Speaker variability on a continuum of spectral sub-bands from 297-speakers' non-contemporaneous cepstra of Japanese vowels

In this paper we describe the results of a preliminary investigation of interand intra-speaker variability in the vowel cepstra extracted from a forensically realistic database of non-contemporaneous telephone recordings of 297 adult, male speakers of Japanese. The methodology adopted to conduct this investigation is based on one-way ANOVA analysis, with the novelty of replacing the standard Euclidean distance with a more flexible band-selective formulation of the cepstral distance (Clermont and Mokhtari, 1994). The results of our investigation show a strong similarity between traditional formant-based F-ratios, and the cepstrally-based F-ratios obtained over consecutive sub-bands. The similarity reinforces the possibility to explore formant-related regions directly from the cepstrum. By virtue of the statistical depth of the data used, the results also provide convincing evidence that the whole-spectrum (i.e., full-band) approach based on the cepstrum, often overvalued in many studies, does not provide complete insights into the spectral regions of speaker-specific importance.