Phonetic realization as an optimization problem

In Flemming (1997, 2001, 2011) I propose that phonetic grammars specify the desired properties for utterances (e.g. minimal articulator velocity, acoustic targets for speech sounds), and realizations are then selected so as to best satisfy these constraints (an optimization problem). This differs from the standard conception in which phonetic grammars consist of complex rules for mapping phonological representations onto different phonetic realizations according to context. The optimization approach yields simpler analyses of compromises between the demands of consecutive speech sounds. For example, where two consecutive sounds require very different positions of the tongue, as in a sequence like 'two' [tu], a constraint against fast articulator movements results in each sound being produced a little closer to the target for the other, reducing the difference between them. This approach is applied to the analysis of tone realization in Flemming and Cho (2017) and Cho and Flemming (2015).

A unified model of phonetics and phonology.

Phonetics and phonology are usually regarded as separate components of grammar, operating in terms of different representations and principles, but an integrated model of phonetics and phonology offers a number of advantages. On the one hand, the constraints posited in Dispersion Theory refer to articulatory effort and perceptual distinctiveness, which are phonetic concepts, so proper formulation of these constraints depends on referring to the detailed phonetic realization of words. On the other hand, the constraints employed in my analyses of phonetic realization are closely related to the constraints posited in Dispersion Theory to account for phonological patterns, so it appears that the same constraints may be applicable at all levels of analysis.
In (2001) I argued that the existence of extensive parallels between phonetic and phonological phenomena provides evidence for this hypothesis, and developed a framework for a unified model. This approach does not deny the existence of phonological categories and categorical phenomena, but derives them from phonetically detailed representations instead of stipulating their existence.
My work on natural classes (2004) is relevant in this connection since restrictions on the natural classes of sounds that pattern together in processes are often claimed to be determined by restrictions on the phonological feature set. I show that these classes are determined more by the nature of the set of phonological constraints than by the contents of the phonological feature set, so enriching the feature set to add phonetic detail does not lead to problematic predictions concerning the range of expected natural classes. This lays the groundwork for deriving generalizations about natural classes from detailed phonetic representations.
In (2005a), I applied the unified approach in an analysis of phonetic and phonological vowel reduction in which both patterns derive from the Dispersion Theory constraints.
For example, Italian contrasts seven vowels in stressed syllables [i, e, E, a, O, o, u], but only five in unstressed syllables [i, e, a, o, u], a pattern referred to as phonological vowel reduction. Phonetically, the low vowel [a] is raised in unstressed syllables, so its quality is closer to the vowel in English 'cut' (phonetic reduction). I argue that these patterns are related: raising of low vowels in unstressed syllables is motivated by effort constraints keep seven vowels distinct from each other, so the number of contrasts is reduced. This analysis can only be formalized in a unified model of phonetics and phonology. I have also tested the predictions of the model of vowel reduction against experimental data on reduced vowels in English (2007, 2009, 2006).