ABSTRACT
This chapter presents a method for building different Markov models (sMMs) that faithfully reflect the structure of the different proposals. It shows how the standard concept of Markov models, in which a single model corresponds to a single segment, can be considered to be the model-theoretic interpretation of a segment-based theory of phonology. The chapter investigates the relation of the typical segment-based and feature-based models from the perspective of model-theoretic formalization of phonology. The major problem is that features lack well-understood acoustic cues. While certain features, such as voicing, are reasonably easy to detect, others, such as rounding, resist acoustic characterization. Since model-theoretic interpretation leads naturally to standard Markov models in the segmental case, it is expected to lead to something even better in the autosegmental case. A key issue in the design of sMMs is to decide which models have direct access to the signal and which rely only on the states of other models.
