Interactions between phonological and articulatory processes in speech 
production: Evidence from speech errors

Matt Goldrick, Northwestern University

 	Recent findings suggest that multiple lexical representations (not just 
the selected target word) activate their phonological representations (e.g., 
Peterson & Savoy, 1998). In spreading-activation theories of speech production 
(e.g., Dell, 1986), this type of interaction is referred to as cascading 
activation between lexical selection and phonological planning processes.  We 
examined whether cascading activation is also found between phonological 
planning and articulatory processes.  When a target segment is replaced in an 
error, it should remain partially active.  Cascading activation from this 
segment can then influence the realization of the error segment.  For example, 
in the error "tab"->"dab," the partially active phonological representation /t/ 
may send cascading activation to articulatory processes.  This cascade may 
cause the acoustic / articulatory properties of the [d] in "dab" to reflect a 
"trace" of the target [t] (e.g., [d] may be produced with a longer voice onset 
time).
 	Method.  Eighty tongue twisters were generated using syllables
made up of four initial obstruents (/k, g, t, d/), five vowels (/i, E, AU,
aI, Oi/), and four final fricatives (/f, v, s, z/; e.g., "keff geff geff
keff"). Five English speakers read each twister aloud 3 times quickly to
induce speech errors.
 	Results.  Productions resulting in perceptible errors (e.g.,
"keff->geff") were paired within each speaker with correct productions
matched to the error (e.g., "geff->geff").  The voice onset time (VOT)
distribution of error tokens reflected a trace of the target segment.
Voiced -> voiceless tokens exhibited a shorter mean VOT (68 msec) compared
to correctly produced voiceless tokens (76 msec; Wilcoxon signed-rank p <
.02); voiceless -> voiced error tokens (e.g., /k/->[g]) exhibited a mean
VOT (26 msec) that was significantly longer than correctly produced voiced
tokens (e.g., /g/->[g]; 24 msec; p < .002).
 	A post-hoc analysis found that the size of the trace was
influenced by the lexical status of the error.  The mean size of the trace
for nonword errors (e.g., "keff -> geff") was significantly larger (6
msec) than the trace for word errors (e.g., "kess -> guess"; 1 msec; ,
t(58) = 2.1, p < .04).
 	Conclusions.  Error tokens reflected an acoustic trace of the
intended target.  This suggests that partially activated target
phonological representations send cascading activation to articulatory
processes.  The effect of lexical status on errors can be attributed to
cascading activation from lexical selection processes.  This cascade
increases the activation of phonological representations that correspond
to words; this activation advantage then cascades to articulatory
processes, reducing the influence of target representations (and the size
of the trace).

References
Peterson, R. R. & Savoy, P. (1998).  Lexical selection and phonological
encoding during language production: Evidence for cascaded processing.
Journal of Experimental Psychology: Learning, Memory, and Cognition, 24,
539-557.