Multiple lexical representations overlapping with the input (cohort neighbors) are temporarily activated in the listener's mental lexicon when speech unfolds in time. Activation for cohort neighbors appears to rapidly decline as soon as there is mismatch with the input. However, it is a matter of debate whether or not they are completely excluded from further processing. We recorded behavioral data and event-related brain potentials (ERPs) in auditory-visual word onset priming during a lexical decision task. As primes we used the first two syllables of spoken German words. In a carrier word condition, the primes were extracted from spoken versions of the target words (ano-ANORAK "anorak"). In a cohort neighbor condition, the primes were taken from words that overlap with the target word up to the second nucleus (ana-taken from ANANAS "pineapple"). Relative to a control condition, where primes and targets were unrelated, lexical decision responses for cohort neighbors were delayed. This reveals that cohort neighbors are disfavored by the decision processes at the behavioral front end. In contrast, left-anterior ERPs reflected long-lasting facilitated processing of cohort neighbors. We interpret these results as evidence for extended parallel processing of cohort neighbors. That is, in parallel to the preparation and elicitation of delayed lexical decision responses to cohort neighbors, aspects of the processing system appear to keep track of those less efficient word candidates.
Keywords: spoken word recognition, word onset priming, ERPs
INTRODUCTIONCurrent theories of speech perception assume that incoming acoustic information activates a set of word candidates or a cohort of words in the mental lexicon that overlap with this input (c.f. McClelland and Elman, 1986;Zwitserlood, 1989;Norris, 1994;Gaskell and Marslen-Wilson, 1997;Norris and McQueen, 2008). If speech was always clear-cut and intelligible and if the phonological parser aimed at processing speech as parsimoniously as possible, it would be ideal to activate only those cohort neighbors that match the acoustic input in all features and remove items from the set of activated word candidates as soon as any mismatching information is available. However, in real life the speech signal is often noisy and degraded and there is high variability within and between talkers. This will often hinder a clear-cut decision for or against a proper lexical candidate. Here we propose that this dilemma might be approached by a twofold recognition strategy. Some processing components of the speech recognition system might handle partially mismatching lexical candidates, while others might more effectively rule out co-activated alternatives in order to end up with a single percept.Classical models of spoken word recognition differently handle co-activated lexical entries. Connectionist models, such as TRACE (McClelland and Elman, 1986) and Shortlist (Norris, 1994) incorporate inhibitory connections between lexical representations. Activated candidates inhibit each other as...