What's the best evidence against a position like that expounded by e.g. Bob Port (or Ken Lodge, for you UK-based phon*ists), which essentially states that phonology (or whatever you want to call the cognitive/internal aspect of the acoustic/auditory/articulatory bits of language) is essentially non-segmental, and that segmental phenomena only emerge post-literacy training?
It depends what you will count as evidence, and which position you want to argue against. Internal evidence would show that aspects of a phonological system cannot be described perspicaciously without recourse to segments, while external evidence would involve arguments based on the performance of speakers to the effect that their behavior could not be explained without positing segments in a mental representation. There are two possible positions to argue against, (a) that not all of phonology is organized into segments, or (b) that none of phonology is organized as segments. Obviously (b) will be easier to argue against.
The strong segmentalist view propounded by Chomsky in the 1960's and still by Halle is that phonological representations are matrices, where each column represents a segment. Arguments about the autosegmental nature of tone and accent have made this position difficult to adhere to. The more reasonable segmentalist position today is that phonological representations are made of timing nodes which are linked to different universal features. The timing slots correspond to segments. The argument made by Lodge (in his 2009 book "Fundamental concepts in phonology") is that the features are not universal, and can therefore be assigned an arbitrary phonetic interpretation. He does not, however, argue that phonological representations lack any sort of timing tier. His model of phonology still requires timing nodes, but varies from generativist work in rejecting the universality of features. Silverman (in his 2006 book "A Critical Introduction to Phonology") brings up external evidence against features, citing a well-known 1979 Study by Morais et al (Cognition 7:(4) 323-331) showing that Portuguese illiterates were unable to perform a task requiring them to add or subtract a segment to/from the beginning of a word. As far as I can tell, there is no opponent of segments who would argue that phonological words are always stored, perceived and produced as indivisible wholes, but this would be a good artificial endpoint to set in evaluating different proposals.
As far as I know, there is no detailed criticism of Lodge's proposal, but I think that a good argument against it on internal grounds would be to show that phonetic interpretation of features must be highly constrained and, at some level, language-independent.
Concerning external evidence, a type of evidence frequently reported in favor of segmentalism is evidence from speech errors, Idsardi (2010), for example,suggests in his review of Silverman's book, refers to a study on speech errors by Wan & Jaeger (1998). On the other hand, speech errors are cited by Articulatory Phonologists as evidence in favor of the gestural organization of speech. A 1992 paper by John Ohala argues for segments on the basis of a thought experiment.
In sum, I think a preliminary to this task is to try to clarify the positions of the antisegmentalists (admittedly not an easy task) and determine exactly at what points these vary from those of segmentalists. It tends to be the case that many arguments put forward on either side are against positions which are not held by anyone, and defenses against such arguments are defenses of positions which were never attacked.
Many opponents of a rule-based approach to language-learning support an exemplar-based approach instead. In these models, the language learner does not form an internal grammar by learning abstract rules that manipulate segments; rather, she stores vast "clouds" of tokens that are committed to long-term memory on a word-by-word basis. Every time we encounter a new token, we compare it with every similar token we've heard before and adjust our mental representation of that token accordingly. It is important to point out that Port et al do not deny the existence of segment-sized units as a cultural construct, or even as part of individual impressions of speech:
So I certainly would not say that our impression that speech has segment-sized units is due solely to alphabet training. Languages do have some segment-sized units, along with many other sized units as well. What is true, however, is that alphabet skills do tend to highlight the segment-like (i.e., letter-like) properties of phonologies for us. (Port, Languages and Sciences 32, p61-62)
That being said, they contend that what we phonologists have been doing for decades is codifying sets of cultural conventions shared among communities of speakers rather than any kind of internal grammars.
I am familiar with three main types of argument against the extreme exemplar view (i.e. the complete abandonment of rule-based approaches to grammar):
- The demands on long-term memory that the pure exemplar view entails are quite extreme.
- The fact that language learners overgeneralize when attempting to produce new forms is evidence that those learners are applying abstract rules.
- The existence of Neogrammarian (i.e. across-the-board) sound changes are very difficult to capture in a model that doesn't allow for abstract units and categories.
Meanwhile, exemplarists cite experimentally confirmed frequency effects and "neighborhood density" (i.e. degree of proximity to similar sounds) effects on production as evidence for exemplars and claim that purely abstractionist approaches cannot explain such effects. Recently, some researchers have supported a middle-ground approach in which insights from both camps are integrated into a model that allows for both the storage of phonetic detail and the manipulation and categorization of abstract units in the internal grammar.