Abstract: Speech, an analog phenomenon, has traditionally been modeled through discrete mechanisms and processes. These methods have given way to dynamic approaches to speech processing that rely on modeling the non-linear and non-contiguous co-production or coarticulation of speech gestures. Acoustic variation due to vowel-to-vowel coarticulation, once phonologized, has been shown to be a conditioning factor for development of vowel harmony patterns (Przezdziecki 2000, Ohala 1994). The resultant reduction in phonetic distinctiveness between vowels, however, is known to be compensated perceptually (Beddor et al. 2002). In this presentation, we report on results from artificial neural network (ANN) and support vector machine (SVM) models trained to predict the identity of the vowels in V1CV2 sequences. We find that our models predict the identity of V1 consistently better than V2 in Telugu non-harmonic V1CV2 sequences, and V2 consistently better than the V1 in Bengali non-harmonic and harmonic sequences. We believe this directionality reflects the relative extent of coarticulatory effects in the anticipatory and carryover directions. We show that while vowel harmony patterns advanced from the anticipatory direction leading to neutralization of vowel contrasts between /iCu/->[uCu] sequences in Telugu, and Advance Tongue Root (ATR) harmony in Bengali, the extent of carryover coarticulation is still greater in non-harmonic contexts for Telugu even though it remains greater in the anticipatory direction in Bengali. Coarticulatory propensity and directionality, therefore, crucially interact with parametric notions of contrast in languages that exhibit vowel harmony.