Roles of dynamic properties of F2 and vowel identity in predicting preceding consonant place in CVX in Korean spontaneous speech

초록

A neural network model was trained and tested on various spectral frequencies sampled at 10% intervals from vowel onset to target to predict the preceding consonant place (bilabial, alveolar, palatal, or velar) in CVX signals in forty-hour-long spontaneous speech in Korean Seoul Corpus. It was found that models trained on F2 samples taken at more than two sections between onset and target across vowel duration predicted place more robustly than those trained on F2 samples taken singly at onset or doubly at onset and target which locus equations are based on. It was also found that vowel identity as an additional predictor enhanced model prediction performance substantially better than static or dynamic F1, F3, F0, or gender. These results show that fine-grained temporal resolution of F2 and vowel identity may constitute major cues for preceding place in CV coarticulation. It is suggested that F2 transitions of a phonemic vowel category preceded by consonant place are allophonically realized as a function of the preceding place.

키워드