In Merge, Norris, McQueen, and Cutler have proposed a somewhat different architecture than has usually been suggested in models of word recognition. In particular, they have split the phonemic level of processing into two separate parts, the "input phonemic" and "phoneme decision" processes. A consequence of this approach is that most of the results in the literature that show lexical effects on phonemic processing can be attributed to the action of the "phoneme decision" mechanism. I will discuss this architecture, and report on the results of experiments that bear on the model.
If one accepts the division of the phonemic level into these two components, then as Norris et al point out, studies showing lexical influences on the identification of phonemes cannot unambiguously be attributed to an effect on encoding the phonemic information. Such a top-down attribution can only be shown through tasks in which the lexical information influences phonemic processing in a manner that does not require the listener to label the phoneme itself.
One such demonstration has been provided by Samuel [Cognitive Psychology, 1997, 32, 97-127]. In that study, lexical context led to the phonemic restoration of a particular phoneme, and the resulting phoneme altered the labeling of other phonemes, through selective adaptation. I will report on a set of new selective adaptation experiments that provide further evidence that lexical activation can affect the activation of phonemes. The new experiments use a "Ganong" type of lexical influence to push the perception of a phonemically-neutral sound toward one lexical item or another, which in turn causes the sound itself to be perceived as one phoneme or another. For example, a fricative noise that is perceptually midway between /s/ and /S/ will be heard as /s/ when placed at the end of "bronchiti", but as /S/ at the end of "aboli". Selective adaptation with words made with this sort of concatenation produces changes in how listeners hear test syllables from an "iss"-"ish" continuum. Appropriate control conditions rule out artifactual explanations for these results. The data thus require top-down lexical-phonemic connections, suggesting that Merges architecture must be modified.