I have been following the discussion on flapping. While I tend to think that Bob is on the right track, there is one thing that really sets my teeth on edge, and that is his term “Underoptimization”. What is that supposed to mean? Sounds like a bad tune-up.
There seems to be a misconception here about what Lexicon Optimization is. Lexicon Optimization doesn’t mean that inputs have to look like outputs, or that inputs have to be fully specified either. It means only that the same grammar that determines outputs also predicts inputs. So depending on the grammar, Lexicon Optimization might imply that underlying forms are underspecified.
In fact one of the earliest papers to employ Lexicon Optimization did that very thing: “Licensing and Underspecification in Optimality Theory” by Itô, Mester, and Padgett. In that paper they adopt a series of constraints called NasVoi, SonVoi, etc. NasVoi says:
[+nasal] ⊃ [+voice]
And then they adopt a principle called Licensing cancellation
Licensing Cancellation: If F ⊃ G, then ¬(F ∧ G). “If the specification F implies the specification G, then it is not the case that F licenses G.”
Which, if you think about it, does exactly what Bob is trying to do. So while Bob might want to call it “Underoptimization”, to me it looks like good ole Lexicon Optimization with a grammar that contains a constraint favoring underspecification.
But actually there is another way to think about the whole problem. What is flapping, and why do we have it? Of course flapping causes a voiceless stop to be voiced. But presumably the real issue is that a coronal stop is short to begin with, and there is pressure to make it even shorter. Once the closure gets very short and fast it is difficult to maintain voicelessness, and even if the segment is devoiced, the devoicing is ineffective. I don’t think the flap is voiced at all phonologically, but rather that the imperative to make the closure short overrides the need to make the voicing distinction perceptible. So one could argue that the output of flap is not /ɾ/, i.e., not a voiced sonorant, at all, but rather /T/. If one does that, then the analysis returns to a position where output = input.
Finally what about Thomason’s and Stampe’s observations? Slow speech creates different conditions. I tend to agree with Eric in thinking that this type of experiment doesn’t show us underlying forms. Rather it shows us the grammar of English. Once the word is slowed each syllable tends to get a bit of stress, and the /T/ becomes a proper closure, a stop. As a stop it must have a voicing specification, and the unmarked voicing for a closure is [-voice]. A [t] before a stress receives aspiration. That is just what we know about English.