Speaker Identification Using Nasal Coarticulation

Abstract

A new approach which used the statistical properties of the nasal spectra was used to quantitatively study the coarticulation of nasal consonants with the vowels in the isolated /h ə′ CVd/ utterances. The Euclidean distance between two nasal spectra means followed by front vowels and by back vowels along each frequency dimension was used as the acoustic measure of the coarticulation of [m,n] and the following vowel context. The coarticulation of [m] was found to to be about 50% larger than that of [n]. Moreover, the strong speaker idiosyncratic characteristics were found for the coarticulation process, which is not likely to be modified in natural speech. Speaker identification was performed using the [m] coarticulation measure and a correlation decision criterion. The results indicated that the coarticulation cues are much better than nasal spectrum alone, which has already been demonstrated to be one of the best acoustic cues for speaker identification. [This work was supported by the AFOSR Grant 69-1776.]