What is really interesting, though, is what it could possibly mean to "learn" a rule like final devoicing...
It means that you learn (since you don't know a priori) that there is in the grammar of German a rule [-sonorant] → [-voice]/__σ]. The basis for learning that is that some stems end with voiced obstruents and others end with voiceless obstruents, but when the voiced obstruents are syllable final (i.e. not followed by a vowel), the obstruent changes voicing. And that is it. That rule could play a role in metalinguistic judgment where you ask a German speaker if they can say [bund] and they say "No", but that doesn't automatically follow from the speaker having learned the rule. It does not entail that German speakers are unable to pronounce words with final voiced obstruents. In the Hale & Reiss "minimalist" approach, grammars are not held to be entirely responsible for all aspects of linguistic behavior: they encourage and even demand independent investigation of metalinguistic knowledge.
I am highly skeptical of the standard view that language learners acquire grammars in the generative sense, i.e. specialized sets of instructions for calculating well-formedness. Rather, I think that the primary goal is performance-oriented and that intuitions of grammaticality are derivative of them. Regardless of maturity, a language learner is always concerned with mastery of linguistic skills. So it is really an open question whether there exists a devoicing rule in the sense that you describe it here. And I think that you have oversimplified the problem. You know as well as I that there is a discrepancy between the sounds that speakers perceive and those that exist superficially in the acoustic signal. And there is a discrepancy between what speakers think they are articulating and the actual articulatory production that they execute. So it is really difficult to interpret what a "no" or "yes" response to the "Can you say..." question really means from a linguistic perspective. Our theories, of course, bias our interpretations. H&R are articulating a standard view of the role of grammars in generative theory, of course--one that has been remarkably resistant to criticism over the past half a century. (I find that depressing, but understandable.)
Let me reiterate the basic dichotomy that I think you are missing, albeit for ideological reasons. The language learner faces two very different questions in learning a new phonological system:
- What sounds does the speaker intend to pronounce?
- How is the speaker articulating those sounds?
The first question is the basis for phonological representation. The second is the basis for motor coordination of the speech tract. These are just two separate cognitive tasks that every language learner has to master. Intuitions of well-formedness--what the speaker
ought to be trying to say--are of secondary importance. Well-formedness can be seen as a sense of perfection of performance, just as we have a sense of perfection in other forms of behavior. That is, well-formedness intuitions are not special to linguistic behavior. They are a generalized cognitive function that arise for every form of coordinated, intentional behavior.
Yet modern generative phonology still treats both types of rules as essentially "phonology".
That is correct, and that is the way it should be...
We disagree quite strongly on that, but we agree on what the standard view is within the generative paradigm.
...There are many, many ways to subclassify phonological rules in terms of speaker behavior with respect to the rule, or even in terms of formal properties of rules (such as "rules that operate between adjacent segments" or "rules that operate between non-adjacent segments"). In the latter case there is an actually interesting question about the nature of phonological computations, since we need to find the proper means of saying "applies between non-adjacent segments". That's because stating that condition correctly is, by definition, what it means to have a theory of phonological computations. Knowing how a given rule relates to second language acquisition or speech errors is not part of the theory of grammatical computations, it is part of the theory of psycholinguistics (in the broadest sense).
I have no trouble thinking of prosodic groupings as the proper level at which to capture non-adjacency, but generative theory blends phonology with morphology, which I think is a fundamental mistake. Morphology deals with the first question above, and phonology with the second. If you are going to bring up psycholinguistics, let's not forget that generative grammar is a purely psycholinguistic approach to language. Unfortunately, it is a myopic approach in that it has a psycholinguistic theory of linguistic competence but lacks a corresponding theory of linguistic performance. One can't have one without the other, although being a generative linguist seems to commit one to believing that one can.
Linguists look at data and analyze patterns of distribution. Language learners are faced with two very different problems--what sounds to try to pronounce and how to coordinate the articulation of those sounds. It is not just about patterns of distribution.
I disagree on what language learners are faced with. I agree that many linguists take statistical patterns of distribution to be part of the object of study, but there is no justification for putting distributional patterns in a grammar. I am reasonably confident though not absolutely certain that H&R also reject reifying distributional observations in grammar (for instance, MSCs which they do specifically reject). For the purpose of this discussion, we can reject distributional patterns as being relevant to the nature of a grammar.
FTR, I said nothing here about putting distributional patterns in the "grammar". We have a disagreement on what the "grammar" is. If you ever engage in a field study of a language, those patterns become part of the record that you analyze, because you want to be very careful about imposing your own linguistic biases on the subject matter. All I meant to say was that we should be wary of confusing language learning with linguistic analysis.
What is relevant is the resolution of contradiction. A child exposed to (American) English will hear that the name of the stuff filling the lakes is [waɾɹ̩], and they will learn that their dictionaries should include the symbolic sequence /waɾɹ̩/ for H2O. They will parse out [tɹʷɪm] from various instances of that verb root, such as [tɹʷɪm], [tɹʷɪmd], [tɹʷɪmz], [tɹʷɪmɪŋ], and likewise register the root as /tɹʷɪm/. They face a superficial contradiction when they encounter "rot" = [rɑt], [rɑts], [rɑɾɪŋ], [rɑʔn̩], that the root could be any of /rɑt, rɑɾ, rɑʔ/. When they learn that there are rules of glottalization and flapping, the contradiction evaporates, and the lexical thing to be learned is /rɑt/.
