Predicting pre- and post-vocalic stop consonant place from the vowel in a Korean spontaneous speech corpus
Soonhyun Hong (Inha University)
Abstract
A neural network model was trained on temporal F2 frequencies sampled singly or
doubly along the first half of vowel articulation in CVX tokens and the second half in
XVC in a Korean spontaneous speech corpus to predict pre- and post-vocalic stop
place, respectively. Model performance on temporal F2 frequencies sampled doubly
at vowel onset/offset and target constituted the best cues for pre- and post-vocalic stop
place, though prevocalic place was predicted slightly better than postvocalic place.
Then, secondary cues were added to the dynamic F2 predictors for further training.
The individual contributions of speaking rate, F0, gender, vowel duration, and static
F1 and F3 frequencies sampled singly at vowel onset/offset were not as big as the
temporal F1 and F3 frequencies sampled at vowel onset/offset and target, the latter of
which enhanced model prediction substantially. The dynamic F1 and F3 predictors
facilitated postvocalic place prediction more than prevocalic place prediction. Vowel
identity enhanced prevocalic place prediction substantially but did not enhance
postvocalic place prediction as much. This was due to the observation that F1/F2
transitions of all vowel categories were substantially more centralized in the F1/F2
vowel space at vowel offset than at onset or target. Vowel quality became less
distinctive at vowel offset. Nevertheless, temporal formant samples plus vowel
identity constituted the best cues for stop place.
Keywords
coarticulation, stop place, formant transitions, vowel category, corpus, neural network