Roles of temporal patterns of vowel-intrinsic cues in model identification of Korean vowels in spontaneous speech

초록

This study investigates the perceptual roles of temporal patterns of vowel-intrinsic cues to identify Korean diphthong and monophthong signals in spontaneous speech from Seoul Corpus. Formant frequencies of vowel signals were sampled at 20%, 40%, 50%, 60%, and 80% of duration, and were averaged to be represented in the acoustic vowel space. Diphthongs showed two different drastic transitions from the onset to 50% of duration and from 50% to the offset, while monophthongs showed one straight-line transition from the onset to the offset. A neural network pattern recognition classifier was trained and tested on vowel signals, which included diphthongs and monophthongs, with various sets of formant samplings along duration coupled with F3, F0, and/or duration. The results show that three- and four-sample F1 and F2 characterized vowel signals better than one- and two-sample F1 and F2. F3, F0 and duration, when combined, had notable perceptual influences on spectral model identification of vowels.

키워드