I get what you are trying to say here, but the interesting thing is that language learners do get a lot of things wrong. Linguistic theory ought to be able to account for the types of errors that they make, especially when errors crop up as patterns in stages of development. However, I still consider your thought experiment here to be flawed in a number of ways, not the least of which is that L1 learners produce flaps and glottal stops at a very early age. And they produce them in interesting patterns which suggest that they have phonological rules that do not seem to be based on observation of adult articulation. Rather, they seem to have more to do with the difficulty of replicating adult pronunciation. IOW, phonology isn't all about recognition of patterns in adult behavior. It is also very much about suppression of misarticulation.
Let me propose a different way of looking at it. Suppose that devoicing, flapping, premature glottal closure, etc., are all natural tendencies that afflict the production of phonetic targets. That makes them equivalent to speech impediments, but only when they impede desired phonetic output. Hence, suppression of those tendencies would be important
only when they contradict desired phonetic output. If those impediments arise naturally in the process of attempted articulation, then they are
not learned. At least part, if not all, of the "phonological system" that emerges is essentially the set of unsuppressed natural tendencies to misarticulate. Except that they aren't technically misarticulations, if they don't impede pronunciation. Indeed, they aid desired articulation. And, BTW, suppression of undesirable movements is exactly what happens when people learn any kind of muscular coordination, so this is not really as radical an idea as it may sound. So I haven't rejected your notion of observation-based "contradiction", but I have rejected your idea that operations such as devoicing, flapping, and glottal substitution are acquired or learned on the basis of observation.
True, but do you have any evidence that this is a significant factor in the behavior of L1 learners? Have you looked at patterns of articulation in longitudinal studies? I think you'll find that there are some rather spectacular substitutions occurring in the early stages of language acquisition. If one proposes a theory of phonological acquisition, one ought to be able to account for them.
Unfortunately, research in that area is at such a primitive state that there is nothing one can say about how phonetic variation becomes phonological variation. As H&R point out, infants don't have adult-style control over their motor systems, and you can't tell from impressionistic transcriptions of acoustic outputs what the pre-motor representation of various tokens of "lamb" is. As with all accounts of historical change, there is a great volume of speculation and few prospects for empirically testing these claims. Currently popular "biased coin toss" theories of phonological change likewise assume that ambiguous signals may tend to be resolved in a particular directions a small percent of the time which eventually leads to sound change, and this also hasn't been verified. My opinion is that the root of the research problem lies in data reduction techniques in child language studies, and adult language studies for that matter.
I do think that David Stampe produced some rather brilliant analyses of some known longitudinal studies back in the 1960s, but he was starting with the assumptions I've taken here--that what was going on was largely evidence of a failure to suppress massive amounts of misarticulations. What he showed was that the careful records kept on changing articulation revealed patterns of global misarticulation that changed on the basis of selective suppression. What emerged in his work was a concept of phonological system as a residual set of benign articulatory constraints on articulation. Unfortunately, his work coincided with the emergence of generative phonology and its concommitant rejection of the dividing line between phonology and morphophonology--a line that he scrupulously maintained. So he pretty much became background noise, especially since he had nothing much to say about the rest of the grammar.
I thought of a nice clear example of the problem. It has been believed by some that there is a rule in English devoicing /l/ after /s/, exemplified by slip, which some claim has the surface form [sl̥ɪp]. My understanding of the phonetic literature on this matter is that there is not any such rule, and instead, linguists have misinterpreted the extent / timing of glottal opening associated with /s/ and other voiceless consonants, where /s/ and aspirated kʰ have a large glottal opening which then overlaps /l/ on the left, leading linguists to sometimes think that there is a rule of devoicing. It is possible that at some point, the raw acoustic data will change enough and be reinterpreted so that this phenomenon is phonologized and English develops a surface segment [l̥]. But before we can detect that happening, we would need to have a better understanding of the difference between physical outputs or articulation, which are not representations, and phonological representations – and we would also need to resolve the question of whether the mind has such a thing as a "phonetic representation". Longitudinal studies of those questions can't be answered until we have some hope of answering those questions non-longitudinally.
Longitudinal studies exist in the literature. The problem is that generative phonology does not have any way to explain the data. However, I would treat the progressive devoicing of /l/ in
slip as essentially the same process that forces English speakers to devoice initial obstruent clusters, thus preventing the occurrence of English minimal pairs like /sbIl/ and /spIl/. Of course, the non-existence of mixed-voice obstruent clusters is sometimes treated as a rule-based, and sometimes as a static-based, constraint on representation, which seems to miss a generalization. However, that is not a problem with generative phonology, but with all alternation-based theories of phonology. The flaw in thinking goes all the way back to Baudouin de Courtenay's seminal conception of the phonology/morphophonology divide as an alternational dichotomy. He had no concept of phonological derivation, and subsequent phonologists stuck with his paradigm. In modern phonological theory, alternations motivate the concept of rules, but constraints on collocations of sounds tend to be seen as static constraints on underlying or superficial representation. A phonological theory based on automatic substitutions would not need static constraints to explain any phonemic or phonetic patterns.
