In connected speech, consonants can adopt the place of articulation of neighbouring consonants. This happens word-internally and at word boundaries. The latter situation, in particular, creates a challenge for word recognition. There are a number of ways to deal with how surface variation affects word recognition.
One is to treat all variation as 'noise' to the system. Whether such variation is harmful to the system depends on the consequences of a less than perfect fit between input and sublexical and lexical representations. In TRACE, word units are not punished by mismatch with the input. Mismatch is essentially the same as lack of information. In both cases, word units will be activated to a lesser extent, compared to a complete match between input and word unit. Simulations with MERGE show that mismatch affects the behaviour of word nodes. In the Cohort model, mismatch can have negative consequences, depending on the legality of the variation.
The idea that mismatch can be legal or illegal comes from a different perspective to surface variation. This perspective has its roots in phonological theory and focuses on lexical specification. If representations of lexical form (or word units) are fully specified - as they are in TRACE and MERGE - all types of mismatch will have an effect. If lexical representations are underspecified - as in Cohort - changes that reflect phonologically legal variation will not harm word recognition, illegal changes will.
Another way to deal with variation due to assimilation is in terms of a context-sensitive undoing of assimilation. If licensed by the local, segmental context, assimilations can be recovered in some way, such that the altered segment can be mapped onto the original one. Gaskell and Marslen-Wilson (1996, 1998) describe such a process in terms of phonological inference. Their recent position, couched in terms of a network model, is that we need abstractness at the level of representation as well as phonological inference processes.
Evidence for abstractness and phonological inference mainly comes from English. We present crossmodal priming data from German, with assimilations in terms of place of articulation and voice, in regressive and progressive directions. We contrasted phonologically legal and illegal assimilations, in isolation and in context, where the context could be viable or unviable. First, the data show that effects depend on the availability of context. With one exception, no reliable effects are found for altered stimuli presented in isolation. With context, both contextual viability and phonological legality have an impact. We discuss implications of our findings for the approaches to word recognition and lexical representation mentioned above.