Assessing feature specification in surface phonological representations through simulation and classification of phonetic data
Jason Shaw, Shigeto Kawahara
February 2018
 

Many previous studies have argued that phonology may leave some phonetic dimensions unspecified in surface representations either because of deletion or lexical underspecification. Lacking a phonologically specified phonetic target, the phonetic signal in these cases can only be structured by phonetic interpolation between segments flanking the targetless element. However, natural variability in the phonetic signal presents a challenge for identifying instances of phonetic interpolation. Our approach to this challenge is to explicitly model phonetic variability providing a computational link between phonological hypotheses and phonetic data. To this end, we set up stochastic generators of competing phonological hypotheses and use them to compute, on a token-by-token basis, the likelihood that a phonetic signal is the consequence of phonetic interpolation, defined as a smooth interpolation between flanking segments, or, alternatively, that it is structured by a phonological target. The empirical material used to demonstrate the approach comes from Electromagnetic Articulography recordings of high vowel devoicing in Japanese. We use Discrete Cosine Transform to express tongue dorsum movement trajectories as a small number of frequency components (cosines differing in frequency and amplitude) that correspond to linguistically meaningful signal modulations, i.e., articulatory gestures. Our stochastic generators operate over this frequency space, generating tongue dorsum movements with realistic variation according to the presence or absence of a lingual articulatory gesture for the devoiced vowel. Finally, a Bayesian classifier trained on simulations of the targetless trajectory assigns posterior probabilities to the data. Results indicate that /u/ is optionally produced without a vowel height target in Tokyo Japanese and that the frequency of targetlessness varies across phonological environments.
Format: [ pdf ]
Reference: lingbuzz/003878
(please use that when you cite this article)
Published in: Phonology, to appear
keywords: phonetics, computational phonology, underspecification, articulatory target, ema, discrete cosine transform, bayesian classification, phonology
Downloaded:493 times

 

[ edit this article | back to article list ]