The relative contributions of non-spectral cues to static and dynamic spectral model identification of Korean monophthong signals in Seoul Corpus

초록

Though the importance of spectral characteristics at the steady-state central sections of Korean monophthongal signals in the hVd syllable has been amply reported in the literature, it has been rarely studied whether dynamic spectral measurements sampled multiply across the temporal dimension can better characterize Korean vowels in spontaneous speech than static spectral measurements at a (steady-state) central section. Furthermore, the perceptual influence of non-spectral cues on the spectral properties of vowels in vowel perception has been frequently reported in the literature, but few reports have been released on the relative amount of the individual perceptual contributions of non-spectral cues (e.g., gender, speaking rate, duration, F0, place and manner of the flanking phones, etc.) on the spectral properties of vowels in vowel perception. Neural Network pattern recognition modeling on spectral identification of Korean monophthong signals in Seoul Corpus showed that dynamic spectral models fitted to non-spectral cues, identified vowel signals better than static spectral models. Furthermore, flanking phone identities, and manner and place of flanking phones (i.e., coarticulation information) were the most contributive to spectral vowel identification. However, F0, speaking rate, duration, gender, and speaker’s age showed little or almost no contribution.

키워드