Assessing surface phonological specification through simulation and classification of phonetic trajectories

Many previous studies have argued that phonology may leave some phonetic dimensions unspecified in surface representations. We introduce computational tools for assessing this possibility though simulation and classification of phonetic trajectories. The empirical material used to demonstrate the approach comes from electromagnetic articulography recordings of high-vowel devoicing in Japanese. Using Discrete Cosine Transform, tongue-dorsum movement trajectories are decomposed into a small number of frequency components (cosines differing in frequency and amplitude) that correspond to linguistically meaningful signal modulations, i.e. articulatory gestures. Stochastic generators of competing phonological hypotheses operate in this frequency space. Distributions over frequency components are used to simulate (i) the vowel-present trajectories and (ii) the vowel-absent trajectories. A Bayesian classifier trained on simulations assigns posterior probabilities to unseen data. Results indicate that /u/ is optionally produced without a vowel-height specification in Tokyo Japanese and that the frequency of such targetlessness varies systematically across phonological environments.