On the basis of abundant evidence afforded by Korean and Turkish, it will be shown that the autosegmental analysis of vowel harmony(VH) the representative model of which is supplied in Clements(1980) is empirically and conceptually inadequate in crucial respects, and thin an alternative analysis of VH will be proposed. For the brevity of exposition claim (1) and principle (2) are presented in advance, and the examples that follow will furnish evidence for them: (1) Root-initial vowels(RIV`s) are lexically specified with harmonic feature(HF), but vowels elsewhere are not specified with it in regular cases. (2) The specified HF may not be changed by a word-level rule. Ex. 1: Alternant pairs like ya1a~yala are found in the 15th-century Korean. The latter alternant is derived by the optional disharmony rule which applies before VH. Its trigger y specifies the second vowel, not lexically specified with HF, with the HF [-ATR], skipping over the RIV lexically specified with the HF [+ATR]. By virtue of (1) and (2), it is inevitable that the disharmony rule should behave as it does. Exs. 2-3: In the 15th-century Korean, *V in the sequence Cy*V can be any vowel specified with HF in root-initial syllables(RIS`s), but it is invariably [+ATR] a elsewhere by a lexical redundancy rule(LRR). In spite of the LRR, the HF lexically specified, in accordance with (1), in RIS`s remains intact in observance of (2). The very same phenomenon also obtains in the present-day Korean. Ex. 4; In Korean, the first disappearance of a of the harmonic pair i : a (i.e. a>i) had taken place only in non-root-initial syllables(NRIS`s). The immunity of the RIS is attributable to (1) and (2). And the reason for the direction of the change a>i is that the feature complex [+voc, -cons, +high, +back, -round, 0 ATR] in NIRS`s, formerly realized as a or i by VH, ceased to be filled in with HF by VH, with the result that the unfilled feature complex as it was became fixed as neutral i, This strongly supports evidence for the latter part of (1). Exs, 5-6: As the direct consequence of a>i, i acts synchronically as neutral exclusively in NRIS`s, whereas the harmonic pair i and a still persist in RIS`s. The asymmetry is derivable from (2), (1) and an extension of (1) which says that every root should bear at least one HF as trigger of VH. In the present-day Korean, u is neutral only in NRIS`s. What is true of the neutral i is equally true of the neutral a. Ex. 7: In Korean, the second disappearance of a (i.e. a>a), which had remained in RIS`s after its first disappearance, cited in Ex 4, occurred in the late 18th century. The [-ATR] a lexically specified, in accordance with (1), with the HF could reasonably change to the [-ATR] vowel a, which is phonetically least different from a in the vowel systam, rather than to its former harmonic counterpart i of the [+ATR] set. It is to be noted that the change is in the different direction from that of the first change. Ex. 8: In the Istanbul dialect of Turkish, a short vowel is unrounded before a palatal exclusively in NRIS`s as in /UsI+mE+yIs/→[usumiyis], and the palatal umlaut rule is assumed to apply before two VH`s. The root-initial it lexically specified with the HF [+ round] is not to be affected by the umlaut rule. This is consistent with (1), and (2) is operative here. Ex. 9: In Turkish, only in NRIS`s are non-high vowels, in regular cases, specified with the HF [-round] by a LRR. In conformity with (2), the LRR cannot affect the HF [+round] lexically specified, consistent with (1), in RIS`s. According to Clements, in lexical representations the autosegments of HF in a separate tier are unbound in regular cases, but they are lexically bound to opaque vowels. The representations where RIV`s are not specified with HF are not empirically justified at all, except that they are an extension from the autosegmental tonal representations; in fact, they run counter to the empirical evidence for (1). Furthermore, in view of the fact that the HF`s of both RIV`s and opaque vowels function equally as triggers, the representations in which the unbound and bound autosegments of HF coexist are conceptually complex, and such represen±ations entail the establishment of two additional priority clauses to implement the well-formedness conditions, The natural corollary of (1) is that VH applies obligatorily and unboundedly affix-ward to the right and left, from the HF (i.e. the trigger) in a RIV (and an opaque vowel). Consequently, we can dispense with the well-formedness conditions, together with the two priority clauses, contingent upon the lexical representations sketched above. It must be added that the domain of VH is also defined by (1) and (2).