Over the years, I've gone back and forth on my view of the schwa. Recently, I had an idea to revolutionize phonological analysis, and I think I'm really onto something. So I figured before I go off the deep-end and abandon segmental phonology altogether, I should give my view of the case for (and against) the schwa.
At first, I thought Why not just let English words have three levels of stress? It seems like you need three levels to explain words like dictionary. Marking syllables with levels of stress from 1 (primary) to 3 (stressless), you get something like this: dic1tio3nar2y3. In the second syllable, you might say you have a schwa, but in the last syllable, it's different. It seems like a stressless high front tense vowel. (A schwi perhaps?)
With three levels of stress, there's no need to postulate the schwa as an extra phoneme when you can just call it the same vowel as in the word stuff (or maybe book) but with less stress. But what about the stressless syllables in spasm and little? There clearly has to be some sort of default vowel to be inject where it's not represented in spelling. The best candidate is probably the same sound as the word uh, but it feels a little arbitrary to elevate just one particular vowel. There's something very intuitively appealing about the idea of a "nothing" vowel that doesn't really count for stress or rhythm or even spelling.
There's another problem with ignoring the schwa and using three levels of stress. One very important extra detail that must be given is that, in a given word, only one syllable may have level 1 stress. As many syllables as you want can have level 2 or 3. Wouldn't it be a bit easier if all stressless syllables used the schwa, and therefore you only needed to point to which single syllable in a word has the primary stress? (All other non-schwa syllables would have secondary stress of course.)
Let me diverge for a minute to say what I've noticed about yod sounds versus R sounds in American English. Both are something like a vowel, but not quite a vowel because there's too much constriction. Both seem to become syllabic (as in pretty and water) and can even take stress (as in feet and bird). Both serve as glides that can intervene between two would-be consecutive vowels. They operate in the same way.
You can extend these observations to the other approximants of English (L and W), but it's not quite so clean and easy. In the word little and normal, L seems to become syllabic. This could easily be called a combination of the schwa and L. With stress, however, it's not clear if a syllabic L exists in words like full or maybe in words like cult. Most dictionaries will tell you something like, "Neither. The first word uses /ul/, and the second uses /Ul/."
With W, it gets really tricky. Does a syllabic W (or schwa plus W) exist? Different vowel sounds lend themselves to different levels of stress. For example, the last syllable of Julio has a little more stress than that of Julia. If we're operating on just three levels of phonologically significant stress (at least on the morpheme level), there has to be some sort of acid test to determine whether a syllable has secondary stress or has a schwa for its nucleus. I say that test involves the phoneme /t/.
If a /t/ falls between some vowel (or diphthong) and another vowel, the latter can be called a schwa only if the /t/ is normally realized as a flap. So the word data does end with a schwa. How about schwa-glide combinations? Pretty ends with /@y/. Butter ends with /@r/. Little ends with /@l/. Photo ends with /@w/. I can't think of anything ending with an /uw/ sound that passes the test, so the /Uw/ sound must have exclusive claim on /ew/. This means that virtue has two stressed syllables, but virtual has only one.
Back to the main topic, if you use a schwa as the nucleus of every stressless syllable, you only need two levels of stress. Because only one syllable in a given word can receive the primary stress, you need only mark a single syllable as "stressed" and leave the rest unmarked.
That's convenient for stressless schwas, but what about situations in which stressed vowels lose their identity? Before R sounds, eng sounds, and semivowels, the identity of vowels phonemes becomes less important, and so vowels are said to "merge". To me, this defeats the purpose of having a phonemic inventory that supposedly holds all the building blocks needed to articulate the language.
So in conclusion, I'd say you need the schwa if you want to use the current model of segmental phonology. But that model has some problems.