Turkish vowel harmony is very systematic, but in a little-studied class of words, appears to break down. I propose that this apparent harmony failure is actually covert harmony, or exceptional transparency.
Background: The backness of vowels in Turkish suffixes is determined by the backness of the nearest vowel in the word they attach to (1).
(1) Nom. Dative Gloss
a. kadran → kadran-a, *kadran-e 'clockface'
b. beden → beden-e, *beden-a 'body'
Harmony within roots is often violated, but harmony between suffixes and the nearest root vowel is extremely robust (2). Nonetheless, in certain roots, /a/ exceptionally behaves as transparent and suffixes must surface with front vowels (3).
(2) Singular Plural Gloss
a. kitap → kitap-lar, *kitap-ler 'book'
b. kahve → kahve-ler, *kahve-lar 'coffee'
a. harf → harf-ler, *harf-lar 'letters'
b. dikkat → dikkat-ler, *dikkat-lar 'attentions'
To account for this exceptional transparency, I extend Smolensky's (2006) headed feature domains theory, and propose that the input to phonology contains headless feature domains.
Proposal: Existing theories of vowel harmony require a segmental harmony trigger. But harf contains no [-back] segment to head a front feature domain and select front suffixes, so harf-ler remains unexplainable. Extending Smolensky (2006), I propose that all features are feature domains, which replace traditional segmental feature specifications.
In Turkish, backness domains normally coincide with roots. Alternating suffixes lack their own backness domains and merge into the stem's domain. Exceptional transparency is driven by faithfulness considerations: the root as a whole is a front domain, but it contains /a/ in an embedded back domain. Suffixes are incorporated into the outermost backness domain as usual, resulting in transparency.
(4) (dik(ka)bt)f) + lEr → (dik(ka)btler)f)
Since vowels must realize the backness of their containing domains, high- and mid- vowels cannot occur in conflicting embedded domains. However, as a low vowel, /a/ is more central than the other back vowels, and unlike the other Turkish back vowels, it lacks an exactly height-matched front counterpart phoneme. Perceptual reasons, then, lead the Turkish constraint-ranking to allow only /a/ as an exceptionally transparent vowel. The feature-as-domain analysis of vowel harmony models the selection of front suffixes by harf and dikkat without the need to posit phonetically unrealized front features on consonants, or any other additional apparatus.