) reports finding a
true lexical effect on phoneme perception. The clear result is
that lexical knowledge affects how we identify consonants
and vowels. It is less clear where in processing the lexical
effect comes in.
Pitt and McQueen’s study introduces another knowledge
variable that can affect phone identification: knowledge of
the relative transition frequencies between phones. Although
this logically could be another manifestation of our lexical
knowledge, Pitt and McQueen’s findings suggest that it is
not, because lexical and transition-probability variables dis-
sociate in their effects on compensation for coarticulation. A
conclusion that transition probability effects arise prelexi-
cally is reinforced by recent findings of Vitevitch and Luce
(1998, 1999).
There are many models of spoken-word recognition. They
include the pioneering TRACE (McClelland & Elman,
1986), Marslen-Wilson’s (e.g., 1987) cohort model, the
neighborhood activation model (NAM; Luce, 1986; Luce &
Pisoni, 1998), the fuzzy logical model of perception (FLMP;
e.g., Massaro, 1987, 1998), and shortlist (e.g., Norris, 1994).
(A more recent model of Norris et al., 1999, Merge, is
currently a model of phoneme identification; it is not a full-
fledged model of word recognition.)
I will describe just two models, TRACE and a nameless
recurrent network model described by Norris (1993); these
models represent extremes along the dimension of interactive
versus feedforward only (autonomous) models.
In TRACE, acoustic signals are mapped onto phonetic fea-
tures, features map to phonemes, and phonemes to words.
Features activated by acoustic information feed activation
forward to the phonemes to which they are linked. Phonemes
activate words that include them. Activation also feeds back
from the word level to the phoneme level and from the
phoneme level to the feature level. It is this feedback that
identifies TRACE as an interactive model. In the model, there
is also lateral inhibition; forms at a given level inhibit forms
at the same level with which they are incompatible. Lexical