Simultaneous prediction of the flanking stop consonant places from the vowel transitions in a Korean spontaneous speech corpus using a neural network classifier

초록

This study aims to predict pre- and post-vocalic stop places simultaneously from the vowel transitions in StopPlace-V-StopPlace tokens in a Korean spontaneous speech corpus. A neural network classifier was trained on F1, F2, and F3 sampled at the vowel onset, target, and offset of the vowel transitions to predict nine types (alveolar-V-alveolar, alveolar-V-bilabial, alveolar-V-velar, bilabial-V-alveolar, bilabial-V-bilabial, bilabial-V-velar, velar-V-alveolar, velar-V-bilabial, and velar-V-velar) for new tokens. Multiple-sampled F2 values proved to be better predictors, achieving the classifier’s prediction accuracy of 32.8%, compared to multiple-sampled F1 or F3. However, both multiple-sampled F1 and F3 predictors contributed to the classifier’s prediction alongside multiple-sampled F2 predictors, resulting in an increased prediction accuracy of 44.9%. The inclusion of vowel identity further increased the prediction to 53.3%. However, gender, F0, speaking rate, and vowel duration did not significantly contribute to the prediction compared to multiple-sampled F1 and F3 predictors and vowel identity. Stop places appear to be robustly manifested by vowel category-specific F1, F2, and F3 transitions from the vowel onset to offset.

키워드