Roles of dynamic lower formant patterns of vowel transitions and vowel identity in predicting the prevocalic consonant manner in Korean spontaneous speech
Soonhyun Hong (Inha University)
Abstract
The present study investigates the role of F1-F3 patterns of temporal vowel transitions and vowel identity in predicting the preceding consonant’s manner of articulation in forty hours of spontaneous Korean speech in Seoul Corpus. A neural network model was trained and tested on selections of raw F1-F3 frequencies sampled at 10% intervals from the onset to the target of the vowel, optionally with(out) vowel identity. It was found that manner could be manifested robustly by spectral patterns of dynamic vowel transitions characterized by F1-F3 measurements sampled at multiple points of the vowel duration in time. However, the full dynamic F1-F3 transition patterns sampled at all 10% intervals from the vowel onset to the target were not necessary. Rather, the dynamic F1-F3 transition patterns from the vowel onset to about 20% of the vowel duration robustly provided enough information to identify the preceding manner. The dynamic spectral pattern model performed robustly better with an additional predictor, vowel identity, but not with F0 or gender.
Keywords
coarticulation, manner contrasts, dynamic F1-F3, vowel identity, corpus