Roles of dynamic patterns of lower formants, vowel identity, and gender in predicting postvocalic consonant place in Korean spontaneous speech
Soonhyun Hong (Inha University)
Abstract
A neural network model was trained and tested on various temporal lower formant
frequencies sampled at 10% intervals from the vowel offset backwards to the target to
predict the postvocalic place (bilabial, alveolar, palatal, or velar) in XVC tokens in
Korean spontaneous speech in Seoul Corpus. Performance of the model, trained on
F2 samples taken doubly at the vowel offset and target or at 70-80% of the vowel
duration and target, performed robustly better than F2 samples taken singly at the
offset. Vowel identity, temporal F1 and F3 samples served to disambiguate places
with ambiguous F2 transitions more than gender, suggesting that gender
normalization was not essential in the present model. Compared to the work on
model prediction of prevocalic place from vowel tokens in Seoul Corpus in Hong
(2022), postvocalic consonants exerted slightly stronger coarticulation resistance over
vowel signals than prevocalic consonants. Furthermore, the contribution of vowel
category information as an additional predictor was substantially smaller in model
prediction of postvocalic place than prevocalic place, which was attributed to vowel
neutralization in the time course of vowel articulation toward the vowel offset.
Keywords
coarticulation, postvocalic place, dynamic F1-F3, corpus, neural network model,
vowel neutralization