This morning, my daughter Mei pronounced the word `butter` in a way that I found quite interesting. She said [bU?t.d@r] (with [?t] being a glottalized /t/ and [@r] being simply a stressless syllabic American /r/). This was in the middle of her saying the word several times, every other time using a flap (with no clear syllable division): [bU^@r].
I know this is only one phonetic form from one speaker, but it's interesting to consider how she could've come to that particular pronunciation. The ambisyllabicity explanation of the flap comes to mind since we have both a coda and an onset in this example where normally we'd have a single sound.
My ears aren't perfect, so it might've been more like [bU?.d@r]. Is it possible that /t/ could be better analyzed as a sequence (or perhaps a "stacked" combination) of /?/ plus /d/? This would explain the lack of aspiration in combinations like [st]; they'd really be /sd/ in the underlying representation. Since clusters within a syllable share the same voicing feature, it's hard to say whether the UR of a voiceless sound in a cluster is really voiced or not!
Maybe you didn't notice, but I made a pretty bold statement just now. It might need some qualification.
For clusters of obstruents, it's obvious that if one's voiceless they all are. Just look at a word like `stacked`, which could easily be /sdakd/ in the UR. And you might also be aware of sonorant devoicing in onset clusters like /kr/, /pl/, /sn/, /str/, /hy/, etc.
For codas, it's not so simple. Postvocalic nasals nasalize the preceding vowel almost obligatorily in English. (If you don't do this, it sounds about as wrong to me as not aspirating your stops.) When this vowel nasalization happens before a voiceless stop, the nasal consonant can delete (or at least devoice), leaving a purely voiceless coda.
Approximants like /r/, /l/, /w/, and /y/ don't devoice in codas, but actually it seems like they don't even occur in codas! Would-be coda glides are pretty much always analyzed as part of diphthongal nuclei. The patterning of /r/ with preceding vowels is pretty much identical to that of /y/ as an off-glide, so it seems pretty clear that /r/ should also be part of the syllabic nucleus.
/l/ is interesting in that (in the most prestigious form of American English at least) it doesn't limit the vowels that can occur before it. In this respect, it seems more like a consonant than an off-glide. But there's a strong tendency for this "dark L" sound to behave more like a vowel or glide across dialects, so I don't think it too outrageous a claim to say that a vowel-L combination should be analyzed in the same way as diphthongs and vowel-R nuclei.
I've wound up a ways off from where I started in this post, but it's all related. Any thoughts?