Second International Conference on Spoken Language Processing (ICSLP'92)
Banff, Alberta, Canada
Speaker variation was examined in cepstral vectors extracted from 24 /CVCq/ nonsense words read by five male and five female speakers. Dynamic Time Warping (DTW) was used for segment synchronization and a ratio of inter/intra speaker distances was employed as a measure of speaker specificity. The following rank order of increasing speaker specificity was found: plosives + /r/; fricatives; short vowels; nasals; long vowels. Transition segments proved to be less speaker specific than vowel steady states. The factor speaker sex was not found to be of relevance. These results may contribute to the fields of automatic speaker identification/verification and automatic speech recognition.
Bibliographic reference. Heuvel, Henk van den / Rietveld, Toni (1992): "Speaker related variability in cepstral representations of dutch speech segments", In ICSLP-1992, 1581-1584.